Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

A review of machine learning applications in wildfire science and management

A review of machine learning applications in wildfire science and management A review of machine learning applications in wild re science and management 1,2 2 y3 Piyush Jain , Sean C P Coogan , Sriram Ganapathi Subramanian , Mark Crowley z3 4 2 , Steve Taylor , and Mike D Flannigan Natural Resources Canada,Canadian Forest Service, Northern Forestry Centre, Edmonton, AB Canadian Partnership for Wildland Fire Science, University of Alberta, Renewable Resources, Edmonton, AB University of Waterloo, Electrical and Computer Engineering, Waterloo, ON Natural Resources Canada,Canadian Forest Service, Paci c Forestry Centre, Victoria, BC August 20, 2020 Abstract Arti cial intelligence has been applied in wild re science and management since the 1990s, with early applications including neural networks and expert systems. Since then the eld has rapidly progressed congruently with the wide adoption of machine learning (ML) methods in the environmental sciences. Here, we present a scoping review of ML applications in wild re science and management. Our overall objective is to improve awareness of ML methods among wild re researchers and managers, as well as illustrate the diverse and challenging range of problems in wild re science available to ML data scientists. To that end, we rst present an overview of popular ML approaches used in wild re science to date, and then review the use of ML in wild re science as broadly categorized into six problem domains, including: 1) fuels characterization, re detection, and mapping; 2) re weather and climate change; 3) re occurrence, susceptibility, and risk; 4) re behavior prediction; 5) re e ects; and 6) re management. Furthermore, we discuss the advantages and limitations of various ML approaches relating to data size, computational requirements, generalizability, and interpretability, as well as identify opportunities for future advances in the science and management of wild res within a data science context. In total, we identi ed 300 relevant publications up to the end of 2019, where the most frequently used ML methods across problem domains included random forests, MaxEnt, arti cial neural networks, decision trees, support vector machines, and genetic algorithms. As such, there exists opportunities to apply more current ML methods | including deep learning and agent based learning | in the wild re sciences, especially in instances involving very large multivariate datasets. We must recognize, however, that despite the ability of ML methods to learn on their own, expertise in wild re science is necessary to ensure realistic modelling of re processes across multiple scales, while the complexity of some ML methods, such as deep learning, requires a dedicated and sophisticated knowledge of their application. Finally, we stress that the wild re research and management communities play an active role in providing relevant, high quality, and freely available wild re data for use by practitioners of ML methods. Keywords: machine learning, wild re science, re management, wildland re, support vector ma- chine, arti cial neural network, decision trees, Bayesian networks, reinforcement learning, deep learning piyush.jain@canada.ca ss@uwaterloo.ca mcrowley@uwaterloo.ca arXiv:2003.00646v2 [cs.LG] 19 Aug 2020 1 Introduction Wildland re is a widespread and critical element of the earth system [Bond and Keeley, 2005], and is a continuous global feature that occurs in every month of the year. Presently, global annual area burned is estimated to be approximately 420 Mha [Giglio et al., 2018], which is greater in area than the country of India. Globally, most of the area burned by wild res occurs in grasslands and savannas. Humans are responsible for starting over 90% of wildland res, and lightning is responsible for almost all of the remaining ignitions. Wildland res can result in signi cant impacts to humans, either directly through loss of life and destruction to communities, or indirectly through smoke exposure. Moreover, as the climate warms we are seeing increasing impacts from wildland re [Coogan et al., 2019]. Consequently, billions of dollars are spent every year on re management activities aimed at mitigating or preventing wild res' negative e ects. Understanding and better predicting wild res is therefore crucial in several important areas of wild re management, including emergency response, ecosystem management, land-use planning, and climate adaptation to name a few. Wildland re itself is a complex process; its occurrence and behaviour are the product of several interrelated factors, including ignition source, fuel composition, weather, and topography. Furthermore, re activity can be examined viewed across a vast range of scales, from ignition and combustion processes that occur at a scale of centimeters over a period of seconds, to re spread and growth over minutes to days from meters to kilometers. At larger extents, measures of re frequency may be measured over years to millennia at regional, continental, and planetary scales (see Simard [1991] for a classi cation of re severity scales, and Taylor et al. [2013] for a review of numerical and statistical models that have been used to characterize and predict re activity at a range of scales). For example, combustion and re behavior are fundamentally physicochemical processes that can be usefully represented in mechanistic (i.e., physics- based) models at relatively ne scales [Coen, 2018]. However, such models are often limited both by the ability to resolve relevant physical processes, as well as the quality and availability of input data [Ho man et al., 2016]. Moreover, with the limitations associated with currently available computing power it is not feasible to apply physical models to inform re management and research across the larger and longer scales that are needed and in near real time. Thus, wild re science and management rely heavily on the development of empirical and statistical models for meso, synoptic, strategic, and global scale processes [Simard, 1991], the utility of which are dependent upon their ability to represent the often complex and non-linear relationships between the variables of interest, as well as by the quality and availability of data. While the complexities of wildland re often present challenges for modelling, signi cant advances have been made in wild re monitoring and observation primarily due to the increasing availability and capability of remote-sensing technologies. Several satellites (eg. NASA TERRA, AQUA and GOES), for instance, have onboard re detection sensors (e.g., Advanced Very High Resolution Radiometer (AVHRR), Moderate Resolution Imaging Spectroradiometer (MODIS), Visible Infrared Imaging Radiometer Suite (VIIRS)), and these sensors along with those on other satellites (e.g., LANDSAT series) routinely monitor vegetation distributions and changes. Additionally, improvements in numerical weather prediction and climate models are simultaneously o ering smaller spatial resolutions and longer lead forecast times [Bauer et al., 2015] which potentially o er improved predictability of extreme re weather events. Such developments make a data-centric approach to wild re modeling a natural evolution for many research problems given sucient data. Consequently, there has been a growing interest in the use of Machine Learning (ML) methodologies in wild re science and management in recent years. Although no formal de nition exists, we adopt the conventional interpretation of ML as the study of computer algorithms that can improve automatically through experience [Mitchell, 1997]. This approach is necessarily data-centric, with the performance of ML algorithms dependent on the quality and quantity of available data relevant to the task at hand. The eld of ML has undergone an explosion of new algorithmic advances in recent years and is deeply connected to the broader eld of Arti cial Intelligence (AI). AI researchers aim to understand and synthesize intelligent agents which can act appropriately to their situation and objectives, adapt to changing environments, and learn from experience [Poole and Mackworth, 2 2010]. The motivations for using AI for forested ecosystem related research, including disturbances due to wild re, insects, and disease, were discussed in an early paper [Schmoldt, 2001], while Olden et al. [2008] further argued for the use of ML methods to model complex problems in ecology. The use of ML models in the environmental sciences has seen a rapid uptake in the last decade, as is evidenced by recent reviews in the geosciences [Karpatne et al., 2017], forest ecology [Liu et al., 2018], extreme weather prediction [McGovern et al., 2017], ood forecasting [Mosavi et al., 2018], statistical downscaling [Vandal et al., 2018], remote sensing [Lary et al., 2016], and water resources [Shen, 2018, Sun and Scanlon, 2019]. Two recent perspectives have also made compelling arguments for the application of deep learning in earth system sciences [Reichstein et al., 2019] and for tackling climate change [Rolnick et al., 2019]. To date, however, no such paper has synthesized the diversity of ML approaches used in the various challenges facing wildland re science. In this paper, we review the current state of literature on ML applications in wild re science and management. Our overall objective is to improve awareness of ML methods among re researchers and managers, and illustrate the diverse and challenging problems in wild re open to data scientists. This paper is organized as follows. In Section 2, we discuss commonly used ML methods, focusing on those most commonly encountered in wild re science. In Section 3, we give an overview of the scoping review and literature search methodology employed in this paper. In this section we also highlight the results of our literature search and examine the uptake of ML methods in wild re science since the 1990s. In Section 4, we review the relevant literature within six broadly categorized wild re modeling domains: (i) Fuels characterization, re detection, and mapping; (ii) re weather and climate change; (iii) re probability and risk; (iv) re behavior prediction; (v) re e ects; and (vi) re management. In Section 5, we discuss our ndings and identify further opportunities for the application of ML methods in wild re science and management. Finally, in Section 6 we o er conclusions. Thus, this review will serve to guide and inform both researchers and practitioners in the wild re community looking to use ML methods, as well as provide ML researchers the opportunity to identify possible applications in wild re science and management. 2 Arti cial Intelligence and Machine Learning \De nition: Machine Learning - (Methods which) detect patterns in data, use the uncov- ered patterns to predict future data or other outcomes of interest" from Machine Learning: A Probabilistic Perspective, 2012 [Murphy, 2012]. ML itself can be seen as a branch of AI or statistics, depending who you ask, that focuses on building predictive, descriptive, or actionable models for a given problem by using collected data, or incoming data, speci c to that problem. ML methods learn directly from data and dispense with the need for a large number of expert rules or the need to model individual environmental variables with perfect accuracy. ML algorithms develop their own internal model of the underlying distributions when learning from data and thus need not be explicitly provided with physical properties of di erent parameters. Take for example, the task of modeling wildland re spread, the relevant physical properties which include fuel composition, local weather and topography. The current state of the art in wild re prediction includes physics-based simulators that re ghters and strategic planners rely on to take many critical decisions regarding allocation of scarce re ghting resources in the event of a wild re [Sullivan, 2007]. These physics- based simulators, however, have certain critical limitations; they normally render very low accuracies, have a prediction bias in regions where they are designed to be used, are often hard to design and implement due to the requirement of large number of expert rules. Furthermore, modelling many complex environmental variables is often dicult due to large resource requirements and complex or heterogeneous data formats. ML algorithms, however, learn their own mappings between parametric rules directly from data and do not require expert rules, which is particularly advantageous when the number of parameters are quite large and their physical properties quite complex, as in the case of wildland re. Therefore, a ML approach to wild re response may help to avoid many of the limitations of physics-based simulators. 3 A major goal of this review is to provide an overview of the various ML methods utilized in wild re sci- ence and management. Importantly, we also provide a generalized framework for guiding wild re scientists interested in applying ML methods to speci c problem domains in wildland re research. This conceptual framework, derived from the approach in [Murphy, 2012] and modi ed to show examples relevant to wild- land re and management is shown in Fig. 1. In general, ML methods can be identi ed as belonging to one of three types: supervised learning; unsupervised learning; or, agent based learning. We describe each of these below. Supervised Learning - In supervised ML all problems can be seen as one of learning a parametrized function, often called a \model", that maps inputs (i.e., predictor variables) to outputs (or \target vari- ables") both of which are known. The goal of supervised learning is to use an algorithm to learn the parameters of that function using available data. In fact, both linear and logistic regression can be seen as very simple forms of supervised learning. Most of the successful and popular ML methods fall into this category. Unsupervised Learning - If the target variables are not available, then ML problems are typically much harder to solve. In unsupervised learning, the canonical tasks are dimensionality reduction and clustering, where relationships or patterns are extracted from the data without any guidance as to the \right" answer. Extracting embedded dimensions which minimize variance, or assigning datapoints to (labelled) classes which maximize some notion of natural proximity or other measures of similarity are examples of unsupervised ML tasks. Agent Based Learning - Between supervised and unsupervised learning are a group of ML methods where learning happens by simulating behaviors and interactions of a single or a group of autonomous agents. These are general unsupervised methods which use incomplete information about the target vari- ables, (i.e., information is available for some instances but not others), requiring generalizable models to be learned. A speci c case in this space is Reinforcement Learning [Sutton and Barto, 1998], which is used to model decision making problems over time where critical parts of the environment can only be observed interactively through trial and error. This class of problems arises often in the real world and require ecient learning and careful de nition of values (or preferences) and exploration strategies. In the next section, we present a brief introduction to commonly used ML methods from the aforemen- tioned learning paradigms. We note that this list is not meant to be exhaustive, and that some methods can accommodate both supervised and unsupervised learning tasks. It should be noted that the classi - cation of a method as belonging to either ML or traditional statistics is often a question of taste. For the purpose of this review | and in the interests of economy | we have designated a number of methods as belonging to traditional statistics rather than ML. For a complete listing see tables 1 and 2. 2.1 Decision Trees Decision Trees (DT) [Breiman, 2017] belong to the class of supervised learning algorithms and are another example of a universal function approximator, although in their basic form such universality is dicult to achieve. DTs can be used for both classi cation and regression problems. A decision tree is a set of if-then- else rules with multiple branches joined by decision nodes and terminated by leaf nodes. The decision node is where the tree splits into di erent branches, with each branch corresponding to the particular decision being taken by the algorithm whereas leaf nodes represent the model output. This could be a label for a classi cation problem or a continuous value in case of a regression problem. A large set of decision nodes is used in this way to build the DT. The objective of DTs are to accurately capture the relationships between input and outputs using the smallest possible tree that avoids over tting. C4.5 [Quinlan, 1993] and Classi cation and Regression Trees (CART, [Breiman et al., 1984]) are examples of common single DT algorithms. Note that while the term CART is also used as an umbrella term for single tree methods, we use DT here to refer to all such methods. The majority of decision tree applications are ensemble decision tree (EDT) models that use multiple trees in parallel (ie. bootstrap aggregation or bagging) or sequentially (ie., boosting) to arrive at a nal model. In this way, EDTs make use of many weak learners to form a strong learner while being robust to over tting. EDTs are well described in many ML/AI textbooks and 4 Machine Learning Types Agent-based learning Supervised Learning Unsupervised Learning Map labelled input to known Understand patterns and Single or multiple agents output discover output interact with environment Continuous Categorical Reward based rather than Target Target Target variable not available target action Variable Variable Dimensionality Decision Regression Classification Clustering Optimization Reduction Making NB, DT, CART, SOM, DQN, A3C, ANN, DT, BRT, KM, SOM, GA, MCTS, A3C RF, DNN, GP autoencoders, MCTS RF, KNN, SVM, autoencoders, ANN, GA, RNN, t-SNE, RF, BRT, K-SVM, LR, LDA GMM, ISODATA, MAXENT, CA, MaxEnt, PCA, HMM, HC, PCA, MLR, GLM, GAM factor analysis DBSCAN - Fire susceptibility - Fuels - Fire Detection - Optimizing fire - Fuel treatment - Landscape - Fire Spread/Burn characterization - Fire mapping simulators - Planning and controls on fire area prediction - Fire detection - burned area - Fire spread and Policy - Fire susceptibility - Fire occurence - Fire mapping prediction growth - Wildfire response - Fire Spread/Burn - Fire severity - Fire weather area prediction - Smoke Prediction prediction - Climate Change Figure 1: A diagram showing the main machine learning types, types of data, and modeling tasks in relation to popular algorithms and potential applications in wild re science and management. Note that the algorithms shown bolded are core ML methods whereas non-bolded algorithms are often not considered ML. are widely available as implemented libraries. 2.1.1 Random Forests A Random Forest (RF) [Breiman, 2001] is an ensemble model composed of a many individually trained DTs, and is the most popular implementation of a bagged decision tree. Each component DT in a RF model makes a classi cation decision where the class with the maximum number of votes is determined to be the nal classi cation for the input data. RFs can also be used for regression where the nal output is determined by averaging over the individual tree outputs. The underlying principle of the RF algorithm is that a random subset of features is selected at each node of each tree; the samples for training each component tree are selected using bagging, which resamples (with replacement) the original set of datapoints. The high performance of this algorithm is achieved by minimizing correlation between trees while reducing model variance so that a large number of di erent trees provides greater accuracy than individual trees. However, this improved performance comes at the cost of an increase in bias and loss of interpretability (although variable importance can still be inferred through permutation tests). Possible Popular Task Data Approach Applications Algorithms Machine Learning Methods A3C Asynchronous Advantage Actor-Critic AdaBoost Adaptive Boosting ANFIS Adaptive Neuro Fuzzy Inference System ANN Arti cial Neural Networks ADP Approximate Dynamic Programming (a.k.a. reinforcement learning) Bag Bagged Decision Trees BN Bayesian Networks BRT Boosted Regression Trees (a.k.a. Gradient Boosted Machine) BULC Bayesian Updating of Land Cover CART Classi cation and Regression Tree CNN Convolutional Neural Network DNN Deep Neural Network DQN Deep Q-Network DT Decision Trees (incl. CART, J48, jRip) EDT Ensemble Decision Trees (incl. bagging and boosting) ELM Extreme Machine Learning (i.e., feedforward network) GA Genetic algorithms (a.k.a evolutionary algorithms) GBM Gradient Boosted Machine (a.k.a. Boosted Regression Trees, incl. XGBoost, AdaBoost, LogitBoost) GMM Gaussian Mixture Models GP Gaussian Processes HCL Hard Competitive Learning HMM Hidden Markov Models ISODATA Iterative Self-Organizing DATA algorithm KNN K Nearest Neighbor KM K-means Clustering LB LogitBoost (incl. AdaBoost) LSTM Long Short Term Memory MaxEnt Maximum Entropy MCMC Markov Chain Monte Carlo MCTS Monte Carlo Tree Search MLP Multilayer Perceptron MDP Markov Decision Process NB Naive Bayes NFM Neuro-Fuzzy models PSO Particle Swarm Optimization RF Random Forest RL Reinforcement Learning RNN Recurrent Neural Network SGB Stochastic Gradient Boosting SOM Self-organizing Maps SVM Support Vector Machines t-SNE T-distributed Stochastic Neighbor Embedding Table 1: Table of acronyms and de nitions for common machine learning algorithms referred to in text. 2.1.2 Boosted Ensembles Boosting describes a strategy where one combines a set of weak learners | usually decision trees | to make a strong learner using a sequential additive model. Each successive model improves on the previous 6 Non-machine learning methods DBSCAN Density-based spatial clustering of applications with noise GAM Generalized Additive Model GLM Generalized Linear Model KLR Kernel Logistic Regression LDA Linear Discriminant Analysis LR Logistic Regression MARS Multivariate Adaptive Regression Splines MLR Multiple Linear Regression PCA Principal Component Analysis SLR Simple Linear regression Table 2: Table of acronyms and de nitions for common data analysis algorithms usually considered as foundational to, or outside of, machine learning itself. by taking into account the model errors from the previous model, which can be done in more than one way. For example, the adaptive boosting algorithm, known as AdaBoost [Freund and Shapire, 1995], works by increasing the weight of observations that were previously misclassi ed. This can in principle reduce the classi cation error leading to a high level of precision [Hastie et al., 2009]. Another very popular implementation for ensemble boosted trees is Gradient Boosting Machine (GBMs), which makes use of the fact that each DT model represents a function that can be di erentiated with re- spect to its parameters, i.e., how much a change in the parameters will change the output of the function. GBMs sequentially build an ensemble of multiple weak learners by following a simple gradient which points in the opposite direction to weakest results of the current combined model [Friedman, 2001]. The details for the GBM algorithm are as follows. Denoting the target output as Y , and given a tree-based ensemble model, represented as a function T (X ) ! Y , after adding i weak learners already, the \perfect" function for the (i + 1)th weak learner would be h(x) = T (x) Y which exactly corrects the previous model (i.e., T (x) = T (x) + h(x) = Y ). In practice, we can only approach this perfect update (i+1) i by performing functional gradient descent where we use an approximation of the true residual (i.e., loss function) at each step. In our case this approximation is simply the sum of the residuals from each weak learner decision tree L(Y; T (X )) = Y T (X ). GBM explicitly uses the gradient r L(Y; T (X ) of the i T i loss function of each tree to t a new tree and add it to the ensemble. In a number of domains, and particularly in the context of ecological modeling GBM is often referred to as Boosted Regression Trees (BRTs) [Elith et al., 2008]. For consistency with the majority of literature reviewed in this paper we henceforth use the latter term. It should be noted that while deep neural networks (DNNs) and EDT methods are both universal function approximators, EDTs are more easily interpretable and faster to learn with less data than DNNs. However, there are fewer and fewer cases where trees-based methods can be shown to provide superior performance on any particular metric when DNNs are trained properly with enough data (see for example, Korotcov et al. [2017]). 2.2 Support Vector Machines Another category of supervised learning includes Support Vector Machines (SVM) [Hearst et al., 1998] and related kernel-based methods. SVM is a classi er that determines the hyper-plane (decision boundary) in an n-dimensional space separating the boundary of each class, for data in n dimensions. SVM nds the optimal hyper-plane in such a way that the distance between the nearest point of each class to the decision boundary is maximized. If the data can be separated by a line then the hyper-plane is de ned to be of the form w x + b = 0 where the w is the weight vector, x is the input vector and b is the bias. The distance of the hyper-plane to the closest data point d, called a support vector, is de ned as the margin of separation. The objective is to nd the optimal hyper-plane that minimizes the margin. If they are 7 not linearly separable, kernel SVM methods such as Radial Basis Functions (RBF) rst apply a set of transformations to the data to a higher dimensional space where nding this hyperplane would be easier. SVMs have been widely used for both classi cation and regression problems, although recently developed deep learning algorithms have proved to be more ecient than SVMs given a large amount of training data. However, for problems with limited training samples, SVMs might give better performances than deep learning based classi ers. 2.3 Arti cial Neural Networks and Deep Learning The basic unit of an Arti cial Neural Network (ANN) is a neuron (also called a perceptron or logistic unit). A neuron is inspired by the functioning of neurons in mammalian brains in that it can learn simple associations, but in reality it is much simpler than its biological counterpart. A neuron has a set of inputs which are combined linearly through multiplication with weights associated with the input. The nal weighted sum forms the output signal which is then passed through a (generally) non-linear activation function. Examples of activation functions include sigmoid, tanh, and the Recti ed Linear Unit (ReLU). This non-linearity is important for general learning since it creates an abrupt cuto (or threshold) between positive and negative signals. The weights on each connection represent the function parameters which are t using supervised learning by optimizing the threshold so that it reaches a maximally distinguishing value. In practice, even simple ANNs, often called Multi-Layered Perceptrons (MLP), combine many neuron units in parallel, each processing the same input with independent weights. In addition, a second layer of hidden neuron units can be added to allow more degrees of freedom to t general functions, see Figure 2(a). MLPs are capable of solving simple classi cation and regression problems. For instance, if the task is one of classi cation, then the output is the predicted class for the input data, whereas in the case of a regression task the output is the regressed value for the input data. Deep learning [LeCun et al., 2015] refers to using Deep Neural Networks (DNNs) which are ANNs with multiple hidden layers (nominally more than 3) and include Convolutional Neural Networks (CNNs) popularized in image analysis and Recurrent Neural Networks (RNNs) which can be used to model dynamic temporal phenomena. The architecture of DNNs can vary in connectivity between nodes, the number of layers employed, the types of activation functions used, and many other types of hyperparameters. Nodes within a single layer can be fully connected, or connected with some form of convolutional layer (e.g., CNNs), recurrent units (e.g., RNNs), or other sparse connectivity. The only requirement of all these connectivity structures and activation functions is that they are di erentiable. Regardless of the architecture, the most common process of training a ANN involves processing input data fed through the network layers and activation functions to produce an output. In the supervised setting, this output is then compared to the known true output (i.e., labelled training data) resulting in an error measurement (loss or cost function) used to evaluate model performance. The error for DNNs are commonly calculated as a cross entropy loss between the predicted output label and the true output label. Since every part of the network is mathematically di erentiable we can compute a gradient for the entire network. This gradient is used to calculate the proportional change in each network weight needed to produce an in nitesimal increase in the likelihood of the network producing the same output for the most recent output. The gradient is then weighted by the computed error, and thereafter all the weights are updated in sequence using a backpropagation algorithm [Hecht-Nielsen, 1992]. ANNs can also be con gured for unsupervised learning tasks. For example, self-organizing maps (SOMs) are a form of ANN adapted for dealing with spatial data and have therefore found widespread use in the atmospheric sciences [Ski c and Francis, 2012]. A SOM is a form of unsupervised learning that consists of a two-dimensional array of nodes as the input layer, representing say, a gridded atmospheric variable at a single time. The algorithm clusters similar atmospheric patterns together and results in a dimensionality reduction of the input data. More recently, unsupervised learning methods from deep learning, such as autoencoder networks, are starting to replace SOMs in the environmental sciences [Shen, 2018]. 8 Figure 2: Logistic regression can be seen as basic building block for neural networks, with no hidden layer and a sigmoid activation function. Classic shallow neural networks (also known as Multi-Layer Perceptrons) have at least one hidden layer and can have a variety of activation functions. Deep neural networks essentially have a much larger number of hidden layers as well as use additional regularization and optimization methods to enhance training. 2.4 Bayesian methods 2.4.1 Bayesian Networks Bayesian networks (Bayes net, belief network; BN) are a popular tool in many applied domains because they provide an intuitive graphical language for specifying the probabilistic relationships between variables as well as tools for calculating the resulting probabilities [Pearl, 1988]. The basis of BNs is Bayes' theorem, which relates the conditional and marginal probabilities of random variables. BNs can be treated as a ML task if one is trying to automatically t the parameters of the model from data, or even more challenging, to learn the best graphical structure that should be used to represent a dataset. BNs have close ties to causal reasoning, but it is important to remember that the relationships encoded in a BN are inherently correlational rather than causal. BNs are acyclic graphs, consisting of nodes and arrows (or arcs), de ning a probability distribution over variables U . The set of parents of a node (variable) X , denoted  , are all nodes with directed arcs going into X . BNs provide compact representation of conditional distributions since p(X jX ; : : : ; X ) = p(X j ) where X ; : : : ; X are arranged to be all of the ancestors of X i 1 i1 i X 1 i1 i other than its direct parents. Each node X is associated with a conditional probability table over X and its parents de ning p(Xj ). If a node has no parents, a prior distribution is speci ed for p(X ). The joint probability distribution of the network is then speci ed by the chain rule P (U ) = p(Xj ). X2U 2.4.2 Na ve Bayes A special case of a BN is the Na ve Bayes (NB) classi er, which assumes conditional independence between input features, which allows the likelihood function to be constructed by a simple multiplication of the conditional probability of each input variable conditional on the output. Therefore, while NB is fast 9 and straightforward to implement, prediction accuracy can be low for problems where the assumption of conditional independence does not hold. 2.4.3 Maximum Entropy Maximum Entropy (MaxEnt), originally introduced by Phillips et al. [2006], is a presence only framework that ts a spatial probability distribution by maximising entropy, consistent with existing knowledge. MaxEnt can be considered a Bayesian method since it is compatible with an application of Bayes Theorem as existing knowledge is equivalent to specifying a prior distribution. MaxEnt has found widespread use in landscape ecology species distribution modeling [Elith, Phillips, Hastie, Dud k, Chee, and Yates, 2011], where prior knowledge consists of occurrence observations for the species of interest. 2.5 Reward based methods 2.5.1 Genetic Algorithms Genetic algorithms (GA) are heuristic algorithms inspired by Darwin's theory of evolution (natural selec- tion) and belong to a more general class of evolutionary algorithms [Mitchell, 1996]. GAs are often used to generate solutions to search and optimization problems by using biologically motivated operators such as mutation, crossover, and selection. In general, GAs involve several steps. The rst step involves creating an initial population of potential solutions, with each solution encoded as a chromosome. Second a tness function appropriate to the problem is de ned, which returns a tness score determining how likely an individual is to be chosen for reproduction. The third step requires the selection of pairs of individuals, denoted as parents. In the fourth step, a new population of nite individuals are created by generating two new o spring from each set of parents using crossover, whereby a new chromosome is created by some random selection process from each parents chromosomes. In the nal step called mutation, a small sample of the new population is chosen and a small perturbation is made to the parameters to maintain diversity. The entire process is repeated many times until the desired results are satisfactory (based on the tness function), or some measure of convergence is reached. 2.5.2 Reinforcement Learning Reinforcement learning (RL) represents a very di erent learning paradigm to supervised or unsupervised learning. In RL, an agent (or actor) interacts with its environment and learns a desired behavior (set of actions) in order to maximize some reward. RL is a solution to a Markov Decision Process (MDP) where the transition probabilities are not explicitly known but need to be learned. This type of learning is well suited to problems of automated decision making, such as required for automated control (e.g., robotics) or for system optimization (e.g., management policies). Various RL algorithms include Monte Carlo Tree Search (MTCS), Q-Learning, and Actor-Critic algorithms. For an introduction to RL see Sutton and Barto [2018]. 2.6 Clustering methods Clustering is the process of splitting a set of points into groups where each point in a group is more similar to its own group than any other group. There are di erent ways in which clustering can be done, for example, the K-means (KM) clustering algorithm [MacQueen et al., 1967], based on a centroid model, is perhaps the most well-known clustering algorithm. In K-means, the notion of similarity is based on closeness to the centroid of each cluster. K-means is an iterative process in which the centroid of a group and points belonging to a group are updated at each step. The K-means algorithm consists of ve steps: (i) specify the number of clusters; (ii) each data point is randomly assigned to a cluster; (iii) the centroids of each cluster is calculated; (iv) the points are reassigned to the nearest centroids, and (v) cluster centroids are recomputed. Steps iv and v repeat until no further changes are possible. Although KM is the most widely 10 used clustering algorithm, several other clustering algorithms exist including, for example, agglomerative Hierarchical Clustering (HC), Gaussian Mixture Models (GMMs) and Iterative Self-Organizing DATA (ISODATA). 2.7 Other methods 2.7.1 K-Nearest Neighbor The K-Nearest Neighbors (KNN) algorithm is a simple but very e ective supervised classi cation algorithm which is based on the intuitive premise that similar data points are in close proximity according to some metric [Altman, 1992]. Speci cally, a KNN calculates the similarity of data points to each other using the Euclidean distance between the K nearest data points. The optimal value of K can be found experimentally over a range values using the classi cation error. KNN is widely used in applications where a search query is performed such that results should be similar to another pre-existing entity. Examples of this include nding similar images to a speci ed image and recommender systems. Another popular application of KNN is outlier (or anomaly) detection, whereby the points (in a multidimensional space) farthest away from their nearest neighbours may be classi ed as outliers. 2.7.2 Neuro-Fuzzy models Fuzzy logic is an approach for encoding expert human knowledge into a system by de ning logical rules about how di erent classes overlap and interact without being constrained to \all-or-nothing" notions of set inclusion or probability of occurrence. Although early implementations of fuzzy logic systems depended on setting rules manually, and therefore are not considered machine learning, using fuzzy rules as inputs or extracting them from ML methods are often described as \neuro-fuzzy" methods. For example, the Adaptive Neuro-Fuzzy Inference System (ANFIS) [Jang, 1993] fuses fuzzy logical rules with an ANN approach, while trying to maintain the bene ts of both. ANFIS is a universal function approximator like ANNs. However, since this algorithm originated in the 1990s, it precedes the recent deep learning revolution so is not necessarily appropriate for very large data problems with complex patterns arising in high-dimensional spaces. Alternatively, human acquired fuzzy rules can be integrated into ANNs learning; however, it is not guaranteed that the resulting trained neural network will still be interpretable. It should be noted that fuzzy rules and fuzzy logic are not a major direction of research within the core ML community. 3 Literature search and scoping review The combination of ML and wild re science and management comprises a diverse range of topics in a rela- tively nascent eld of multidisciplinary research. Thus, we employed a scoping review methodology [Arksey and O'Malley, 2005, Levac et al., 2010] for this paper. The goal of a scoping review is to characterize the existing literature in a particular eld of study, particularly when a topic has yet to be extensively reviewed and the related concepts are complex and heterogeneous [Pham, Raji c, Greig, Sargeant, Papadopoulos, and Mcewen, 2014]. Furthermore, scoping reviews can be particularly useful for summarizing and dissem- inating research ndings, and for identifying research gaps in the published literature. A critical review of methodological advances and limitations and comparison with other methods is left for future work. We performed a literature search using the Google Scholar and Scopus databases and the key words \wild- re" or \wildland re" or\forest re" or \bush re" in combination with \machine learning" or \random forest" or \decision trees" or \regression trees" or \support vector machine" or \maximum entropy" or \neural network" or \deep learning" or \reinforcement learning". We also used the Fire Research Institute online database (http://fireresearchinstitute.org) using the following search terms: \Arti cial In- telligence"; \Machine Learning"; \Random Forests"; \Expert Systems"; and \Support Vector Machines". 11 50 1996 2000 2005 2010 2015 2019 Year of publication Figure 3: Number of publications by year for 300 publications on topic of ML and wild re science and management as identi ed in this review. SVM RF Other MAXENT KNN GA DT 60 DL BRT BN ANN 1996 2000 2005 2010 2015 2019 Year of publication Figure 4: Number of ML applications by category and by year for 300 publications on topic of ML and wild re science and management as identi ed in this review. Furthermore, we obtained papers from references cited within papers we had obtained using literature databases. After performing our literature search, we identi ed a total of 300 publications relevant to the topic of ML applications in wild re science and management (see supplementary material for a full bibliography). Furthermore, a search of the Scopus database revealed a dramatic increase in the number of wild re and Number of applications Number of publications ML articles published in recent years (see Fig. 3). After identifying publications for review, we further applied the following criteria to exclude non-relevant or unsuitable publications, including: (i) conference submissions where a journal publication describing the same work was available; (ii) conference posters; (iii) articles in which the methodology and results were not adequately described to conduct an assessment of the study; (iv) articles not available to as either by open access or by subscription; and (v) studies that did not present new methodologies or results. 4 Wild re applications In summary, we found a total of 300 journal papers or conference proceedings on the topic of ML applica- tions in wild re science and management. We found the problem domains with the highest application of ML methods was Fire Occurrence , Susceptibility and Risk (127 papers) followed by Fuels Characterization , Fire Detection And Mapping (66 papers), Fire Behaviour Prediction (43 papers), Fire E ects (35 papers), Fire Weather and Climate Change (20 papers), and Fire Management (16 papers). Within Fire Occur- rence, Susceptibility and Risk, the subdomains with the most papers were Fire Susceptibility Mapping (71 papers) and Landscape Controls on Fire (101 papers). Refer to table 3 and the supplementary material for a break-down of each problem subdomain and ML methods used, as well as study areas considered. 4.1 Fuels Characterization, Fire Detection, and Mapping 4.1.1 Fuels characterization Fires ignite in a few fuel particles; subsequent heat transfer between particles through conduction, radiation and convection, and the resulting re behavior (fuel consumption, spread rate, intensity) is in uenced by properties of the live and dead vegetative fuels, including moisture content, biomass, and vertical and horizontal distribution. Fuel properties are a required input in all re behavior models, whether it be a simple categorical vegetation type, as in the Canadian FBP System, or as physical quantities in a 3 dimensional space (eg. see FIRETEC model). Research to predict fuel properties has been carried out at two di erent scales 1) regression applications to predict quantities such as the crown biomass of single trees from more easily measured variables such as height and diameter, and 2) classi cation applications to map fuel type descriptors or fuel quantities over a landscape from visual interpretation of air photographs or by interpretation of the spectral properties of remote sensing imagery. However, relatively few studies have employed ML to wild re fuel prediction, leaving the potential for substantially more research in this area. In an early study, Riano ~ et al. [2005] used an ANN to predict and map the equivalent water thickness and dry matter content of wet and dry leaf samples from 49 species of broad leaf plants using re ectance and transmittance values in the Ispra region of Italy. Pierce et al. [2012] used RF to classify important canopy fuel variables (e.g. canopy cover, canopy height, canopy base height, and canopy bulk density) related to wildland re in Lassen Volcanic National Park, California, using eld measurements, topographic data, and NDVI to produce forest canopy fuel maps. Likewise, Viegas et al. [2014] used RF with Land re and biophysical variables to perform fuel classi cation and mapping in Eastern Oregon. The authors of the aforementioned study achieved relatively high overall modelling accuracy, for example, 97% for forest height, 86% for forest cover, and 84% for existing vegetation group (i.e. fuel type). L opez-Serrano et al. [2016] compared the performance of three common ML methods (i. SVM; ii. KNN; and iii. RF) and multiple linear regression in estimating above ground biomass in the Sierra Madre Occidental, Mexico. The authors reported the advantages and limitations of each method, concluding that that the non-parametric ML methods had an advantage over multiple linear regression for biomass estimation. Garc a et al. [2011] used SVM to classify LiDAR and multispectral data to map fuel types in Spain. Chirici et al. [2013] compared the use of CART, RF, and Stochastic Gradient Boosting SGB, an ensemble tree method that uses both boosting and bagging, for mapping forest fuel types in Italy, and found that SGB had the highest overall accuracy. 13 Section Domain NFM SVM KM GA BN BRT ANN DT RF KNN MAXENT DL NB Other 1.1 Fuels characterization - 2 - - - 1 1 1 4 1 - - - - 1.2 Fire detection 2 3 1 1 1 - 12 - - - - 18 - 3 1.3 Fire perimeter and severity 1 12 1 2 - 1 6 1 4 2 1 - - 6 mapping 2.1 Fire weather prediction - - 1 - - - - - 1 - - - - 3 2.2 Lightning prediction - - - - - - - 1 2 - - - - - 2.3 Climate change - 1 - - - 6 2 2 5 - 7 - - - 3.1 Fire occurrence prediction - 3 - - 1 - 7 1 5 1 2 - 1 4 3.2 Landscape-scale Burned area - 1 1 1 - - 1 1 2 - 1 1 - 1 prediction 3.3 Fire Susceptibility Mapping 2 12 1 3 2 8 16 9 26 - 27 1 2 3 3.4 Landscape controls on re 2 10 1 3 2 19 11 15 40 1 30 1 1 2 4.1 Fire Spread and Growth - - - 13 2 - 4 - 1 1 - 3 - 2 4.2 Burned area and re severity - 7 - 1 1 3 10 7 6 3 - 2 1 5 prediction 5.1 Soil erosion and deposits - - 1 - - - 1 1 - - 1 - - - 5.2 Smoke and particulate levels - 2 - - - 3 3 - 5 2 - - - 2 5.3 Post- re regeneration and - 1 - 1 1 6 1 2 10 - 2 - 1 - ecology 5.4 Socioeconomic e ects - - - - 1 - - - - - - - - - 6.1 Planning and policy - - - 1 1 - - - 2 - - - - 2 6.2 Fuel treatment - - - 1 1 - - - - - - - - 1 6.3 Wild re preparedness and re- - - - 1 2 1 1 - - - 1 1 - 1 sponse 6.4 Social factors - - - - 1 - - - - - - - - - Table 3: Summary of application of ML methods applied to di erent problem domains in wild re science and management. A table of acronyms for the ML methods are given in 1. Note that in some cases a paper may use more than one ML method and/or appear in multiple problem domains. 14 4.1.2 Fire detection Detecting wild res as soon as possible after they have ignited, and therefore while they are still relatively small, is critical to facilitating a quick and e ective response. Traditionally, res have mainly been detected by human observers, by distinguishing smoke in the eld of view directly from a re tower, or from a video feed from a tower, aircraft, or from the ground. All of these methods can be limited by spatial or temporal coverage, human error, the presence of smoke from other res and by hours of daylight. Automated detection of heat signatures or smoke in infra-red or optical images can extend the spatial and temporal coverage of detection, the detection eciency in smoky conditions, and remove bias associated with human observation. The analytical task is a classi cation problem that is quite well suited to ML methods. For example, Arrue et al. [2000] used ANNs for infrared (IR) image processing (in combination with visual imagery, meteorological and geographic data used in a decision function using fuzzy logic), to identify true wild res. Several researchers have similarly employed ANNs for re detection [Al-Rawi et al., 2001, Angayarkkani and Radhakrishnan, 2010, Fernandes et al., 2004a,b, Li et al., 2015, Soliman et al., 2010, Utkin et al., 2002, Sayad et al., 2019]. In addition, Liu et al. [2015] used ANNs on wireless sensor networks to build a re detection system, where multi-criteria detection was used on multiple attributes (e.g. ame, heat, light, and radiation) to detect and raise alarms. Other ML methods used in re detection systems include SVM to automatically detect wild res from videoframes [Zhao et al., 2011], GA for multi-objective optimization of a LiDAR-based re detection system [Cordoba et al., 2004], BN in a vision-based early re detection system [Ko et al., 2010], ANFIS [Angayarkkani and Radhakrishnan, 2011, Wang et al., 2011], and KM [Srinivasa et al., 2008]. CNNs (ie. deep learning), which are able to extract features and patterns from spatial images and are nding widespread use in object detection tasks, have recently been applied to the problem of re detection. Several of these applications trained the models on terrestrial based images of re and/or smoke [Zhang et al., 2016, 2018a,b, Yuan et al., 2018, Akhlou et al., 2018, Barmpoutis et al., 2019, Jakubowski et al., 2019, Jo~ ao Sousa et al., 2019, Li et al., 2018b, 2019, Muhammad et al., 2018, Wang et al., 2019]. Of particular note, Zhang et al. [2018b] found CNNs outperformed a SVM-based method and Barmpoutis et al. [2019] found a Faster region-based CNN outperformed another CNN based on YOLO (\you only look once"). Yuan et al. [2018] used CNN combined with optical ow to include time-dependent information. Li et al. [2018b] similarly used a 3D CNN to incorporate both spatial and temporal information and so were able to treat smoke detection as a segmentation problem for video images. Another approach by Cao et al. [2019] used convolutional layers as part of a Long Short Term Memory (LSTM) Neural network for smoke detection from a sequence of images (ie. video feed). They found the LSTM method achieved 97.8% accuracy, a 4.4% improvement over a single image-based deep learning method. Perhaps of greater utility for re management were re/smoke detection models trained on either unmanned aerial vehicle (UAV) images [Zhao et al., 2018, Alexandrov et al., 2019] or satellite imagery including GOES-16 [Phan and Nguyen, 2019] and MODIS [Ba et al., 2019]. Zhao et al. [2018] compared SVM, ANN and 3 CNN models and found their 15-layer CNN performed best with an accuracy of 98%. By comparison, the SVM based method, which was unable to extract spatial features, only had an accuracy of 43%. Alexandrov et al. [2019] found YOLO was both faster and more accurate than a region-based CNN method in contrast to Barmpoutis et al. [2019]. 4.1.3 Fire perimeter and severity mapping Fire maps have two management applications: 1) Accurate maps of the location of the active re perimeter are important for daily planning of suppression activities and/or evacuations, including modeling re growth 2) Maps of the nal burn perimeter and re severity are important for assessing and predicting the economic and ecological impacts of wildland re and for recovery planning. Historically, re perimeters were sketch-mapped from the air, from a ground or aerial GPS or other traverse, or by air-photo interpretation. Developing methods for mapping re perimeters and burn severity from remote sensing imagery has been an area of active research since the advent of remote sensing in the 1970s, and is mainly concerned with 15 classifying active re areas from inactive or non burned areas, burned from unburned areas (for extinguished res), or re severity measures such as the Normalized Burn Ratio [Lutes et al., 2006]. In early studies using ML methods for re mapping Al-Rawi et al. [2001] and Al-Rawi et al. [2002] used ANNs (speci cally, the supervised ART-II neural network) for burned scar mapping and re detection. Pu and Gong [2004] compared Logistic Regression (LR) with ANN for burned scar mapping using Landsat images; both methods achieved high accuracy (> 97%). Interestingly, however, the authors found that LR was more ecient for their relatively limited data set. The authors in Zammit et al. [2006] performed burned area mapping for two large res that occurred in France using satellite images and three ML algorithms, including SVM, K-nearest neighbour, and the K-means algorithm; overall SVM had the best performance. Likewise, E. Dragozi, I. Z. Gitas, D.G. Stavrakoudis [2011] compared the use of SVM against a nearest neighbour method for burned area mapping in Greece and found better performance with SVM. In fact, a number of studies [Alonso-Benito et al., 2008, Cao et al., 2009, Petropoulos et al., 2010, 2011, Zhao et al., 2015, Pereira et al., 2017, Branham et al., 2017, Hamilton et al., 2017] have successfully used SVM for burned scar mapping using satellite data. Mitrakis et al. [2012] performed burned area mapping in the Mediterranean region using a variety of ML algorithms, including a fuzzy neuron classi er (FNC), ANN, SVM, and AdaBoost, and found that, while all methods displayed similar accuracy, the FNC performed slightly better. Dragozi et al. [2014] applied SVM and a feature selection method (based on fuzzy logic) to IKONOS imagery for burned area mapping in Greece. Another approach to burned area mapping in the Mediterranean used an ANN and MODIS hotspot data [G omez and Pilar Mart n, 2011]. Pereira et al. [2017] used a one class SVM, which requires only positive training data (i.e. burned pixels), for burned scar mapping, which may o er a more sample ecient approach than general SVMs { the one class SVM approach may be useful in cases where good wild re training datasets are dicult to obtain. In Mithal et al. [2018], the authors developed a three-stage framework for burned area mapping using MODIS data and ANNs. Crowley et al. [2019] used Bayesian Updating of Landcover (BULC) to merge burned-area classi cations from three remote sensing sources (Landsat-8, Sentinel-2 and MODIS). Celik [2010] used GA for change detection in satellite images, while Sunar and Ozkan [2001] used the interactive Iterative Self-Organizing DATA algorithm (ISODATA) and ANN to map burned areas. In addition to burned area mapping, ML methods have been used for burn severity mapping, including GA [Brumby et al., 2001], MaxEnt [Quintano et al., 2019], bagged decision trees [S a et al., 2003], and others. For instance, Hultquist et al. [2014] used three popular ML approaches (Gaussian Process Regression (GPR) [Rasmussen and Williams, 2006], RF, and SVM) for burn severity assessment in the Big Sur ecoregion, California. RF gave the best overall performance and had lower sensitivity to di erent combinations of variables. All ML methods, however, performed better than conventional multiple regression techniques. Likewise, Hultquist et al. [2014] compared the use of GPR, RF, and SVM for burn severity assessment, and found that RF displayed the best performance. Another recent paper by Collins et al. [2018] investigated the applicability of RF for re severity mapping, and discussed the advantages and limitations of RF for di erent re and land conditions. One recent paper by Langford et al. [2019] used a 5-layer deep neural network (DNN) for mapping res in Interior Alaska with a number of MODIS derived variables (eg. NDVI and surface re ectance). They found that a validation-loss (VL) weight selection strategy for the unbalanced data set (i.e., the no- re class appeared much more frequently than re) allowed them to achieve better accuracy compared with a XGBoost method. However, without the VL approach, XGBoost outperformed the DNN, highlighting the need for methods to deal with unbalanced datasets in re mapping. 4.2 Fire Weather and Climate Change 4.2.1 Fire weather prediction Fire weather is a critical factor in determining whether a re will start, how fast it will spread, and where it will spread. Fire weather observations are commonly obtained from surface weather station networks operated by meteorological services or re management agencies. Weather observations may be interpolated 16 from these point locations to a grid over the domain of interest, which may include diverse topographical conditions; the interpolation task is a regression problem. Weather observations may subsequently be used in the calculation of meteorologically based re danger indices, such as the Canadian Fire Weather Index (FWI) System [Van Wagner, 1987]. Future re weather conditions and danger indices are commonly forecast using the output from numerical weather prediction (NWP) models (e.g., The European Forest Fire Information System [San-Miguel-Ayanz et al., 2012]). However, errors in the calculation of re danger indices that have a memory (such as the moisture indices of the FWI System) can accumulate in such projections. It is noteworthy that surface re danger measures may be correlated with large scale weather and climatic patterns. To date there has been relatively few papers that address re weather and danger prediction using ma- chine learning. The rst e ort [Crimmins, 2006] used self-organizing maps (SOMs) to explore the synoptic climatology of extreme re weather in the southwest USA. He found three key patterns representing south- westerly ow and large geopotential height gradients that were associated with over 80% of the extreme re weather days as determined by a re weather index. Nauslar et al. [2019] used SOMs to determine the timing of the North American Monsoon that plays a major role on the length of the active re season in the southwest USA. Lagerquist et al. [2017] also used SOMs to predict extreme re weather in northern Alberta, Canada. Extreme re weather was de ned by using extreme values of the Fine Fuel Moisture Code (FFMC), Initial Spread Index (ISI) and the Fire Weather Index (FWI), all components of the Cana- dian Fire Weather Index (FWI) System [Van Wagner, 1987]. Good performance was achieved with the FFMC and the ISI and this approach has the potential to be used in near real time, allowing input into re management decision systems. Other e orts have used a combination of conventional and machine learning approaches to interpolate meteorological re danger in Australia [Sanabria et al., 2013]. 4.2.2 Lightning prediction Lightning is second most common cause of wild res (behind human causes); thus predicting the location and timing of future storms/strikes is of great importance to predicting re occurrence. Electronic lightning detection systems have been deployed in many parts of the world for several decades and have accrued rich strike location/time datasets. Lightning prediction models have employed these data to derive regression relationships with atmospheric conditions and stability indices that can be forecast with NWP. Ensemble forecasts of lightning using RF is a viable modelling approach for Alberta, Canada [Blouin et al., 2016]. Bates et al. [2017] used two machine learning methods (CART and RF) and three statistical methods to classify wet and dry thunderstorms (lightning associated with dry thunderstorms are more likely to start res) in Australia. 4.2.3 Climate Change Transfer modeling, whereby a model produced for one study region and/or distribution of environmental conditions is applied to other cases [Phillips et al., 2006], is a common approach in climate change science. Model transferability should be considered when using ML methods to estimated projected quantities due to climate change or other environmental changes. With regards to climate change, transfer modeling is essentially an extrapolation task. Previous studies in the context of species distribution modeling have indicated ML approaches may be suitable for transfer modeling under future climate scenarios. For exam- ple, Heikkinen et al. [2012] indicated MaxEnt and generalized boosting methods (GBM) have the better transferability than either ANN and RF, and that the relatively poor transferability of RF may be due to over tting. There are several publications on wild res and climate change that use ML approaches. Amatulli et al. [2013] found that Multivariate Adaptive Regression Splines (MARS) were better predictors of future monthly area burned for 5 European countries as compared to Multiple Linear Regression and RF. [Parks et al., 2016] projected re severity for future time periods in Western USA using BRT. Young et al. [2017] similarly used BRT to project future re intervals in Alaska and found up to a fourfold increase in (30 17 year) re occurrence probability by 2100. Several authors used MaxEnt to project future re probability globally [Moritz et al., 2012], for Mediterranean ecosystems [Batllori et al., 2013], in Southwest China [Li et al., 2017], the paci c northwestern USA [Davis et al., 2017], and for south central USA [Stroh et al., 2018]. An alternative approach for projecting future potential burn probability was employed by Stralberg et al. [2018] who used RF to determine future vegetation distributions as inputs to ensemble Burn-P3 simulations. Another interesting paper of note was by Boulanger et al. [2018] who built a consensus model with 2 di erent predictor datasets and 5 di erent regression methods (generalised linear models, RF, BRT, CART and MARS) to make projections of future area burned in Canada. The consensus model can be used to quantify uncertainty in future area burned estimates. The authors noted that model uncertainty for future periods (> 200%) can be higher than that of di erent climate models under di erent carbon forcing scenarios. This highlights the need for further work in the application of ML methods for projecting future re danger under climate change. 4.3 Fire Occurrence, Susceptibility and Risk Papers in this domain include prediction of re occurrence and area burned (at a landscape or seasonal scales), mapping of re susceptibility (or similar de nitions of risk) and analysis of landscape or environ- mental controls on re. 4.3.1 Fire occurrence prediction Predictions of the number and location of re starts in the upcoming day(s) are important to preparedness planning | that is, the acquisition of resources, including the relocation of mobile resources and readiness for expected re activity. The origins of re occurrence prediction (FOP) models go back almost 100 years [Nadeem et al., 2020]. FOP models typically use regression methods to relate the response variable ( re reports or hotspots) to weather, lightning, and other covariates for a geographic unit, or as a spatial probability. The seminal work of Brillinger and others in developing the spatio-temporal FOP framework is reviewed in Taylor et al. [2013] The most commonly used ML method in studies predicting re occurrence were ANNs. As early as 1996, Vega-Garcia et al. [1996] used an ANN for human-caused wild re prediction in Alberta, Canada, correctly predicting 85% of no- re observations and 78% of re observations. Not long after, Alonso-Betanzos et al. [2002] and Alonso-Betanzos et al. [2003] used ANN to predict a daily re occurrence risk index using temperature, humidity, rainfall, and re history, as part of a larger system for real-time wild re management system in the Galicia region of Spain. Vasilakos et al. [2007] used separate ANNs for three di erent indices representing re weather (Fire Weather Index; FWI), hazard (Fire Hazard Index; FHI), and risk (Fire Risk Index) to create a composite re ignition index (FII) for estimating the probability of wild re occurrence on the Greek island of Lesvos. Sakr et al. [2010] used meteorological variables in a SVM to create a daily re risk index corresponding to the number of res that could potentially occur on a particular day. Sakr et al. [2011] then compared the use of SVM and ANN for re occurrence prediction based only on relative humidity and cumulative precipitation up to the speci c day. While Sakr et al. [2011] reported low errors for the number of res predicted by both the SVM and ANN models, ANN models outperformed SVM; however, the SVM performed better on binary classi cation of re/no re. It is important to note, however, that ANNs encompass a wide range of possible network architectures. In an Australian study, Dutta et al. [2013] compared the use of ten di erent types of ANN models for estimating monthly re occurrence from climate data, and found that an Elman RNN performed the best. After 2012, RF became the more popular method for predicting re occurrence among the papers reviewed here. Stojanova et al. [2012] evaluated several machine learning methods for predicting re outbreaks using geographical, remote sensed, and meteorological data in Slovenia, including single classi er methods (i.e., KNN, Naive Bayes, DT (using the J48 and jRIP algorithms), LR, SVM, and BN), and ensemble methods (AdaBoost, DT with bagging, and RF). The ensemble methods DT with bagging and RF displayed the best predictive performance with bagging having higher precision and RF having better 18 recall. Vec n-Arias et al. [2016] found that RF performed slightly better than LR for predicting lightning re occurrence in the Iberian Peninsula, based on topography, vegetation, meteorology, and lightning characteristics. Similarly, Cao et al. [2017] found that a cost-sensitive RF analysis outperformed GLM and ANN models for predicting wild re ignition susceptibility. In recent non-comparative studies, Yu et al. [2017] used RF to predict re risk ratings in Cambodia using publicly available remote sensed products, while Van Beusekom et al. [2018] used RF to predict re occurrence in Puerto Rico and found precipitation was found to be the most important predictor. The maximum entropy (MaxEnt) method has also been used for re occurrence prediction [De Angelis et al., 2015, Chen et al., 2015]. For example, De Angelis et al. [2015] used MaxEnt to evaluate di erent meteorological variables and re-indices (e.g. the Canadian Fire Weather Index, FWI) for daily re risk forecasting in the mountainous Canton Ticino region of Switzerland. The authors of that study found that combinations of such variables increased predictive power for identifying daily meteorological conditions for wild res. Dutta et al. [2016] use a two- stage machine learning approach (ensemble of unsupervised deep belief neural networks with conventional supervised ensemble machine learning) to predict bush- re hot spot incidence on a weekly time-scale. In the rst unsupervised deep learning phase, Dutta et al. [2016] used Deep Belief Networks (DBNet; an ensemble deep learning method) to generate simple features from environmental and climatic surfaces. In the second supervised ensemble classi cation stage, features extracted from the rst stage were fed as training inputs to ten ML classi ers (i.e., conventional supervised Binary Tree, Linear Discriminant Analyser, Na ve Bayes, KNN, Bagging Tree, AdaBoost, Gentle Boosting Tree, Random Under-Sampling Boosting Tree, Subspace Discriminant, and Subspace KNN) to establish the best classi er for bush re hotspot estimation. The authors found that bagging and the conventional KNN classi er were the two best classi ers with 94.5% and 91.8% accuracy, respectively. 4.3.2 Landscape scale burned area prediction The use of ML methods in studies of burned area prediction have only occurred relatively recently compared to other wild re domains, yet such studies have incorporated a variety of ML methods. For example, Cheng and Wang [2008] used an RNN to forecast annual average area burned in Canada, while Archibald et al. [2009] used RF to evaluate the relative importance of human and climatic drivers of burnt area in Southern Africa. Arnold et al. [2014] used Hard Competitive Learning (HCL) to identify clusters of unique pre- re antecedent climate conditions in the interior western US which they then used to construct re danger models based on MaxEnt. Mayr et al. [2018] evaluated ve common statistical and ML methods for predicting burned area and re occurrence in Namibia, including GLM, Multivariate Adaptive Regression Splines (MARS), Regres- sion Trees from Recursive Partitioning (RPART), RF, and SVMs for Regression (SVR). The RF model performed best for predicting burned area and re occurrence; however, adjusted R values were slightly higher for RPART and SVR in both cases. Likewise, de Bem et al. [2018] compared the use of LR and ANN for modelling burned area in Brazil. Both LR and ANN showed similar performance; however, the ANN had better accuracy values when identifying non-burned areas, but displayed lower accuracy when classifying burned areas. 4.3.3 Fire Susceptibility Mapping A considerable number of references (71) used various ML algorithms to map wild re susceptibility, cor- responding to either the spatial probability or density of re occurrence (or other measures of re risk such as burn severity) although other terms such as re vulnerability and risk have also been used. The general approach was to build a spatial re susceptibility model using either remote sensed or agency reported re data with some combination of landscape, climate, structural and anthropogenic variables as explanatory variables. In general, the various modeling approaches used either a presence only framework (e.g., MaxEnt) or a presence/absence framework (e.g., BRT or RF). 19 Early attempts at re susceptibility mapping used CART [Amatulli et al., 2006, Amatulli and Camia, 2007, Lozano et al., 2008]. Amatulli and Camia [2007] compared re density maps in central Italy using CART and multivariate adaptive regression splines (MARS) and found while CART was more accurate that MARS led to smoother density model. More recent work has used ensemble based classi ers, such as RF and BRT, or ANNs (see table S.3.3 in supplementary material for a full list) Several of these papers also compared ML and non-ML methods for re susceptibility mapping and in general found superior performance from the ML methods. Speci cally, Adab [2017] mapped re hazard in the Northeast of Iran, and found ANN performed better than binary logistic regression (BLR) with an AUC of 87% compared with 81% for BLR. Bisquert et al. [2012] found ANN outperformed logistic regression for mapping re risk in the North-west of Spain. Goldarag et al. [2016] also compared ANN and linear regression for re susceptibility mapping in Northern Iran and found ANN had much better accuracy (93.49%) than linear regression (65.76%). Guo et al. [2016b] and Guo et al. [2016a] compared RF and logistic regression for re susceptibility mapping in China and found RF led to better performance. Oliveira et al. [2012] compared RF and LR for re density mapping in Mediterranean Europe and found RF outperformed linear regression. De Vasconcelos et al. [2001] found ANN had better classi cation accuracy than logistic regression for ignition probability maps in parts of Portugal. Referring to table 3 and section S.3.3 of the supplementary material a frequently used ML method for re susceptibility mapping was Maximum Entropy (MaxEnt) which is extensively used in landscape ecology for species distribution modeling [Elith et al., 2011]. In particular, Vilar et al. [2016] found MaxEnt performed better than GLM for re susceptibility mapping in central Spain with respect to sensitivity (i.e., true positive rate) and commission error (i.e., false positive rate), even though the AUC was lower. Of further note, Duane et al. [2015] partitioned their re data into topography-driven, wind-driven and convection-driven res in Catalonia and mapped the re susceptibility for each re type. Other ML methods used for regional re susceptibility mapping include Bayesian networks [Bashari et al., 2016, Dlamini, 2011] and novel hybrid methods such as Neuro-Fuzzy systems [Jaafari et al., 2019, Tien Bui et al., 2017]. Bashari et al. [2016] noted that Bayesian networks may be useful because it allows probabilities to be updated when new observations become available. SVM was also used by a number of authors as a benchmark for other ML methods [Ghorbanzadeh et al., 2019b, Gigovi c et al., 2019, Hong et al., 2018, Jaafari, 2019, Ngoc Thach et al., 2018, Rodrigues and De la Riva, 2014, Sachdeva et al., 2018, Tehrany et al., 2018, Tien Bui et al., 2017, van Breugel et al., 2016, Zhang et al., 2019] but as we discuss below, it did not perform as well as other methods to which it was being compared. There were two applications of ML for mapping global re susceptibility including Moritz et al. [2012] who used MaxEnt and Luo et al. [2013] who used RF. Both of these papers found that at a global scale, precipitation was one of the most important predictors of re risk. The majority of papers considered thus far used the entire study period (typically 4 or more years) to map re susceptibility, therefore neglecting the temporal aspect of re risk. However, a few authors have considered various temporal factors to map re susceptibility. Mart n et al. [2019] included seasonality and holidays as explanatory variables for re probability in northeast Spain. Vacchiano et al. [2018] predicted re susceptibility separately for the winter and summer seasons. Several papers produced maps of re susceptibility in the Eastern US by month of year [Peters et al., 2013, Peters and Iverson, 2017]. Parisien et al. [2014] examined di erences in annual re susceptibility maps and a 31 year climatology for the USA, highlighting the role of climate variability as a driver of re occurrence. In particular, they found FWI90 (the 90th percentile of the Canadian Fire Weather Index) was the dominant factor for annual re risk but not for climatological re risk. Cao et al. [2017] considered a 10 day resolution (corresponding to the available re data) for re risk mapping, which makes their approach similar to re occurrence prediction. In addition to re susceptibility mapping, a few papers focused on other aspects of re risk including mapping probability of burn severity classes [Holden et al., 2009, Parks et al., 2018, Tracy et al., 2018]. Parks et al. [2018] additionally considered the role of fuel treatments on re probability which has obvious implications for re management. Additionally Ghorbanzadeh et al. [2019a] combined re susceptibility maps with vulnerability and infrastructure indicators to produce a re hazard map. 20 A number of papers directly compared three or more ML (and sometimes non-ML) methods for re susceptibility mapping. Here we highlight some of these papers, which elucidate the performance and advantages/disadvantages of various ML methods. Cao et al. [2017] found a cost-sensitive RF model outperformed a standard RF model, ANN as well as probit and logistic regression. Ghorbanzadeh et al. [2019b] compared ANN, SVM and RF and found the best performance with RF. Gigovi c et al. [2019] compared SVM and RF for re susceptibility mapping in combination with Bayesian averaging to generate ensemble models. They found the ensemble model led to marginal improvement (AUC = 0.848) over SVM (AUC=0.834) and RF (AUC=0.844). For mapping both wild re ignitions and potential natural vegetation in Ethiopia van Breugel et al. [2016] also considered ensemble models consisting of a weighted combination of ML methods (RF, SVM, BRT, MaxEnt, ANN, CART) and non-ML methods (GLM and MARS) and concluded the ensemble member performed best over a number of metrics. However, in this paper RF showed the best overall performance of all methods including the ensemble model. Jaafari et al. [2018] compared 5 decision tree based classi ers for wild re susceptibility mapping in Iran. Here, the Alternating Decision tree (ADT) classi er achieved the highest performance (accuracy 94.3%) in both training and validation sets. Ngoc Thach et al. [2018] compared SVM, RF and a Multilayer Perceptron (MLP) neural network for forest re danger mapping in the region of Tjuan chau in Vietnam. They found the performance of all models were comparable although MLP had the highest AUC values. Interestingly Pourtaghi et al. [2016] found that a generalized additive model (GAM) outperformed RF and BRT for re susceptibility mapping in the Golestan province in Iran. This was one of the few examples we found where a non-ML method outperformed ML methods. Rodrigues and De la Riva [2014] compared RF, BRT, SVM and logistic regression for re susceptibility mapping and found RF led to the highest accuracy as well as the most parsimonious model. Tehrany et al. [2018] compared a LogitBoost ensemble-based decision tree (LEDT) algorithm with SVM, RF and Kernel logistic regression (KLR) for re susceptibility mapping in Lao Cai region of Vietnam and found the best performance with LEDT, closely followed by RF. Finally, of particular note, Zhang et al. [2019] compared CNN, RF, SVM, ANN and KLR for re susceptibility mapping in the Yunnan Province of China. This was the only application of deep learning we could nd for re susceptibility mapping. The authors found that CNN outperformed the other algorithms with overall accuracy of 87.92% compared with RF (84.36%), SVM (80.04%), MLP (78.47%), KLR (81.23%). They noted that the bene t of CNN is that it incorporates spatial correlations so that it can learn spatial features. However, the downside is that deep learning models are not as easily interpretted as other ML methods (such as RF and BRT). 4.3.4 Landscape controls on re Many of the ML methods used in re susceptibility mapping have also been used to examine landscape controls { ie. the relative importance of weather, vegetation, topography, structural and anthropogenic variables { on re activity, which may facilitate hypothesis formation and testing or model building. From table 3 the most commonly used methods in this section were MaxEnt, RF, BRT and ANN. These methods all allow for the determination of variable importance (i.e. the relative in uence of predictor variables in a given model of a response variable). A commonly used method to ascertain variable importance is through the use of partial dependence plots [Hastie et al., 2009]. This method works by averaging over models that exclude the predictor variable of interest, with the resulting reduction in AUC (or other performance metrics) representing the marginal e ect of the variable on the response. Partial dependence plots have the advantage of being able to be applied to a wide range of ML methods. A related method for determining variable importance, often used for RFs, is a permutation test which involves random permutation of each predictor variable [Strobl et al., 2007]. Another model-dependent approach used for ANN is the use of partial derivatives (of the activation functions of hidden and output nodes) as outlined by Vasilakos et al. [2009]. It should be noted that while many other methods for model interpretation and variable dependence exist, a discussion of these methods is outside the scope of this paper. In general, the drivers of re occurrence or area burned varied greatly by the study area considered (including the size of area) and the methods used. Consistent with other work on \top down" and \bottom 21 up" drivers of re activity, at large scales climate variables were often determined to be the main drivers of re activity whereas at smaller scales anthropogenic or structural factors exerted a larger in uence. Here we discuss some of the papers that highlight the diversity of results for di erent study areas and spatial scales (global, country, ecoregion, urban) but refer the reader to section S.3.4 of the supplementary material for a full listing of papers in this section. Note that many of the papers listed under section S.3.4 also belong to the re susceptibility mapping section and have already been discussed there. Aldersley et al. [2011] considered drivers of monthly area burned at global and regional scales using both regression trees and RF. They found climate factors (high temperature, moderate precipitation, and dry spells) were the most important drivers at the global scale, although at the regional scale the models exhibited higher variability due to the in uence of anthropogenic factors. At a continental scale Mansuy et al. [2019] used MaxEnt to show that climate variables were the dominant controls (over landscape and human factors) on area burned for most ecoregions for both protected areas and outside these areas, although anthropogenic factors exerted a stronger in uence in some regions such as the Tropical Wet Forests ecoregion. [Masrur et al., 2018] used RF to investigate controls on circumpolar arctic re and found June surface temperature anomalies were the most important variable for determining the likelihood of wild re occurrence on an annual scale. Chingono and Mbohwa [2015] used MaxEnt to model re occurrences in Southern Africa where most res are human-caused and found vegetation (i.e., dry mass productivity and NDVI) were the main drivers of biomass burning. Curt et al. [2015] used BRT to examine drivers of re in New Caledonia. Interestingly, they found that human factors (such as distance to villages, cities or roads) were dominant in uences for predicting re ignitions whereas vegetation and weather factors were most important for area burned. Curt et al. [2016] modeled re probabilities by di erent re ignition causes (lightning, intentional, accidental, negligence professional and negligence personal) in Southeastern France. They found socioeconomic factors (eg. housing and road density) were the dominant factors for ignitions and area burned for human-caused res. Fernandes et al. [2016] used BRT to examine large res in Portugal and found high pyrodiversity (ie. spatial structure due to re recurrence) and low landscape fuel connectivity were important drivers of area burned. Curt et al. [2016] modeled re probabilities by di erent re ignition causes (lightning, intentional, accidental, negligence professional and negligence personal) in Southeastern France. They found socioeconomic factors (eg. housing and road density) were the dominant factors for ignitions and area burned for human-caused res. Leys et al. [2017] used RF to nd the drivers that determine sedimentary charcoal counts in order to reconstruct grass re history in the Great Plains, USA. Not surprisingly, they found re regime characteristics (eg. area burned and re frequency) were the most important variables and concluded that charcoal records can therefore be used to reconstruct re histories. Li et al. [2009] used ANNs to show that wild re probability was strongly in uenced by population density in Japan, with a peak determined by the interplay of positive and negative e ects of human presence. This relationship, however, becomes more complex when weather parameters and forest cover percentage are added to the model. Liu et al. [2013] used BRT to study factors in uencing re size in the Great Xingan Mountains in Northeastern China. Their method included a \moving window" resampling technique that allowed them to look at the relative in uence of variables at di erent spatial scales. They showed that the most dominant factors in uencing re size were fuel and topography for small res, but re weather became the dominant factor for larger res. For regions of high population density, anthropogenic or structural factors are often dominant for re susceptibility. For example Molina et al. [2019] used MaxEnt to show distance to roads, settlements or powerlines were the dominant factors for re occurrence probability in the Andalusia region in southern Spain. MaxEnt has also been used for estimating spatial re probability under di erent scenarios such as future projections of housing development and private land conservation [Syphard et al., 2016]. One study in China using RF found mean spring temperature was the most important variable for re occurrence whereas forest stock was most important for area burned [Ying et al., 2018]. Some authors examined controls on re severity using high resolution data for a single large re. For example, several authors used RF to examine controls on burn severity for the 2013 Rim re in the Sierra Nevada [Lydersen et al., 2014, Kane et al., 2015, Lydersen et al., 2017]. At smaller spatial scales re 22 weather was the most important variable for re severity, whereas fuel treatments were most important at larger spatial scales [Lydersen et al., 2017]. A similar study by Harris and Taylor [2017] showed that previous re severity was an important factor in uencing re severity for the Rim re. For the 2005 Riba de Saelices re, Viedma et al. [2015] looked at factors contributing to burn severity using a BRT model and found burning conditions (including re weather variables) were more important compared than stand structure and topography. For burn severity these papers all used the Relativized di erenced Normalized Burn Ratio (RdNBR) metric, derived from Landsat satellite images, which allowed spatial modeling at high resolutions (eg. 30m by 30m). In addition to the more commonly used ML methods one paper by Wu et al. [2015] used KNN to identify spatially homogeneous re environment zones by clustering climate, vegetation, topography, and human activity related variables. They then used CART to examine variable importance for each of three re environment zones in south-eastern China. For landscape controls on re there were few studies comparing multiple ML methods. One such study by Nelson et al. [2017] compared CART, BRT and RF for classifying di erent re size classes in British Columbia, Canada. For both central and periphery regions they found the best performing model was BRT followed by CART and RF. For example, in the central region BRT achieved a classi cation accuracy of 88% compared with 82.9% and 49.6% for the CART and RF models respectively. It is not clear from the study why RF performed poorly, although it was noted that variable importance di ers appreciably between the three models. 4.4 Fire Behavior Prediction In general, re behavior includes physical processes and characteristics at a variety of scales including combustion rate, aming, smouldering residence time fuel consumption, ame height, and ame depth. However, the papers in this section deal mainly with larger scale processes and characteristics such as the prediction of re spread rates, re growth, burned area, and re severity, conditional on the occurrence (ignition) of one, or more, wild res. Here, our emphasis is on prognostic applications, in contrast to the Fuels Characterization, Fire Detection and Mapping problem domain, in which we focused on diagnostic applications. 4.4.1 Fire spread and growth Predicting the spread of a wildland re is an important task for re management agencies, particularly to aid in the deployment of suppression resources or to anticipate evacuations one or more days in advance. Thus, a large number of models have been developed using di erent approaches. In a series of reviews Sullivan [2009a,b,c] described re spread models he classi ed as being of physical or quasi-physical nature, or empirical or quasi-empirical nature, as well as mathematical analogues and simulation models. Many re growth simulation models convert one dimensional empirical or quasi-empirical spread rate models to two dimensions and then propagate a re perimeter across a modelled landscape. A wide range of ML methods have been applied to predict re growth. For example, Markuzon and Kolitz [2009] tested several classi ers (RF, BNs, and KNN) to estimate if a re would become large either one or two days following its observation; they found each of the tested methods performed similarly with RF correctly classifying large res at a rate over 75%, albeit with a number of false positives. Vakalis et al. [2004] used a ANN in combination with a fuzzy logic model to estimate the rate of spread in the mountainous region of Attica in Greece. A number of papers used genetic algorithms (GAs) to optimize input parameters to a physics or empirically based re simulator in order to improve re spread predictions [Abdalhaq et al., 2005, Rodriguez et al., 2008, Rodr guez et al., 2009, Art es et al., 2014, 2016, Carrillo et al., 2016, Denham et al., 2012, Cencerrado et al., 2012, 2013, 2014, Art es et al., 2017, Denham and Laneri, 2018]. For example, Cencerrado et al. [2014] developed a framework based on GAs to shorten the time needed to run deterministic re spread simulations. They tested the framework using the FARSITE [Finney, 2004] re spread simulator with di erent input scenarios sampled from distributions of vegetation models, wind speed/direction, and dead/live fuel moisture content. The algorithm used a tness function which discarded the most time-intensive simulations, but did not lead to an appreciable decrease in the 23 accuracy of the simulations. Such an approach is potentially useful for re management where it is desirable to predict re behavior as far in advance as possible so that the information can be enacted upon. This approach may greatly reduce overall simulation time by reducing the input parameter space as also noted by Art es et al. [2016] and Denham et al. [2012], or through parallelization of simulation runs for stochastic approaches [Art es et al., 2017, Denham and Laneri, 2018]. A di erent goal was considered by Ascoli et al. [2015] who used a GA to optimize fuel models in Southern Europe by calibrating the model with respect to rate of spread observations. Kozik et al. [2013] presented a re spread model that used a novel ANN implementation that incorpo- rated a Kalman lter for data assimilation that could potentially be run in real-time, the resulting model more closely resembling that of complex cellular automata than a traditional ANN. The same authors later implemented this model and simulated re growth under various scenarios with di erent wind speeds and directions, or both, although a direct comparison with real re data was not possible [Kozik et al., 2014]. Zheng et al. [2017] simulated re spread by integrating a cellular automata (CA) model with an Extreme Learning Machine (ELM; a type of feedforward ANN). Transition rules for the CA were determined by the ELM trained with data from historical res, as well as vegetation, topographic, and meteorological data. Likewise, Chetehouna et al. [2015] used ANNs to predict re behavior, including rate of spread, and ame height and angle. In contrast, Subramanian and Crowley [2017] formulated the problem of re spread prediction as a Markov Decision Process, where they proposed solutions based on both a classic reinforcement learning algorithm and a deep reinforcement learning algorithm { the authors found the deep learning approach improved on the traditional approach when tested on two large res in Alberta, Canada. The authors further developed this work to compare ve widely used reinforcement learning algorithms [Subramanian and Crowley, 2018], and found that the Asynchronous Advantage Actor-Critic (A3C) and Monte Carlo Tree Search (MCTS) algorithms achieved the best accuracy. Meanwhile, Khakzad [2019] developed a re spread model to predict the risk of re spread in Wildland-Industrial Interfaces, using Dynamic Bayesian Networks (DBN) in combination with a deterministic re spread model. The Canadian Fire Behavior Prediction (FBP) system, which uses meteorological and fuel conditions data as inputs, determined the re spread probabilities from one node to another in the aforementioned DBN. More recently Hodges and Lattimer [2019] trained a (deep learning) CNN to predict re spread using environmental variables (topography, weather and fuel related variables). Outputs of the CNN were spatial grids corresponding to the probability the burn map reached a pixel and the probability the burn map did not reach a pixel. Their method achieved a mean precision of 89% and mean sensitivity of 80% with reference 6 hourly burn maps computed using the physics-based FARSITE simulator. Radke et al. [2019] also used a similar approach to predict daily re spread for the 2016 Beaver Creek re in Colorado. 4.4.2 Burned area and re severity prediction There are a number of papers that focus on using ML approaches to directly predict the nal area burned from a wild re. Cortez and Morais [2007] compared multiple regression and four di erent ML methods (DT, RF, ANN, and SVM) to predict area burned using re and weather (i.e., temperature, precipitation, relative humidity and wind speed) data from the Montesinho natural park in northeastern Portugal, and found that SVM displayed the best performance. A number of publications subsequently used the data from Cortez and Morais [2007] to predict area burned using various ML methods, including ANN [Sa and Bouroumi, 2013, Storer and Green, 2016], genetic algorithms [Castelli et al., 2015], both ANN and SVM [Al Janabi et al., 2018], and decision trees [Alberg, 2015, Li et al., 2018a]. Notably, Castelli et al. [2015] found that a GA variant outperformed other ML methods including SVM. Xie and Shi [2014] used a similar set of input variables with SVM to predict burned area in for Guangzhou City in China. In addition to these studies, Toujani et al. [2018] used hidden Markov models (HMM) to predict burned area in the north- west of Tunisia, where the spatiotemporal factors used as inputs to the model were initially clustered using self-organizing maps (SOMs). Liang et al. [2019] compared back-propagation neural networks, recurrent neural networks (RNN) and Long Short Term Memory (LSTM) neural networks to predict wild re scale, a quantity related to area burned and re duration, in Alberta Canada. They found the highest accuracy 24 (90.9%) was achieved with LSTM. Most recently, Xie and Peng [2019] compared a number of machine learning methods for estimating area burned (regression) and binary classi cation of re sizes (> 5 Ha) in Montesinho natural park, Portugal. For the regression task, they found a tuned RF algorithm performed better than standard RF, tuned and standard gradient boosted machines, tuned and standard generalized linear models (GLMs) and deep learning. For the classi cation problem they found extreme gradient boosting and deep learning had a higher accuracy than CART, RF, SVM, ANN, and logistic regression. By attempting to predict membership of burned area size classes, a number of papers were able to recast the problem of burned area prediction as a classi cation problem. For example, Yu et al. [2011] used a combination of SOMs and back-propagation ANNs to classify forest res into size categories based on meteorological variables. This approach gave Yu et al. [2011] better accurary ( 90%) when compared with a rules-based method ( 82%). Ozbayo glu and Bozer [2012] estimated burned area size classes us- ing geographical and meteorological data using three di erent machine learning methods: i) Multilayer Perceptron (MLP); ii) Radial Basis Function Networks (RBFN); and iii) SVM. Overall, the best perform- ing method was MLP, which achieved a 65% success rate, using humidity and windspeed as predictors. Zwirglmaier et al. [2013] used a BN to predict area burned classes using historical re data, re weather data, re behaviour indices, land cover, and topographic data. Shidik and Mustofa [2014] used a hybrid model (Fuzzy C-Means and Back-Propagation ANN) to estimate re size classes using data from Cortez and Morais [2007], where the hybrid model performed best with an accuracy of 97.50% when compared with Naive Bayes (55.5%), DT (86.5%), RF (73.1%), KNN (85.5%) and SVM (90.3%). Mitsopoulos and Mallinis [2017] compared BRT, RF and Logistic Regression to predict 3 burned area classes for res in Greece. They found RF led to the best performance of the three tested methods and that re suppression and weather were the two most important explanatory variables. Coeld et al. [2019] compared CART, RF, ANN, KNN and gradient boosting to predict 3 burned area classes at time of ignition in Alaska. They found a parsimonious model using CART with Vapor Pressure De cit (VPD) provided the best performance of the models and variables considered. We found only one study that used ML to predict re behavior related to re severity, which is important in the context of re ecology, suggesting that there are opportunities to apply ML in this domain of wild re science. In that paper, Zald and Dunn [2018] used RF to determine that the most important predictor of re severity was daily re weather, followed by stand age and ownership, with less predictability given by topographic features. 4.5 Fire E ects Fire E ects prediction studies have largely used regression based approaches to relate costs, losses, or other impacts (e.g., soils, post- re ecology, wildlife, socioeconomic factors) to physical measures of re severity and exposure. Importantly, this category also includes wild re smoke and particulate modelling (but not smoke detection which was previously discussed in the re detection section). 4.5.1 Soil Erosion and Deposits Mallinis et al. [2009] modelled potential post- re soil erosion risk following a large intensive wild re in the Mediterranean area using CART and k-means algorithms. In that paper, before wild re, 55% of the study area was classi ed as having severe or heavy erosion potential, compared to 90% post- re, with an overall classi cation accuracy of 86%. Meanwhile, Buckland et al. [2019] used ANNs to examine the relationships between sand deposition in semi-arid grasslands and wild re occurrence, land use, and climatic conditions. The authors then predicted soil erosion levels in the future given climate change assumptions. 25 4.5.2 Smoke and Particulate Levels Smoke emitted from wild res can seriously lower air quality with adverse e ects on the health of both human and non-human animals, as well as other impacts. Thus, it is not surprising that ML methods have been used to understand the dynamics of smoke from wildland re. For example, Yao et al. [2018b] used RF to predict the minimum height of forest re smoke using data from the CALIPSO satellite. More commonly, ML methods have also been used to estimate population exposure to ne particulate matter (e.g., PM2.5: atmospheric particulate matter with diameter less than 2.5m), which can be useful for epidemiological studies and for informing public health actions. One such study by Yao et al. [2018a] also used RF to estimate hourly concentrations of PM2.5 in British Columbia, Canada. Zou et al. [2019] compared RF, BRT and MLR to estimate regional PM2.5 concentrations in the Paci c Northwest and found RF performed much better than the other algorithms. In another very broad study covering several datasets and ML methods, Reid et al. [2015] estimated spatial distributions of PM2.5 concentrations during the 2008 northern California wild res. The authors of the aforementioned study used 29 predictor variables and compared 11 di erent statistical models, including RF, BRT, SVM, and KNN. Overall, the BRT and RF models displayed the best performance. Emissions other than particulate matter have also been modelled using ML, as Lozhkin et al. [2016] used an ANN to predict carbon monoxide concentrations emitted from a peat re in Siberia, Russia. In another study, the authors used ten di erent statistical and ML methods and 21 covariates (including weather, geography, land-use, and atmospheric chemistry) to predict ozone exposures before and after wild re events [Watson et al., 2019]. Here, gradient boosting gave the best results with respect to both root mean square error and R values, followed by RF and SVM. In a di erent application related to smoke, Fuentes et al. [2019] used ANNs to detect smoke in several di erent grape varietals used for wine making. 4.5.3 Post- re regeneration, succession, and ecology The study of post- re regeneration is an important aspect of understanding forest and ecosystem responses and resilience to wild re disturbances, with important ecological and economic consequences. RF, for example, has been a popular ML method for understanding the important variables driving post- re regeneration [Jo~ ao et al., 2018, Vijayakumar et al., 2016]. Burn severity (a measure of above and below ground biomass loss due to re) is an important metric for understanding the impacts of wild re on vegetation and post- re regeneration, soils, and potential successional shifts in forest composition, and as such, has been included in many ML studies in this section, including [Barrett et al., 2011, Cai et al., 2013, Cardil et al., 2019, Chapin et al., 2014, Divya and Vijayalakshmi, 2016, Fairman et al., 2017, Han et al., 2015, Johnstone et al., 2010, Liu and Yang, 2014, Mart n-Alc on and Coll, 2016, Sherrill and Romme, 2012, Thompson and Spies, 2010]. For instance, Cardil et al. [2019] used BRT to demonstrate that remotely- sensed data (i.e., Relative Di erenced Normalized Burn Ratio index; RdNBR) can provide an acceptable assessment of re-induced impacts (i.e., burn severity) on forest vegetation, while [Fairman et al., 2017] used RF to identify the variables most important in explaining plot-level mortality and regeneration of Eucalyptus pauci ora in Victoria, Australia, a ected by high-severity wild res and subsequent re-burns. Debouk et al. [2013] assessed post- re vegetation regeneration status using eld measurements, a canopy height model, and Lidar (i.e., 3D laser scanning) data with a simple ANN. Post- re regeneration also has important implications for the successional trajectories of forested areas, and a few studies have examined this using ML approaches [Barrett et al., 2011, Cai et al., 2013, Johnstone et al., 2010]. For example, Barrett et al. [2011] used RF to model re severity, from which they made an assessment of the area susceptible to a shift from coniferous to deciduous forest cover in the Alaskan boreal forest, while Cai et al. [2013] used BRT to assess the in uence of environmental variables and burn severity on the composition and density of post- re tree recruitment, and thus the trajectory of succession, in northeastern China. In other studies not directly related to post- re regeneration, Hermosilla et al. [2015] used RF to attribute annual forest change to one of four categories, including wild re, in Saskatchewan, Canada, while [Jung et al., 2013] used GA and RF to estimate the basal area of post- re residual spruce (Picea obovate ) and r 26 (Abies sibirica ) stands in central Siberia using remotely sensed data. Magadzire et al. [2019] used MaxEnt to demonstrate that re return interval and species life history traits a ected the distribution of plant species in South Africa. ML has also been used to examine re e ects on the hydrological cycle, as Poon et al. [2018] used SVM to estimate both pre- and post-wild re evapotranspiration using remotely sensed variables. Considering the potential impacts of wild res on wildlife, it is perhaps surprising that relatively few of such studies have adopted ML approaches. However, ML methods have been used to predict the impacts of wild re and other drivers on species distributions and arthropod communities. Hradsky et al. [2017], for example, used non-parametric BNs to describe and quantify the drivers of faunal distributions in wild re- a ected landscapes in southeastern Australia. Similarly, Reside et al. [2012] used MaxEnt to model bird species distributions in response to re regime shifts in northern Australia, which is an important aspect of conservation planning in the region. ML has also been used to look at the e ects of wild re on fauna at the community level, as Luo et al. [2017] used DTs, Association Rule Mining, and AdaBoost to examine the e ects of re disturbance on spider communities in Cangshan Mountain, China. 4.5.4 Socioeconomic e ects ML methods have been little used to model socio-economic impacts of re to date. We found one study in which BNs were used to predict the economic impacts of wild res in Greece from 2006-2010 due to housing losses [Papakosta et al., 2017]. The authors did this by rst de ning a causal relationship between the participating variables, and then using BNs to estimate housing damages. It is worth noting that the problem of detecting these causal relationships from data is a dicult task and remains an active area of research in arti cial intelligence. 4.6 Fire management The goal of contemporary re management is to have the appropriate amount of re on the landscape, which may be accomplished through the management of vegetation including prescribed burning, the management of human activities (prevention), and re suppression. Fire management is a form of risk management that seeks to maximize re bene ts and minimize costs and losses [Finney, 2005]. Fire management decisions have a wide range of scales, including long-term strategic decisions about the acquisition and location of resources or the application of vegetation management in large regions, medium-term tactical decisions about the acquisition of additional resources, relocation, or release of resources during the re season, and short-term real time operational decisions about the deployment and utilization of resources on individual incidents. Fire preparedness and response is a supply chain with a hierarchical dependence. Taylor [2020] describes 20 common decision types in re management and maps the spatial-temporal dimensions of their decision spaces. Fire management models can be predictive, such as the probability of initial attack success, or pre- scriptive such as to maximize/minimize an objective function (e.g., optimal helicopter routing to minimize travel time in crew deployment). While advances have been made in the domain of wild re management using ML techniques, there have been relatively few studies in this area compared to other wild re problem domains. Thus, there appears to be great potential for ML to be applied to wild re management problems, which may lead to novel and innovative approaches in the future. 4.6.1 Planning and policy An important area of re management is planning and policy, where various ML methods have been applied to address pertinent challenges. For example, Bao et al. [2015] used GA, which are useful for solving multi-objective optimization problems, to optimize watchtower locations for forest re monitoring. Bradley et al. [2016] used RF to investigate the relationship between the protected status of forest in the western US and burn severity. Likewise, Ru ault and Mouillot [2015] also used BRTs to assess the impact 27 of re policy introduced in the 1980s on re activity in southern France and the relationships between re and weather, and Penman et al. [2011] used BNs to build a framework to simultaneously assess the relative merits of multiple management strategies in Wollemi National Park, NSW, Australia. McGregor et al. [2016] used Markov decision processes (MDP) and model free Monte Carlo method to create fast running simulations (based on the FARSITE simulator) to create interactive visualizations of forest futures over 100 years based on alternate high-level suppression policies. McGregor et al. [2017] demonstrated ways in which a variety of ML and optimization methods can be used to create an interactive approximate simulation tool for re managers. The authors of the aforementioned study utilized a modi ed version of the FARSITE re-spread simulator, which was augmented to run thousands of simulation trajectories while also including new models of lightning strike occurrences, re duration, and a forest vegetation simulator. McGregor et al. [2017] also clearly show how decision trees can be used to analyze a hierarchy of decision thresholds for deciding whether to suppress a re or not; their hierarchy splits on fuel levels, then intensity estimations, and nally weather predictors to arrive at a generalizable policy. 4.6.2 Fuel treatment ML methods have also been used to model the e ects of fuel treatments in order to mitigate wild re risk. For example, Penman et al. [2014] used a BN to examine the relative risk reduction of using prescribed burns on the landscape versus within the 500m interface zone adjacent to houses in the Sydney basin, Australia. Lauer et al. [2017] used approximate dynamic programming (also known as reinforcement learning) to determine the optimal timing and location of fuel treatments and timber harvest for a re-threatened landscape in Oregon, USA, with the objective of maximizing wealth through timber management. Similarly, Arca et al. [2015] used GA for multi-objective optimization of fuel treatments. 4.6.3 Wild re preparedness and response Wild re preparedness and response issues have also been examined using ML techniques. Costafreda- Aumedes et al. [2015] used ANNs to model the relationships between daily re load, re duration, re type, re size, and response time, as well as personnel and terrestrial/aerial units deployed for individual wild res in Spain. Most of the models in Costafreda-Aumedes et al. [2015] highlighted the positive correlation of burned area and re duration with the number of resources assigned to each re, and some highlighted the negative in uence of daily re load. In another study, Penman et al. [2015] used Bayesian Networks to assess the relative in uence of preventative and suppression management strategies on the probability of house loss in the Sydney basin, Australia. O'Connor et al. [2017] used BRT to develop a predictive model of re control locations in the Northern Rocky Mountains, USA, based on the likelihood of nal re perimeters, while Homchaudhuri et al. [2010] used GAs to optimize reline generation. Rodrigues et al. [2019] modelled the probability that wild re will escape initial attack using a RF model trained with re location, detection time, arrival time, weather, fuel types, and available resources data. Important variables in Rodrigues et al. [2019] included re weather and simultaneity of events. Julian and Kochenderfer [2018a] used two di erent RL algorithms to develop a system for autonomous control of one or more aircraft in order to monitor active wild res. 4.6.4 Social factors Recently, the use of ML in re management has grown to encompass more novel aspects of re management, even including the investigation of criminal motives related to arson, as Delgado et al. [2018] used BNs to characterize wild re arsonists in Spain thereby identifying ve motivational archetypes (i.e., slight negligence; gross negligence; impulsive; pro t; and revenge). 28 5 Discussion ML methods have seen a spectacular evolution in development, accuracy, computational eciency, and application in many elds since the 1990s. It is therefore not surprising that ML has been helpful in providing new insights into several critical sustainability and social challenges in the 21st century [Gomes, 2009, Sullivan et al., 2014, Butler, 2017]. The recent uptake and success of ML methods has been driven in large part by ongoing advances in computational power and technology. For example, the recent use of bandwidth optimized Graphics Processing Units (GPUs) takes advantage of parallel processing for simul- taneous execution of computationally expensive tasks, which has facilitated a wider use of computationally demanding but more accurate methods like DNNs. The advantages of powerful but ecient ML methods are therefore widely anticipated as being useful in wild re science and management. However, despite some early papers suggesting that data driven techniques would be useful in forest re management [Latham, 1987, Kourtz, 1990, 1993], our review has shown that there was relatively slow adoption of ML-based research in wild re science up to the 2000s compared with other elds, followed by a sharp increase in publication rate in the last decade. In the early 2000s, data mining techniques were quite popular and classic ML methods such as DTs, RF, and bagging and boosting techniques began to appear in the wild re science literature (e.g., Stojanova et al. [2006]). In fact, some researchers started using simple feed forward ANNs for small scale applications as early as the mid 1990s and early 2000s (e.g., Mccormick et al. [1999], Al-Rawi et al. [2002]). In the last three decades, almost all major ML methods have been used in some way in wild re applications, although some more computationally demanding methods, such as SOMs and cellular automatons, have only been actively experimented with in the last decade [Toujani et al., 2018, Zheng et al., 2017]. Furthermore, the recent development of DL algorithms, with a particular focus on extracting spatial features from images, has led to a sharp rise in the application of DL for wild re applications in the last decade. It is evident, however, from our review that while an increasing number of ML methodologies have been used across a variety of re research domains over the past 30 years, this research is unevenly distributed among ML algorithms, research domains and tasks, and has had limited application in re management. Many re science and management questions can be framed within a re risk context. Xi et al. [2019] discussed the advantages of adopting a risk framework with regard to statistical modeling of wild res. There the risk components of \hazard", \vulnerability" and \exposure" are replaced respectively by re probability, re behavior and re e ects. Most re management activities can be framed as risk controls to mitigate these components of risk. Traditionally, methods used in wild re re science to address these various questions have included physical modeling (e.g., Sullivan [2009a,b,c]), statistical methods (e.g., Taylor et al. [2013], Xi et al. [2019]), simulation modeling (e.g., Keane et al. [2004]), and operations research methods (Martell [2015], Minas et al. [2012]). In simple terms, any analytical study begins with one or more of four questions: \what happened?"; \why did it happen?"; \what will happen?"; or \what to do?" Corresponding data driven approaches to address these questions are respectively called descriptive, diagnostic, predictive, and prescriptive analyt- ics. The type of analytical approach adopted then circumscribes the types of methodological approaches (e.g., regression, classi cation, clustering, dimensionality reduction, decision making) and sets of possible algorithms appropriate to the analysis. In our review, we found that studies incorporating ML methods in wildland re science were predomi- nantly associated with descriptive or diagnostic analytics, re ecting the large body of work on re detection and mapping using classi cation methods, and on re susceptibility mapping and landscape controls on re using regression approaches. In many cases, the ML methods identi ed in our review are an alternative to statistical methods used for clustering and regression. While the aforementioned tasks are undoubtedly very important for understanding wildland re, we found much less work associated with predictive or pre- scriptive analytics, such as re occurrence prediction (predictive), re behaviour prediction (predictive), and re management (prescriptive). This may be because: a) particular domain knowledge is required to frame re management problems; b) re management data are often not publicly available, need a lot of 29 work to transform into an easily analyzable form, or do not exist at the scale of the problem; and c) some re management problems are not suited or can't be fully addressed by ML approaches. We note that much of the work on re risk in the re susceptibility and mapping domain used historical re and environmental data to map re susceptibility; therefore, while that work aims to inform future re risk, it cannot be considered to be predictive analytics, except, for example, in cases where it was used in combination with climate change projections. It appears then that, in general, wild re science research is currently more closely aligned with descriptive and diagnostic analytics, whereas wild re management goals are aligned with predictive and prescriptive analytics. This fundamental di erence identi es new opportunities for research in re management, which we discuss later in this paper. In the remainder of the paper, we examine some considerations for the use of ML methods, including: data considerations, model selection and accuracy, implementation challenges, interpretation, opportuni- ties, and implications for re management. 5.1 Data considerations ML is a data-centric modeling paradigm concerned with nding patterns in data. Importantly, data scientists need to determine, often in collaboration with re managers or domain experts, whether there are suitable and sucient data for a given modeling task. Some of the criteria for suitable data include whether: a) the predictands and covariates are or can be wrangled into the same temporal and spatial scale; b) the observations are a representative sample of the full range of conditions that may occur in application of a model to future observations; and c) whether the data are at spatiotemporal scale appropriate to the re science or management question. The rst of these criteria can be relaxed in some ML models such as ANNs and DNNs, where inputs and outputs can be at di erent spatial or temporal scales for appropriately designed network architectures, although data normalization may still be required. The second criterion also addresses the important question of whether enough data exists for training a given algorithm for a given problem. In general, this question depends on the nature of the problem, complexity of the underlying model, data uncertainty and many other factors (see Roh et al. [2018] for a further discussion of data requirements for ML). In any case, many complex problems require a substantive data wrangling e ort, to acquire, perform quality assurance, and fuse data into sampling units at the appropriate spatiotemporal scale. An example of this in daily re occurrence prediction, where observations of a variety of features (e.g., continuous measures such as re arrival time and location, or lightning strike times and locations) are discretized into three-dimensional (e.g., longitude, latitude, and day) cells called voxels. Another important consideration for the collection and use of data in machine learning is selection bias. A form of spatial selection bias called preferential sampling occurs when sampling occurs preferentially in locations where one expects a certain response [Diggle et al., 2010]. For example, preferential sampling may occur in air monitoring, because sensors may be placed in locations where poor air quality is expected [Shaddick and Zidek, 2014]. In general, preferential sampling or other selection biases may be avoided altogether by selecting an appropriate sampling strategy at the experimental design phase, or, where this is not possible, to take it into account in model evaluation [Zadrozny, 2004]. For the problem domain re detection and mapping, most applications of ML used some form of im- agery (e.g., remote sensed satellite images or terrestrial photographs). In particular, many papers used satellite data (e.g., Landsat, MODIS) to determine vegetation di erences before and after a re and so were able to map area burned. For re detection, many applications considered either remote sensed data for hotspot or smoke detection, or photographs of wild res (used as inputs to an image classi cation problem). For re weather and climate change, the three main sources of data were either weather station observa- tions, climate reanalyses (modelled data that include historical observations), or GCMs for future climate projections. Reanalyses and GCMs are typically highly dimensional large gridded spatiotemporal datasets which require careful feature selection and/or dimensional reduction for ML applications. Fire occurrence prediction, susceptibility, and risk applications used a large number of di erent environmental variables as predictors, but almost all used re locations and associated temporal information as predictands. Fire data itself is usually collated from re management agencies in the form of georeferenced points or perimeter 30 data, along with reported dates, ignition cause, and other related variables. Care should be taken using such data because changes in reporting standards or accuracy may lead to data inhomogeneity. As well as re locations and perimeters, re severity is an attribute of much interest to re scientists. Fire severity is often determined from remotely sensed data and represented using variables such as the Di erenced Nor- malized Burn Ratio (dNBR) and variants, or through eld sampling. However, remote sensed estimates of burn severity should be considered as proxies as they have low skill in some ecosystems. Other re ecology research historically relies on in situ eld, sampling although many of the ML applications attempt to resolve features of interest using remote sensed data. Smoke data can also be derived from remote sensed imagery or from air quality sensors (e.g., PM2.5, atmospheric particulate matter less than 2.5 m). Continued advances in remote sensing, as well as the quality and availability of remote sensed data prod- ucts, in weather and climate modeling have led to increased availability of large spatiotemporal datasets, which presents both an opportunity and challenge for the application of ML methods in wild re research and management. The era of \big data" has seen the development of cloud computing platforms to provide the computing and data storage facilities to deal with these large datasets. For example, in our review we found two papers [Crowley et al., 2019, Quintero et al., 2019] that used Google Earth Engine which inte- grates geospatial datasets with a coding environment [Gorelick et al., 2017]. In any case, data processing and management plays an important role in the use of large geospatial datasets. 5.2 Model selection and accuracy Given a wild re science question or management problem and available relevant data, a critical question to ask is what is the most appropriate modeling tool to address the problem? Is it a standard statistical model (e.g., linear regression or LR), a physical model (e.g., FIRETEC or other re simulator), a ML model, or a combination of approaches? Moreover, which speci c algorithm will yield the most accurate classi cation or regression. Given the heterogeneity of research questions, study areas, and datasets considered in the papers reviewed here, it is not possible to comprehensively answer these questions with respect to ML approaches. Even in the case where multiple studies used the same dataset [Cortez and Morais, 2007, Sa and Bouroumi, 2013, Storer and Green, 2016, Castelli et al., 2015, Al Janabi et al., 2018, Alberg, 2015, Li et al., 2018a, Castelli et al., 2015] the di erent research questions considered meant a direct comparison of ML methods was not possible between research studies. However, a number of individual studies did make comparisons between multiple ML methods, or between ML and statistical methods for a given wild re modeling problem and dataset. Here we highlight some of their ndings to provide some guidance with respect to model selection. In our review (see section 4 and the supplementary material), we found 29 papers comparing ML and statistical methods, where in the majority of these cases ML methods were found to be more accurate than traditional statistical methods (e.g., GLMs), or displayed similar performance [Pu and Gong, 2004, Bates et al., 2017, de Bem et al., 2018]. In only one study on climate change by Amatulli et al. [2013], MARS was found to be superior to RF for their analytical task. A sizable number of the comparative studies (14) involved classi cation problems that used LR as a benchmark method against ANN or ensemble tree methods. For studies comparing multiple ML methods, there was considerable variation in the choice of most accurate method; however, in general ensemble methods tended to outperform single classi er methods (e.g., Stojanova et al. [2012], Dutta et al. [2016], Mayr et al. [2018], Nelson et al. [2017], Reid et al. [2015], Watson et al. [2019]), except in one case where the most accurate model (CART) was also the most parsimonious [Coeld et al., 2019]. A few more recent papers also highlighted the advantages of DL over other methods. In particular, for re detection, Zhang et al. [2018b] compared CNNs with SVM and found that CNNs were more accurate, while Zhao et al. [2018] similarly found CNNs superior to SVMs and ANNs. For re susceptibility mapping, Zhang et al. [2019] found CNNs were more accurate than RF, SVMs, and ANNs. For time series forecasting problems, Liang et al. [2019] found LSTMs outperformed ANNs. Finally, Cao et al. [2019] found that using an LSTM combined with a CNN led to better re detection performance from video compared with CNNs alone. In any case, more rigorous inter-model comparisons are needed to reveal in which conditions, and in what sense particular methods are more accurate, as well as to establish procedures for evaluating accuracy. 31 ML methods are also prone to over tting, so it is important to evaluate models with robust test datasets using appropriate cross-validation strategies. For example, the na ve application of cross-validation to data that have spatial or spatio-temporal dependencies may lead to overly optimistic evaluations [Roberts et al., 2017]. In general, one also desires to minimise errors associated with either under-speci cation or over- speci cation of the model, a problem known as the bias-variance trade-o [Geman et al., 1992]. However, several recent advances have been made to reduce over tting in ML models, for instance, regularization techniques in DNNs [Kuka cka et al., 2017]. Moreover, when interpreting comparisons between ML and statistical methods, we should be cognizant that just as some ML methods require expert knowledge, the accuracy of statistical methods can also vary with the skill of the practitioner. Thompson and Calkin [2011] also emphasize the need for identifying sources of uncertainty in modeling so that they can better managed. 5.3 Implementation Challenges Beyond data and model selection, two important considerations for model speci cation are feature selection and spatial autocorrelation. Knowledge of the problem domain is extremely important in identifying a set of candidate features. However, while many ML methods are not limited by the number of features, more variables do not necessarily make for a more accurate, interpretable, or easily implemented model [Schoenberg, 2016, Breiman, 2001] and can lead to over tting and increased computational time. Two di erent ML methods to enable selection of a reduced and more optimal set of features include GAs and PSO. Sachdeva et al. [2018] used a GA to select input features for BRT and found this method gave the best accuracy compared with ANN, RF, SVM, SVM with PSO (PSO-SVM), DTs, logistic regression, and NB. Hong et al. [2018] employed a similar approach for re susceptibility mapping and found this led to improvements for both SVM and RF compared with their non-optimized counterparts. Tracy et al. [2018] used a novel random subset feature selection algorithm for feature selection, which they found led to higher AUC values and lower model complexity. Jaafari et al. [2019] used a NFM combined with the imperialist competitive algorithm (a variant of GA) for feature selection which led to very high model accuracy (0.99) in their study. Tien Bui et al. [2017] used PSO to choose inputs to a NFN and found this improved results. [Zhang et al., 2019] also considered the information gain ratio for feature selection. As noted in Moritz et al. [2012] and Mayr et al. [2018], one should also take spatial autocorrelation into account when modeling re probabilities spatially. In general, the presence of spatial autocorrelation violates the assumption of independence for parametric models, which can degrade model performance. One approach to deal with autocorrelation requires subsampling to remove any spatial autocorrelation Moritz et al. [2012]. It is also often necessary to subsample from non- re locations due to class imbalance between ignitions and non- ignitions (e.g., Cao et al. [2017], Zhang et al. [2019]). Song et al. [2017] considered spatial econometric models and found a spatial autocorrelation model worked better than RF, although Kim et al. [2019] note that RF may be robust to spatial autocorrelation with large samples. In contrast to many ML methods, a strength of CNNs is its ability to exploit spatial correlation in the data to enable the extraction of spatial features. 5.4 Interpretation A major obstacle for the adoption of ML methods to re modeling tasks is the perceived lack of inter- pretability or explainability of such methods, which are often considered to be \black box" models. Users (in this case re ghters and managers) need to trust ML model predictions, and so have the con dence and justi cation to apply these models, particularly in cases where proposed solutions are considered novel. Model intepretability should therefore be an important aspect of model development if models are to be selected and deployed in re management operations. Model interpretability varies signi cantly across the di erent types of ML. For example, conventional thinking is that tree-based methods are more inter- pretable than neural network methods. This is because a single decision tree classi er can be rendered as a ow chart corresponding to if-then-else statements, whereas an ANN represents a nonlinear function 32 approximated through a series of nonlinear activations. However, because they combine multiple trees in an optimized way, ensemble tree classi ers are less interpretable than single tree classi ers. On the other hand, BNs are one example of an ML technique where good explanations for results can be inferred due to their graphical representation; however, full Bayesian learning on large-scale data is very computation- ally expensive which may have limited early applications; however, as computational power has increased we have seen an increase in the popularity of BNs in wild re science and management applications (e.g., Penman et al. [2015], Papakosta et al. [2017]). DL-based architectures are widely considered to be among the least interpretable ML models, despite the fact that they can achieve very accurate function approximation [Chakraborty et al., 2017]. In fact, this is demonstrative of the well-known trade-o between prediction accuracy and interpretability (see Kuhn and Johnson [2013] for an in-depth discussion). The ML community, however, recognizes the problem of interpretability and work is underway to develop methods that allow for greater interpretability of ML methods, including methods for DL (see for example, McGovern et al. [2019]) or model-agnostic approaches [Ribeiro et al., 2016]. Runge et al. [2019] further argue that casual inference methods should be used in conjunction with predictive models to improve our understanding of physical systems. Finally, it is worth noting that assessing variable importance (see Sec. 4.3.4) for a given model can play a role in model interpretation. 5.5 Opportunities Our review highlights a number of potential opportunities in wild re science and management for ML applications where ML has not yet been applied or is under-utilized. Here we examine ML advances in other areas of environmental science that have analogous problems in wildland re science and which may be useful for identifying further ML applications. For instance, Li et al. [2011] compared ML algorithms for spatial interpolation and found that a RF model combined with geostatistical methods yielded good results; a similar method could be used to improve interpolation of re weather observations from weather stations, and so enhance re danger monitoring. Rasp and Lerch [2018] showed that ANNs could improve weather forecasts by post-processing ensemble forecasts, an approach which could similarly be applied to improve short-term forecasts of re weather. Belayneh et al. [2014] used ANNs and SVMs combined with wavelet transforms for long term drought forecasting in Ethiopia; such methods could also be useful for forecasting drought in the context of re danger potential. In the context of numerical weather prediction, Cohen et al. [2019] found better predictability using ML methods than dynamical models for subseasonal to seasonal weather forecasting, suggesting similar applications for long-term re weather forecasting. McGovern et al. [2017] discussed how AI techniques can be leveraged to improve decision making around high-impact weather. More recently, Reichstein et al. [2019] have further argued for the use of DL in the environmental sciences, citing its potential to extract spatiotemporal features from large geospatial datasets. Kussul et al. [2017] used CNNs to classify land cover and crop types and found that CNNs improved the results over standard ANN models; a similar approach could be used for fuels classi cation, which is an important input to re behaviour prediction models. Shi et al. [2016] also used CNNs to detect clouds in remote sensed imagery and were able to di erentiate between thin and thick cloud. A similar approach could be used for smoke detection, which is important for re detection, as well as in determining the presence of false negatives in hotspot data (due to smoke or cloud obscuration). Finally, recent proposals have called for hybrid models that combine process-based models and ML methods [Reichstein et al., 2019]. For example, ML models may replace user-speci ed parameterizations in numerical weather prediction models [Brenowitz and Bretherton, 2018]. Other recent approaches use ML methods to determine the solutions to nonlinear partial di erential equations Raissi and Karniadakis [2018], Raissi et al. [2019]. Such methods could nd future applications in improving re behaviour prediction models based on computationally expensive physics-based re simulators, in coupled re-atmosphere models, or in smoke dispersion modeling. In any case, the applications of ML that we have outlined are meant for illustrative purposes and are not meant to represent an exhaustive list of all possible applications. 33 5.6 Implications for re management We believe ML has been under-utilized in re management, particularly with respect to problems belonging to either predictive or prescriptive analytics. Fire management comprises a set of risk control measures, which are often cast in the framework of the emergency response phases: prevention; mitigation; prepared- ness; response; recovery; and review [Tymstra et al., 2019]. In terms of nancial expenditure, by far the largest percentage spent in the response phase [Stocks and Martell, 2016]. In practice, re management is largely determined by the need to manage resources in response to active or expected wild res, typically for lead times of days to weeks, or to manage vegetative fuels. This suggests the opportunity for increased research in areas of re weather prediction, re occurrence prediction, and re behaviour prediction, as well as optimizing re operations and fuel treatments. The identi cation of these areas, as well as the fact that wild re is both a spatial and temporal process, further reiterate the need for ML applications for time series forecasting. From this review, there were few papers that used time series ML methods for forecasting problems, suggesting an opportunity for further work in this area. In particular, recurrent neural networks (RNNs) were used for re behavior prediction [Cheng and Wang, 2008, Kozik et al., 2013, 2014] and re occurrence prediction [Dutta et al., 2013]. The most common variant of RNNs are Long Short Term Memory (LSTM) networks [Hochreiter and Schmidhuber, 1997], which have been used for burned area prediction [Liang et al., 2019] and re detection [Cao et al., 2019]. Because these methods implicitly model dynamical processes, they should lead to improve forecasting models compared with standard ANNs. For example Gensler et al. [2017] have used LSTMs to forecast solar power and Kim et al. [2017] used CNNs combined with LSTM for forecasting precipitation. We anticipate that these methods could also be employed for re weather, re occurrence, and re behaviour prediction. We note that there are a number of operational research and management science methods used in re management research including queuing, optimization, and simulation of complex system dynamics (e.g., Martell [2015]) where ML algorithms don't seem to provide an obvious alternative. For example, planning models to simulate the interactions between re management resource con gurations and re dynamics reviewed by [Mavsar et al., 2013]. From our review, a few papers used agent-based learning methods for re management. In particular, reinforcement learning was used for optimizing fuel treatments [Lauer et al., 2017] or for autonomous control of aircraft for re monitoring [Julian and Kochenderfer, 2018a]. GAs were used for generating optimal relines for active res [Homchaudhuri et al., 2010] and for reducing the time for re simulation [Cencerrado et al., 2014]. However, more work is needed to identify where ML methods could contribute to tactical, operational, or strategic re management decision making. An important challenge for the re research and management communities is enabling the transition of potentially useful ML models to re management operations. Although we identi ed several papers that emphasized their ML models could be deployed in re management operations [Art es et al., 2016, Alonso-Betanzos et al., 2002, Iliadis, 2005, Stojanova et al., 2012, Davis et al., 1989, 1986, Liu et al., 2015], it can be dicult to assess whether and how a study has been adopted by, or in uenced, re management agencies. This challenge is often exacerbated by a lack of resources and/or funding, as well as the di erent priorities and institutional cultures of researchers and re managers. One possible solution to this problem would be the formation of working groups dedicated to enabling this transition, preferably at the research proposal phase. In general, enabling operational ML methods will require tighter integration and greater collaboration between the research and management communities, particularly with regards to project design, data compilation and variable selection, implementation, and interpretation. However, it is worth noting that this is not a problem unique to ML, it is a long-standing and common issue in many areas of re research and other applied science disciplines, where continuous e ort is required to maintain communications and relationships between researchers and practitioners. Finally, we would like to stress that we believe the wild re research and management communities should play an active role in providing relevant, high quality, and freely available wild re data for use by practitioners of ML methods. For example, burned area and re weather data made available by Cortez and Morais [2007] was subsequently used by a number of authors in their work. It is imperative that the 34 quality of data collected by management agencies be as robust as possible, as the results of any modelling process are dependent upon the data used for analysis. It is worth considering how new data on, for example, hourly re growth or the daily use of re management resources, could be used in ML methods to yield better predictions or management recommendations | using new tools to answer new questions may require better or more complete data. Conversely, we must recognize that despite ML models being able to learn on their own, expertise in wild re science is necessary to ensure realistic modelling of wild re processes, while the complexity of some ML methods (e.g., DL) requires a dedicated and sophisticated knowledge of their application (we note that many of the most popular ML methods used in this study are fairly easy to implement, such as RF, MaxEnt, and DTs). The observation that no single ML algorithm is superior for all classes of problem, an idea encapsulated by the \no free lunch" theorem [Wolpert, 1996], further reinforces the need for domain-speci c knowledge. Thus, the proper implementation of ML in wild re science is a challenging endeavor, often requiring multidisciplinary teams and/or interdisciplinary specialists to e ectively produce meaningful results. 5.7 A word of caution ML holds tremendous potential for a number of wild re science and management problem domains. As indicated in this review, much work has already been undertaken in a number of areas, although further work is clearly needed for re management speci c problems. Despite this potential, ML should not be considered a panacea for all re research areas. ML is best suited to problems where there is sucient high- quality data, and this is not always the case. For example, for problems related to re management policy, data is needed at large spatiotemporal scales (i.e., ecosystem/administrative spatial units at timescales of decades or even centuries), and such data may simply not yet exist in current inventories. At the other extreme, data is needed at very ne spatiotemporal scales for re spread and behavior modeling, including high resolution fuel maps and surface weather variables which are often not available at the required scale and are dicult to acquire even in an experimental context. Another limitation of ML may occur when one attempts make predictions where no analog exists in the observed data, such as may be the case with climate change prediction. 6 Conclusions Our review shows that the application of ML methods in wild re science and management has been steadily increasing since their rst use in the 1990s, across core problem domains using a wide range ML methods. The bulk of work undertaken thus far has used traditional methods such as RF, BRT, MaxEnt, SVM and ANNs, partly due to the ease of application and partly due to their simple interpretability in many cases. However, problem domains associated with predictive (e.g., predicted re behavior) or prescriptive analytics (e.g. optimizing re management decisions) have seen much less work with ML methods. We therefore suggest opportunities exist for both the wild re community and ML practitioners to apply ML methods in these areas. Moreover, the increasing availability of large spatio-temporal datasets, from climate models or remote sensing for example, may be amenable to the use of deep learning methods, which can eciently extract spatial or temporal features from data. Another major opportunity is the application of agent based learning to re management operations, although many other opportunities exist. However, we must recognize that despite ML models being able to learn on their own, expertise in wild re science is necessary to ensure realistic modelling of wild re processes across multiple scales, while the complexity of some ML methods (e.g. DL) requires a dedicated and sophisticated knowledge of their application. Furthermore, a major obstacle for the adoption of ML methods to re modeling tasks is the perceived lack of interpretability of such methods, which are often considered to be black box models. The ML community, however, recognizes this problem and work is underway to develop methods that allow for greater interpretability of ML methods (see for example, [McGovern et al., 2019]). Data driven approaches are by de nition data dependent | if the re management community wants to more fully exploit powerful 35 ML methods, we need to consider data as a valuable resource and examine what further information on re events or operations are needed to apply ML approaches to management problems. Thus, wildland re science is a diverse multi-faceted discipline that requires a multi-pronged approach, a challenge made greater by the need to mitigate and adapt to current and future re regimes. Acknowledgments The motivation for this paper arose from the \Not the New Normal" BC AI Wild re Symposium held in Vancouver, BC, on 12 October 2018. The authors would also like to thank Intact Insurance, the Canadian Partnership for Wildland Fire Science, the NSERC Discovery Grants program and the Microsoft AI for Social Good program for their support. References Baker Abdalhaq, Ana Cort es, Tom as Margalef, and Emilio Luque. Enhancing wildland re prediction on cluster systems applying evolutionary optimization techniques. Future Generation Computer Systems, 21(1):61{67, 2005. ISSN 0167739X. doi: 10.1016/j.future.2004.09.013. Hamed Adab. Land re hazard assessment in the Caspian Hyrcanian forest ecoregion with the long-term MODIS active re data. Natural Hazards, 87(3):1807{1825, jul 2017. ISSN 0921-030X. doi: 10.1007/ s11069-017-2850-2. URL http://link.springer.com/10.1007/s11069-017-2850-2. Hamed Adab, Azadeh Atabati, Sandra Oliveira, and Ahmad Moghaddam Gheshlagh. Assessing re hazard potential and its main drivers in Mazandaran province, Iran: a data-driven approach. Environmental Monitoring and Assessment, 190(11):670, nov 2018. ISSN 0167-6369. doi: 10.1007/s10661-018-7052-1. URL http://link.springer.com/10.1007/s10661-018-7052-1. Moulay A. Akhlou , Roger Booto Tokime, and Hassan Elassady. Wildland res detection and segmentation using deep learning. In Mohammad S. Alam, editor, Pattern Recognition and Track- ing XXIX, page 11. SPIE, apr 2018. ISBN 9781510618091. doi: 10.1117/12.2304936. URL https://www.spiedigitallibrary.org/conference-proceedings-of-spie/10649/2304936/ Wildland-fires-detection-and-segmentation-using-deep-learning/10.1117/12.2304936. full. K. R. Al-Rawi, J. L. Casanova, and A. Calle. Burned area mapping system and re detection system, based on neural networks and NOAA-AVHRR imagery. International Journal of Remote Sensing, 22(10):2015{ 2032, jan 2001. ISSN 0143-1161. doi: 10.1080/01431160117531. URL https://www.tandfonline.com/ doi/full/10.1080/01431160117531. K R Al-Rawi, J L Casanova, A Romo, and E M Louakfaoui. Integrated re evolution monitoring system (IFEMS) for monitoring spatial-temporal behaviour of multiple re phenomena. International Journal of Remote Sensing, 23(10):1967{1983, 2002. ISSN 01431161. doi: 10.1080/01431160110069809. URL http://www.tandfonline.com/action/journalInformation?journalCode=tres20. Dima Alberg. An Interval Tree Approach to Predict Forest Fires using Meteorological Data. International Journal of Computer Applications, 132(4):17{22, 2015. doi: 10.5120/ijca2015907398. Andrew Aldersley, Steven J. Murray, and Sarah E. Cornell. Global and regional analysis of climate and human drivers of wild re. Science of the Total Environment, 409(18):3472{3481, 2011. ISSN 00489697. doi: 10.1016/j.scitotenv.2011.05.032. URL http://dx.doi.org/10.1016/j.scitotenv.2011.05.032. 36 Dmitriy Alexandrov, Elizaveta Pertseva, Ivan Berman, Igor Pantiukhin, and Aleksandr Kapitonov. Anal- ysis of machine learning methods for wild re security monitoring with an unmanned aerial vehicles. In Conference of Open Innovation Association, FRUCT, volume 2019-April, pages 3{9. IEEE Computer Society, may 2019. ISBN 9789526865386. doi: 10.23919/FRUCT.2019.8711917. Samaher Al Janabi, Ibrahim Al Shourbaji, and Mahdi A. Salman. Assessing the suitability of soft comput- ing approaches for forest res prediction. Applied Computing and Informatics, 14(2):214{224, jul 2018. ISSN 2210-8327. doi: 10.1016/J.ACI.2017.09.006. URL https://www.sciencedirect.com/science/ article/pii/S2210832717301539. A. Alonso-Benito, P. A. Hernandez-Leal, A. Gonzalez-Calvo, M. Arbelo, and A. Barreto. Analysis of Di erent Methods for Burnt Area Estimation using Remote Sensing and Ground Truth Data. In IGARSS 2008 - 2008 IEEE International Geoscience and Remote Sensing Symposium, pages III { 828{III { 831. IEEE, 2008. ISBN 978-1-4244-2807-6. doi: 10.1109/IGARSS.2008.4779477. URL http://ieeexplore. ieee.org/document/4779477/. Amparo Alonso-Betanzos, Oscar Fontenla-Romero, Bertha Guijarro-BerdinasBerdi, ~ Elena Hern andez- Pereira, Juan Canda, Eulogio Jimenez, Jos e Luis Legido, Susana MunizMu, ~ Cristina Paz-Andrade, and Mar a Inmaculada Paz-Andrade. A Neural Network Approach for Forestal Fire Risk Estimation. Technical report, 2002. Amparo Alonso-Betanzos, Oscar Fontenla-Romero, Bertha Guijarro-Berdin ~as, Elena Hern andez-Pereira, Mar a Inmaculada Paz Andrade, Eulogio Jim enez, Jose Luis Legido Soto, and Tarsy Carballas. An intelligent system for forest re risk prediction and re ghting management in Galicia. Expert Systems with Applications, 25(4):545{554, 2003. ISSN 09574174. doi: 10.1016/S0957-4174(03)00095-2. N. S. Altman. An introduction to kernel and nearest-neighbor nonparametric regression. American Statis- tician, 46(3):175{185, 1992. ISSN 15372731. doi: 10.1080/00031305.1992.10475879. Giuseppe Amatulli and Andrea Camia. Exploring the relationships of re occurrence variables by means of CART and MARS models. In Proceedings of the 4th International Wildland Fire Conference, pages 1{11, 2007. Giuseppe Amatulli, Maria Jo~ ao Rodrigues, Marco Trombetti, and Ra aella Lovreglio. Assessing long- term re risk at local scale by means of decision tree technique. Journal of Geophysical Research: Biogeosciences, 111(G4), dec 2006. ISSN 01480227. doi: 10.1029/2005JG000133. URL http://doi. wiley.com/10.1029/2005JG000133. Giuseppe Amatulli, Andrea Camia, and Jesus  San-Miguel-Ayanz. Estimating future burned areas un- der changing climate in the EU-Mediterranean countries. Science of The Total Environment, 450- 451:209{222, apr 2013. ISSN 0048-9697. doi: 10.1016/J.SCITOTENV.2013.02.014. URL https: //www.sciencedirect.com/science/article/pii/S0048969713001770. K. Angayarkkani and N. Radhakrishnan. An Intelligent System For E ective Forest Fire Detection Using Spatial Data. feb 2010. URL http://arxiv.org/abs/1002.2199. K. Angayarkkani and N. Radhakrishnan. An e ective technique to detect forest re region through AN- FIS with spatial data. In ICECT 2011 - 2011 3rd International Conference on Electronics Computer Technology, volume 3, pages 24{30, 2011. ISBN 9781424486779. doi: 10.1109/ICECTECH.2011.5941794. Bachisio Arca, Tiziano Ghisu, and Giuseppe A. Trun o. GPU-accelerated multi-objective optimization of fuel treatments for mitigating wild re hazard. Journal of Computational Science, 11:258{268, 2015. ISSN 18777503. doi: 10.1016/j.jocs.2015.08.009. 37 Sally Archibald, David P. Roy, Brian W. van Wilgen, and Robert J. Scholes. What limits re? An examination of drivers of burnt area in Southern Africa. Global Change Biology, 15(3):613{630, 2009. ISSN 13541013. doi: 10.1111/j.1365-2486.2008.01754.x. Juan P. Arganaraz, ~ Gregorio Gavier Pizarro, Marcelo Zak, Marcos A. Landi, and Laura M Bellis. Hu- man and biophysical drivers of res in Semiarid Chaco mountains of Central Argentina. Science of The Total Environment, 520:1{12, jul 2015. ISSN 0048-9697. doi: 10.1016/J.SCITOTENV.2015. 02.081. URL https://www-sciencedirect-com.login.ezproxy.library.ualberta.ca/science/ article/pii/S0048969715002338. Hilary Arksey and Lisa O'Malley. Scoping studies: towards a methodological framework. International Journal of Social Research Methodology, 8(1):19{32, feb 2005. ISSN 1364- 5579. doi: 10.1080/1364557032000119616. URL http://www.tandfonline.com/doi/abs/10.1080/ James D. Arnold, Simon C. Brewer, and Philip E. Dennison. Modeling Climate-Fire Connections within the Great Basin and Upper Colorado River Basin, Western United States. Fire Ecology, 10(2):64{75, aug 2014. ISSN 19339747. doi: 10.4996/ reecology.1002064. URL http://fireecologyjournal.org/ journal/abstract/?abstract=220. A. Arpaci, B. Malowerschnig, O. Sass, and H. Vacik. Using multi variate data mining techniques for estimating re susceptibility of Tyrolean forests. Applied Geography, 53:258{270, sep 2014. ISSN 0143-6228. doi: 10.1016/J.APGEOG.2014.05.015. URL https://www.sciencedirect.com/science/ article/abs/pii/S0143622814001106. B.C. Arrue, A. Ollero, and J.R. Matinez de Dios. An intelligent system for false alarm reduction in infrared forest- re detection. IEEE Intelligent Systems, 15(3):64{73, may 2000. ISSN 1094-7167. doi: 10.1109/5254.846287. URL http://ieeexplore.ieee.org/document/846287/. Tom as Art es, Andr es Cencerrado, Ana Cort es, and Tom as Margalef. Core Allocation Policies on Multicore Platforms to Accelerate Forest Fire Spread Predictions. PPAM 2013: Parallel Processing and Applied Mathematics, pages 151{160, 2014. ISSN 16113349. doi: 10.1007/978-3-642-55195-6. URL https: //link.springer.com/chapter/10.1007/978-3-642-55195-6{_}14{#}enumeration. Tom as Art es, Andr es Cencerrado, Ana Cort es, and Tom as Margalef. Real-time genetic spatial opti- mization to improve forest re spread forecasting in high-performance computing environments. In- ternational Journal of Geographical Information Science, 30(3):594{611, 2016. ISSN 13623087. doi: 10.1080/13658816.2015.1085052. URL http://dx.doi.org/10.1080/13658816.2015.1085052. Tom as Art es, Andr es Cencerrado, Ana Cort es, and Tom as Margalef. Time aware genetic algorithm for forest re propagation prediction: exploiting multi-core platforms. Concurrency Computation, 29(9): 1{18, 2017. ISSN 15320634. doi: 10.1002/cpe.3837. Davide Ascoli, Giorgio Vacchiano, Renzo Motta, and Giovanni Bovio. Building Rothermel re behaviour fuel models by genetic algorithm optimisation. International Journal of Wildland Fire, 24(3):317{328, 2015. ISSN 10498001. doi: 10.1071/WF14097. Rui Ba, Chen Chen, Jing Yuan, Weiguo Song, and Siuming Lo. SmokeNet: Satellite Smoke Scene Detection Using Convolutional Neural Network with Spatial and Channel-Wise Attention. Remote Sensing, 11(14): 1702, jul 2019. ISSN 2072-4292. doi: 10.3390/rs11141702. URL https://www.mdpi.com/2072-4292/ 11/14/1702. Shitai Bao, Ningchuan Xiao, Zehui Lai, Heyuan Zhang, and Changjoo Kim. Optimizing watchtower locations for forest re monitoring using location models. Fire Safety Journal, 71(December 2013):100{ 109, 2015. ISSN 03797112. doi: 10.1016/j. resaf.2014.11.016. URL http://dx.doi.org/10.1016/j. firesaf.2014.11.016. 38 Avi Bar Massada, Alexandra D. Syphard, Susan I. Stewart, and Volker C. Radelo . Wild re ignition- distribution modelling: a comparative study in the Huron?Manistee National Forest, Michigan, USA. International Journal of Wildland Fire, 22(2):174, apr 2013. ISSN 1049-8001. doi: 10.1071/WF11178. URL http://www.publish.csiro.au/?paper=WF11178. Panagiotis Barmpoutis, Kosmas Dimitropoulos, Kyriaki Kaza, and Nikos Grammalidis. Fire Detection from Images Using Faster R-CNN and Multidimensional Texture Analysis. In ICASSP, IEEE International Conference on Acoustics, Speech and Signal Processing - Proceedings, volume 2019-May, pages 8301{ 8305. Institute of Electrical and Electronics Engineers Inc., may 2019. ISBN 9781479981311. doi: 10.1109/ICASSP.2019.8682647. K. Barrett, A. D. McGuire, E. E. Hoy, and E. S. Kasischke. Potential shifts in dominant forest cover in interior Alaska driven by variations in re severity. Ecological Applications, 21(7):2380{2396, 2011. ISSN 10510761. doi: 10.1890/10-0896.1. Hossein Bashari, Ali Asghar Naghipour, Seyed Jamaleddin Khajeddin, Hamed Sangoony, and Pejman Tah- masebi. Risk of re occurrence in arid and semi-arid ecosystems of Iran: an investigation using Bayesian belief networks. Environmental Monitoring and Assessment, 188(9):531, sep 2016. ISSN 0167-6369. doi: 10.1007/s10661-016-5532-8. URL http://link.springer.com/10.1007/s10661-016-5532-8. Bryson C. Bates, Andrew J. Dowdy, Richard E. Chandler, Bryson C. Bates, Andrew J. Dowdy, and Richard E. Chandler. Classi cation of Australian Thunderstorms Using Multivariate Analyses of Large- Scale Atmospheric Variables. Journal of Applied Meteorology and Climatology, 56(7):1921{1937, jul 2017. ISSN 1558-8424. doi: 10.1175/JAMC-D-16-0271.1. URL http://journals.ametsoc.org/doi/ 10.1175/JAMC-D-16-0271.1. Enric Batllori, Marc-Andr e Parisien, Meg A. Krawchuk, and Max A. Moritz. Climate change-induced shifts in re for Mediterranean ecosystems. Global Ecology and Biogeography, 22(10):1118{1129, oct 2013. ISSN 1466822X. doi: 10.1111/geb.12065. URL http://doi.wiley.com/10.1111/geb.12065. Peter Bauer, Alan Thorpe, and Gilbert Brunet. The quiet revolution of numerical weather prediction. Nature, 525(7567):47{55, sep 2015. ISSN 0028-0836. doi: 10.1038/nature14956. URL http://www. nature.com/articles/nature14956. A. Belayneh, J. Adamowski, B. Khalil, and B. Ozga-Zielinski. Long-term SPI drought forecasting in the Awash River Basin in Ethiopia using wavelet neural networks and wavelet support vector regression models. Journal of Hydrology, 508:418{429, jan 2014. ISSN 00221694. doi: 10.1016/j.jhydrol.2013.10.052. Mar Bisquert, Eduardo Caselles, Juan Manuel S anchez, and Vicente Caselles. Application of arti cial neural networks and logistic regression to the prediction of forest re danger in Galicia using MODIS data. International Journal of Wildland Fire, 21(8):1025, dec 2012. ISSN 1049-8001. doi: 10.1071/WF11105. URL http://www.publish.csiro.au/?paper=WF11105. Karen D. Blouin, Mike D. Flannigan, Xianli Wang, and Bohdan Kochtubajda. Ensemble lightning predic- tion models for the province of Alberta, Canada. International Journal of Wildland Fire, 25(4):421{432, apr 2016. ISSN 1049-8001. doi: 10.1071/WF15111. URL http://www.publish.csiro.au/?paper= WF15111. William J. Bond and Jon E. Keeley. Fire as a global 'herbivore': The ecology and evolution of ammable ecosystems. Trends in Ecology and Evolution, 20(7):387{394, jul 2005. ISSN 01695347. doi: 10.1016/j. tree.2005.04.025. Yan Boulanger, Marc-Andr e Parisien, and Xianli Wang. Model-speci cation uncertainty in future area burned by wild res in Canada. International Journal of Wildland Fire, 27(3):164, apr 2018. ISSN 1049-8001. doi: 10.1071/WF17123. URL http://www.publish.csiro.au/?paper=WF17123. 39 Curtis M. Bradley, Chad T. Hanson, and Dominick A. DellaSala. Does increased forest protection corre- spond to higher re severity in frequent- re forests of the western United States? Ecosphere, 7(10):1{13, 2016. ISSN 21508925. doi: 10.1002/ecs2.1492. Jonathan Branham, Nicholas Hamilton, Dale Hamilton, and Barry Myers. Evaluation of Image Spatial Resolution for Machine Learning Mapping of Wildland Fire E ects, jul 2017. URL https://scholarworks.boisestate.edu/icur/2017/Poster{_}Session/26https:// link.springer.com/chapter/10.1007/978-3-030-01054-6{_}29. Leo Breiman. Statistical modeling: The two cultures. Statistical Science, 16(3):199{215, 2001. ISSN 08834237. doi: 10.1214/ss/1009213726. Leo Breiman. Classi cation and regression trees. Routledge, 2017. Leo Breiman, Jerome Friedman, Richard A Olshen, and Charles J Stone. Classi cation and regression trees Chapman & Hall. New York, 1984. N. D. Brenowitz and C. S. Bretherton. Prognostic Validation of a Neural Network Uni ed Physics Parameterization. Geophysical Research Letters, 45(12):6289{6298, jun 2018. ISSN 00948276. doi: 10.1029/2018GL078510. URL http://doi.wiley.com/10.1029/2018GL078510. Steven P. Brumby, Neal R. Harvey, Je rey J. Bloch, James P. Theiler, Simon J. Perkins, Aaron C. Young, and John J. Szymanski. Evolving forest re burn severity classi cation algorithms for multispectral imagery. Algorithms for Multispectral, Hyperspectral, and Ultraspectral Imagery VII, 4381(August 2001): 236{245, 2001. doi: 10.1117/12.437013. C. E. Buckland, R. M. Bailey, and D. S. G. Thomas. Using arti cial neural networks to predict future dryland responses to human and climate disturbances. Scienti c Reports, 9(1):3855, dec 2019. ISSN 2045-2322. doi: 10.1038/s41598-019-40429-5. URL http://www.nature.com/articles/ s41598-019-40429-5. Declan Butler. AI summit aims to help world's poorest, jun 2017. ISSN 14764687. Wenhua Cai, Jian Yang, Zhihua Liu, Yuanman Hu, and Peter J. Weisberg. Post- re tree recruitment of a boreal larch forest in Northeast China. Forest Ecology and Management, 307:20{29, 2013. ISSN 03781127. doi: 10.1016/j.foreco.2013.06.056. URL http://dx.doi.org/10.1016/j.foreco.2013.06.056. X. Cao, J. Chen, B. Matsushita, H. Imura, and L. Wang. An automatic method for burn scar mapping using support vector machines. International Journal of Remote Sensing, 30(3):577{594, feb 2009. ISSN 0143-1161. doi: 10.1080/01431160802220219. URL https://www.tandfonline.com/doi/full/ 10.1080/01431160802220219. Yichao Cao, Feng Yang, Qingfei Tang, and Xiaobo Lu. An Attention Enhanced Bidirectional LSTM for Early Forest Fire Smoke Recognition. IEEE Access, pages 1{1, oct 2019. doi: 10.1109/access.2019. Yinxue Cao, Ming Wang, and Kai Liu. Wild re Susceptibility Assessment in Southern China: A Compar- ison of Multiple Methods. International Journal of Disaster Risk Science, 8(2):164{181, jun 2017. ISSN 21926395. doi: 10.1007/s13753-017-0129-6. Adri an Cardil, Blas Mola-Yudego, Angela Bl azquez-Casado, and Jos e Ram on Gonz alez-Olabarria. Fire and burn severity assessment: Calibration of Relative Di erenced Normalized Burn Ratio (RdNBR) with eld data. Journal of Environmental Management, 235(January):342{349, 2019. ISSN 10958630. doi: 10.1016/j.jenvman.2019.01.077. URL https://doi.org/10.1016/j.jenvman.2019.01.077. 40 Carlos Carrillo, Tom as Art es, Ana Cort es, and Tom as Margalef. Error function impact in dynamic data- driven framework applied to forest re spread prediction. Procedia Computer Science, 80:418{427, 2016. ISSN 18770509. doi: 10.1016/j.procs.2016.05.342. URL http://dx.doi.org/10.1016/j.procs.2016. 05.342. Mauro Castelli, Leonardo Vanneschi, and Ale s Popovi c. PREDICTING BURNED AREAS OF FOR- EST FIRES: AN ARTIFICIAL INTELLIGENCE APPROACH. Fire Ecology, 11(1):106{118, apr 2015. ISSN 19339747. doi: 10.4996/ reecology.1101106. URL http://fireecologyjournal.org/journal/ abstract/?abstract=236. Turgay Celik. Change detection in satellite images using a genetic algorithm approach. IEEE Geoscience and Remote Sensing Letters, 7(2):386{390, 2010. ISSN 1545598X. doi: 10.1109/LGRS.2009.2037024. Andr es Cencerrado, Ana Cort es, and Tom as Margalef. Genetic algorithm characterization for the quality assessment of forest re spread prediction. Procedia Computer Science, 9:312{320, 2012. ISSN 18770509. doi: 10.1016/j.procs.2012.04.033. Andr es Cencerrado, Ana Cort es, and Tom as Margalef. Applying probability theory for the quality as- sessment of a wild re spread prediction framework based on genetic algorithms. The Scienti c World Journal, 2013, 2013. ISSN 1537744X. doi: 10.1155/2013/728414. Andr es Cencerrado, Ana Cort es, and Tom as Margalef. Response time assessment in forest re spread simulation: An integrated methodology for ecient exploitation of available prediction time. Environ- mental Modelling & Software, 54:153{164, apr 2014. ISSN 1364-8152. doi: 10.1016/J.ENVSOFT.2014. 01.008. URL https://www-sciencedirect-com.login.ezproxy.library.ualberta.ca/science/ article/pii/S1364815214000176. Supriyo Chakraborty, Richard Tomsett, Ramya Raghavendra, Daniel Harborne, Moustafa Alzantot, Federico Cerutti, Mani Srivastava, Alun Preece, Simon Julier, Raghuveer M Rao, and Others. In- terpretability of deep learning models: a survey of results. In 2017 IEEE SmartWorld, Ubiq- uitous Intelligence & Computing, Advanced & Trusted Computed, Scalable Computing & Commu- nications, Cloud & Big Data Computing, Internet of People and Smart City Innovation (Smart- World/SCALCOM/UIC/ATC/CBDCom/IOP/SCI), pages 1{6. IEEE, 2017. Fs Chapin, T.N. Hollingsworth, and Re Hewitt. Fire e ects on seedling establishment success across treeline: implications for future tree migration and ammability in a changing climate. 2014. URL http://digitalcommons.unl.edu/jfspresearch/82/. Feng Chen, Yongsheng Du, Shukui Niu, and Jinlong Zhao. Modeling Forest Lightning Fire Occurrence in the Daxinganling Mountains of Northeastern China with MAXENT. Forests, 6(12):1422{1438, apr 2015. ISSN 1999-4907. doi: 10.3390/f6051422. URL http://www.mdpi.com/1999-4907/6/5/1422. Tao Cheng and Jiaqiu Wang. Integrated Spatio-temporal Data Mining for Forest Fire Prediction. Trans- actions in GIS, 12(5):591{611, sep 2008. ISSN 13611682. doi: 10.1111/j.1467-9671.2008.01117.x. URL http://doi.wiley.com/10.1111/j.1467-9671.2008.01117.x. Khaled Chetehouna, Eddy El Tabach, Loubna Bouazaoui, and Nicolas Gascoin. Predicting the ame characteristics and rate of spread in res propagating in a bed of Pinus pinaster using Arti cial Neu- ral Networks. Process Safety and Environmental Protection, 98:50{56, nov 2015. ISSN 0957-5820. doi: 10.1016/J.PSEP.2015.06.010. URL https://www.sciencedirect.com/science/article/pii/ S0957582015001111. Tatenda T Chingono and C Mbohwa. Fire Hazard Modelling in Southern Africa. In Proceedings of the World Congress on Engineering and Computer Science, San Francisco, 2015. URL http://www.iaeng. org/publication/WCECS2015/WCECS2015{_}pp514-519.pdf. 41 G. Chirici, R. Scotti, A. Montaghi, A. Barbati, R. Cartisano, G. Lopez, M. Marchetti, R. E. Mcroberts, H. Olsson, and P. Corona. Stochastic gradient boosting classi cation trees for forest fuel types mapping through airborne laser scanning and IRS LISS-III imagery. International Journal of Applied Earth Observation and Geoinformation, 25(1):87{97, 2013. ISSN 15698432. doi: 10.1016/j.jag.2013.04.006. Emilio Chuvieco, F. Javier Salas, Luis Carvacho, and Francisco Rodr guez-Silva. Integrated re risk mapping. In Remote Sensing of Large Wild res, pages 61{100. Springer Berlin Heidelberg, Berlin, Heidelberg, 1999. doi: 10.1007/978-3-642-60164-4 5. URL http://link.springer.com/10.1007/ 978-3-642-60164-4{_}5. Hamish Clarke, Rebecca Gibson, Brett Cirulis, Ross A Bradstock, and Trent D Penman. Developing and testing models of the drivers of anthropogenic and lightning-caused wild re ignitions in south- eastern Australia. Journal of Environmental Management, 235:34{41, apr 2019. ISSN 0301-4797. doi: 10.1016/J.JENVMAN.2019.01.055. URL https://www-sciencedirect-com.login.ezproxy. library.ualberta.ca/science/article/pii/S0301479719300568. J. Coen. Some Requirements for Simulating Wildland Fire Behavior Using Insight from Coupled Weather|Wildland Fire Models. Fire, 1(1):6, feb 2018. ISSN 2571-6255. doi: 10.3390/ re1010006. URL http://www.mdpi.com/2571-6255/1/1/6. Shane R. Coeld, Casey A. Gra , Yang Chen, Padhraic Smyth, E Foufoula-Georgiou, and James T. Randerson. Machine learning to predict nal re size at the time of ignition. International Journal of Wildland Fire, 28(11):861, 2019. ISSN 1049-8001. doi: 10.1071/WF19023. URL http://www.publish. csiro.au/?paper=WF19023. Judah Cohen, Dim Coumou, Jessica Hwang, Lester Mackey, Paulo Orenstein, Sonja Totz, and Eli Tziper- man. S2S reboot: An argument for greater inclusion of machine learning in subseasonal to seasonal forecasts. Wiley Interdisciplinary Reviews: Climate Change, 10(2), mar 2019. ISSN 1757-7780. doi: 10.1002/wcc.567. URL https://onlinelibrary.wiley.com/doi/abs/10.1002/wcc.567. L. Collins, P. Grioen, G. Newell, and A. Mellor. The utility of Random Forests for wild re severity mapping. Remote Sensing of Environment, 216:374{384, oct 2018. ISSN 0034-4257. doi: 10.1016/J.RSE. 2018.07.005. URL https://www.sciencedirect.com/science/article/pii/S0034425718303328. Sean C.P. Coogan, Fran cois Nicolas Robinne, Piyush Jain, and Mike D. Flannigan. Scientists' warning on wild re | a canadian perspective. Canadian Journal of Forest Research, 49(9):1015{1023, 2019. ISSN 12086037. doi: 10.1139/cjfr-2019-0094. Michelle Coppoletta, Kyle E. Merriam, and Brandon M. Collins. Post- re vegetation and fuel development in uences re severity patterns in reburns. Ecological Applications, 26(3):686{699, apr 2016. ISSN 1051-0761. doi: 10.1890/15-0225. A. Cordoba, R. Vilar, A. Lavrov, A. B. Utkin, and A. Fernandes. Multi-objective optimisation of lidar parameters for forest- re detection on the basis of a genetic algorithm. Optics and Laser Technology, 36 (5):393{400, 2004. ISSN 00303992. doi: 10.1016/j.optlastec.2003.10.010. Paulo Cortez and An bal de Jesus Raimundo Morais. A data mining approach to predict forest res using meteorological data. 2007. URL https://repositorium.sdum.uminho.pt/handle/1822/8039. Sergi Costafreda-Aumedes, Adrian Cardil, Domingo M. Molina, Sarah N. Daniel, Robert Mavsar, and Cristina Vega-Garcia. Analysis of factors in uencing deployment of re suppression resources in Spain using arti cial neural networks. iForest, 9(Feb 2016):138{145, 2015. ISSN 19717458. doi: 10.3832/ ifor1329-008. 42 Michael A. Crimmins. Synoptic climatology of extreme re-weather conditions across the southwest United States. International Journal of Climatology, 26(8):1001{1016, jun 2006. ISSN 0899-8418. doi: 10.1002/ joc.1300. URL http://doi.wiley.com/10.1002/joc.1300. Morgan A. Crowley, Je rey A. Cardille, Joanne C. White, and Michael A. Wulder. Multi-sensor, multi- scale, Bayesian data synthesis for mapping within-year wild re progression. Remote Sensing Letters, 10(3):302{311, 2019. ISSN 2150-704X. doi: 10.1080/2150704X.2018.1536300. URL https://www. tandfonline.com/doi/full/10.1080/2150704X.2018.1536300. Thomas Curt, Laurent Borgniet, Thomas Ibanez, Vincent Moron, and Christelle H ely. Understanding re patterns and re drivers for setting a sustainable management policy of the New-Caledonian biodiversity hotspot. Forest Ecology and Management, 337:48{60, 2015. ISSN 03781127. doi: 10.1016/j.foreco.2014. 10.032. URL http://dx.doi.org/10.1016/j.foreco.2014.10.032. Thomas Curt, Thibaut Fr ejaville, and S ebastien Lahaye. Modelling the spatial patterns of ignition causes and re regime features in southern France: Implications for re prevention policy. International Journal of Wildland Fire, 25(7):785{796, 2016. ISSN 10498001. doi: 10.1071/WF15205. James Richard Davis, J R L Hoare, and P M Nanninga. Developing a re man- agement expert system for Kakadu National Park, Australia. Journal of Environmental Management, 1986. URL https://publications.csiro.au/rpr/pub?list=BRO{&}pid=procite: f221b911-2e97-4c9f-be9b-f1155bf48c24. J.R. Davis, P.M. Nanninga, J.R.L. Hoare, and A.J. Press. Transferring scienti c knowledge to natu- ral resource managers using arti cial intelligence concepts. Ecological Modelling, 46(1-2):73{89, jul 1989. ISSN 0304-3800. doi: 10.1016/0304-3800(89)90070-7. URL https://www.sciencedirect.com/ science/article/pii/0304380089900707. Raymond Davis, Zhiqiang Yang, Andrew Yost, Cole Belongie, and Warren Cohen. The normal re en- vironment|Modeling environmental suitability for large forest wild res using past, present, and fu- ture climate normals. Forest Ecology and Management, 390:173{186, apr 2017. ISSN 0378-1127. doi: 10.1016/J.FORECO.2017.01.027. URL https://www.sciencedirect.com/science/article/ pii/S0378112716309318. Antonella De Angelis, Carlo Ricotta, Marco Conedera, and Gianni Boris Pezzatti. Modelling the Meteoro- logical Forest Fire Niche in Heterogeneous Pyrologic Conditions. PLOS ONE, 10(2):e0116875, feb 2015. ISSN 1932-6203. doi: 10.1371/journal.pone.0116875. URL https://dx.plos.org/10.1371/journal. pone.0116875. P. P. de Bem, O. A. de Carvalho Ju nior, E. A. T. Matricardi, R. F. Guimar~ aes, and R. A. T. Gomes. Predicting wild re vulnerability using logistic regression and arti cial neural networks: a case study in Brazil's Federal District. International Journal of Wildland Fire, 28(1):35, feb 2018. ISSN 1049-8001. doi: 10.1071/wf18018. URL http://www.publish.csiro.au/?paper=WF18018. Marla Jose Perestrello De Vasconcelos, Sara Sllva, Margarlda Tome, Margarlda Alvim, Jose Mlguel, and Cardoso Perelra. Spatial Prediction of Fire Ignition Probabilities: Comparing Logistic Regression and Neural Networks. Photogrammetric Engineering & Remote Sensing, 67(1):73{81, 2001. Haifa Debouk, Ramon Riera-Tatch e, and Cristina Vega-Garc a. Assessing Post-Fire Regeneration in a Mediterranean Mixed Forest Using Lidar Data and Arti cial Neural Networks. Photogrammetric Engi- neering & Remote Sensing, 2013. ISSN 00991112. doi: 10.14358/PERS.79.12.1121. Rosario Delgado, Jos e-Luis Gonz alez, Andr es Sotoca, and Xavier-Andoni Tibau. Archetypes of Wild re Arsonists: An Approach by Using Bayesian Networks. In Forest Fire. InTech, may 43 2018. doi: 10.5772/intechopen.72615. URL http://www.intechopen.com/books/forest-fire/ archetypes-of-wildfire-arsonists-an-approach-by-using-bayesian-networks. M onica Denham and Karina Laneri. Using ecient parallelization in Graphic Processing Units to param- eterize stochastic re propagation models. Journal of Computational Science, 25:76{88, 2018. ISSN 18777503. doi: 10.1016/j.jocs.2018.02.007. URL https://doi.org/10.1016/j.jocs.2018.02.007. M onica Denham, Kerstin Wendt, Germ an Bianchini, Ana Cort es, and Tom as Margalef. Dynamic Data- Driven Genetic Algorithm for forest re spread prediction. Journal of Computational Science, 3(5): 398{404, 2012. ISSN 18777503. doi: 10.1016/j.jocs.2012.06.002. URL http://dx.doi.org/10.1016/j. jocs.2012.06.002. Peter J. Diggle, Raquel Menezes, and Ting-li Su. Geostatistical inference under preferential sam- pling. Journal of the Royal Statistical Society: Series C (Applied Statistics), 59(2):191{232, mar 2010. ISSN 00359254. doi: 10.1111/j.1467-9876.2009.00701.x. URL http://doi.wiley.com/10.1111/j. 1467-9876.2009.00701.x. Luca Antonio Dimuccio, Rui Ferreira, Lucio Cunha, and Antonio Campar de Almeida. Regional forest- re susceptibility analysis in central Portugal using a probabilistic ratings procedure and arti cial neural network weights assignment. International Journal of Wildland Fire, 20(6):776, oct 2011. ISSN 1049- 8001. doi: 10.1071/WF09083. URL http://www.publish.csiro.au/?paper=WF09083. T. L. Divya and M. N. Vijayalakshmi. Inference of Replanting in Forest Fire A ected Land Using Data Mining Technique. pages 121{129. Springer, New Delhi, 2016. doi: 10.1007/978-81-322-2734-2 13. URL http://link.springer.com/10.1007/978-81-322-2734-2{_}13. Wisdom M. Dlamini. A Bayesian belief network analysis of factors in uencing wild re occurrence in Swaziland. Environmental Modelling and Software, 25(2):199{208, 2010. ISSN 13648152. doi: 10.1016/ j.envsoft.2009.08.002. URL http://dx.doi.org/10.1016/j.envsoft.2009.08.002. Wisdom Mdumiseni Dlamini. Application of Bayesian networks for re risk mapping using GIS and remote sensing data. GeoJournal, 76(3):283{296, jun 2011. ISSN 0343-2521. doi: 10.1007/s10708-010-9362-x. URL http://link.springer.com/10.1007/s10708-010-9362-x. E. Dragozi, I. Gitas, D. Stavrakoudis, and J. Theocharis. Burned Area Mapping Using Support Vector Machines and the FuzCoC Feature Selection Method on VHR IKONOS Imagery. Remote Sensing, 6 (12):12005{12036, dec 2014. ISSN 2072-4292. doi: 10.3390/rs61212005. URL http://www.mdpi.com/ 2072-4292/6/12/12005. Andrea Duane, M riam Piqu e, Marc Castellnou, and Llu s Brotons. Predictive modelling of re occurrences from di erent re spread patterns in Mediterranean landscapes. International Journal of Wildland Fire, 24(3):407, jun 2015. ISSN 1049-8001. doi: 10.1071/WF14040. URL http://www.publish.csiro.au/ ?paper=WF14040. Ritaban Dutta, Jagannath Aryal, Aruneema Das, and Jamie B. Kirkpatrick. Deep cognitive imaging systems enable estimation of continental-scale re incidence from climate data. Scienti c Reports, 3, 2013. ISSN 20452322. doi: 10.1038/srep03188. Ritaban Dutta, Aruneema Das, and Jagannath Aryal. Big data integration shows Australian bush- re frequency is increasing signi cantly. Royal Society Open Science, 3(2):150241, feb 2016. ISSN 2054-5703. doi: 10.1098/rsos.150241. URL https://royalsocietypublishing.org/doi/10.1098/rsos.150241. Francis K. Dwomoh and Michael C. Wimberly. Fire regimes and their drivers in the Upper Guinean Region of West Africa. Remote Sensing, 9(11), 2017. ISSN 20724292. doi: 10.3390/rs9111117. 44 J.B. Theocharis E. Dragozi, I. Z. Gitas, D.G. Stavrakoudis. A Performance Evaluation Of Sup- port Vector Machines And The Nearest Neighbor Classi er In Classifying Image Objects For Burned Area Mapping. In Proceedings of the 8th International EARSeL FF-SIG Workshop, Stresa, Italy, 2011. URL https://publications.europa.eu/en/publication-detail/-/publication/ 5e90be79-efe8-43dd-8670-2518a43155f4/language-en. Hamid Ebrahimy, Aliakbar Rasuly, *, and Davoud Mokhtari. Development of a Web GIS Sys- tem Based on the MaxEnt Approach for Wild re Management: A Case Study of East Azer- baijan. ECOPERSIA, 5(3):1859{1873, 2017. URL https://pdfs.semanticscholar.org/46b2/ fd74419232dbe2dedccaaca40bab6dbf50b8.pdf. J. Elith, J. R. Leathwick, and T. Hastie. A working guide to boosted regression trees. Journal of Animal Ecology, 77(4):802{813, jul 2008. ISSN 0021-8790. doi: 10.1111/j.1365-2656.2008.01390.x. URL http: //doi.wiley.com/10.1111/j.1365-2656.2008.01390.x. Jane Elith, Steven J. Phillips, Trevor Hastie, Miroslav Dud k, Yung En Chee, and Colin J. Yates. A statistical explanation of MaxEnt for ecologists. Diversity and Distributions, 17(1):43{57, jan 2011. ISSN 13669516. doi: 10.1111/j.1472-4642.2010.00725.x. URL http://doi.wiley.com/10.1111/j. 1472-4642.2010.00725.x. Thomas A. Fairman, Lauren T. Bennett, Shauna Tupper, and Craig R. Nitschke. Frequent wild res erode tree persistence and alter stand structure and initial composition of a re-tolerant sub-alpine forest. Journal of Vegetation Science, 28(6):1151{1165, 2017. ISSN 16541103. doi: 10.1111/jvs.12575. Lei Fang, Jian Yang, Megan White, and Zhihua Liu. Predicting potential re severity using vegetation, topography and surface moisture availability in a Eurasian boreal forest landscape. Forests, 9(3):1{26, 2018. ISSN 19994907. doi: 10.3390/f9030130. Armando M. Fernandes, Andrei B. Utkin, Alexander V. Lavrov, and Rui M. Vilar. Neural Network Based Recognition of Smoke Signatures from Lidar Signals. Neural Processing Letters, 19(3):175{189, jun 2004a. ISSN 1370-4621. doi: 10.1023/B:NEPL.0000035598.19042.42. URL http://link.springer. com/10.1023/B:NEPL.0000035598.19042.42. Armando M. Fernandes, Andrei B. Utkin, Alexander V. Lavrov, and Rui M. Vilar. Development of neural network committee machines for automatic forest re detection using lidar. Pattern Recognition, 37(10): 2039{2047, oct 2004b. ISSN 00313203. doi: 10.1016/j.patcog.2004.04.002. URL http://linkinghub. elsevier.com/retrieve/pii/S0031320304001360. Paulo M. Fernandes, Tiago Monteiro-Henriques, Nuno Guiomar, Carlos Loureiro, and Ana M.G. Barros. Bottom-Up Variables Govern Large-Fire Size in Portugal. Ecosystems, 19(8):1362{1375, 2016. ISSN 14350629. doi: 10.1007/s10021-016-0010-2. Alfonso Fernandez-Manso, Carmen Quintano, and Dar A. Roberts. Burn severity analysis in Mediterranean forests using maximum entropy model trained with EO-1 Hyperion and LiDAR data. ISPRS Journal of Photogrammetry and Remote Sensing, 155(July):102{118, 2019. ISSN 09242716. doi: 10.1016/j.isprsjprs. 2019.07.003. URL https://doi.org/10.1016/j.isprsjprs.2019.07.003. Mark A Finney. FARSITE: Fire Area Simulator-Model Development and Evaluation. Technical report, Rocky Mountain Research Station, 2004. URL http://www.farsite.org. Mark A. Finney. The challenge of quantitative risk analysis for wildland re. In Forest Ecology and Management, volume 211, pages 97{108, jun 2005. doi: 10.1016/j.foreco.2005.02.010. 45 Marisa G. Fonseca, Luiz Eduardo O. C. Arag~ ao, Andr e Lima, Yosio E. Shimabukuro, Egidio Arai, and Liana O. Anderson. Modelling re probability in the Brazilian Amazon using the maximum entropy method. International Journal of Wildland Fire, 25(9):955, sep 2016. doi: 10.1071/WF15216. URL http://www.publish.csiro.au/?paper=WF15216. Matthias Forkel, Wouter Dorigo, Gitta Lasslop, Irene Teubner, Emilio Chuvieco, and Kirsten Thonicke. A data-driven approach to identify controls on global re activity from satellite and climate observations (SOFIA V1). Geoscienti c Model Development, 10(12):4443{4476, dec 2017. ISSN 1991-9603. doi: 10.5194/gmd-10-4443-2017. URL https://www.geosci-model-dev.net/10/4443/2017/. Matthias Forkel, Niels Andela, Sandy P Harrison, Gitta Lasslop, Margreet Van Marle, Emilio Chuvieco, Wouter Dorigo, Matthew Forrest, Stijn Hantson, Angelika Heil, Fang Li, Joe Melton, Stephen Sitch, Chao Yue, and Almut Arneth. Emergent relationships with respect to burned area in global satellite observations and re-enabled vegetation models. Biogeosciences, 16(1):57{76, 2019. ISSN 17264189. doi: 10.5194/bg-16-57-2019. Yoav Freund and Robert E Shapire. A decision-theoretic generalization of on-line learning and an ap- plication to boosting. In Computational Learning Theory: Eurocolt '95, pages 23{37. Springer-Verlag, Jerome H. Friedman. Greedy function approximation: A gradient boosting machine. Ann. Statist., 29(5): 1189{1232, 10 2001. doi: 10.1214/aos/1013203451. URL https://doi.org/10.1214/aos/1013203451. Sigfredo Fuentes, Eden Jane Tongson, Roberta De Bei, Claudia Gonzalez Viejo, Renata Ristic, Stephen Tyerman, and Kerry Wilkinson. Non-Invasive Tools to Detect Smoke Contamination in Grapevine Canopies, Berries and Wine: A Remote Sensing and Machine Learning Modeling Approach. Sensors, 19 (15):3335, jul 2019. ISSN 1424-8220. doi: 10.3390/s19153335. URL https://www.mdpi.com/1424-8220/ 19/15/3335. Mariano Garc a, David Riano, ~ Emilio Chuvieco, Javier Salas, and F. Mark Danson. Multispectral and LiDAR data fusion for fuel type mapping using Support Vector Machine and decision rules. Remote Sensing of Environment, 115(6):1369{1379, 2011. ISSN 00344257. doi: 10.1016/j.rse.2011.01.017. URL http://dx.doi.org/10.1016/j.rse.2011.01.017. Paula Garc a-Llamas, Susana Su arez-Seoane, Angela Taboada, Alfonso Fern andez-Manso, Carmen Quin- tano, V ctor Fern andez-Garc a, Jos e Manuel Fern andez-Guisuraga, Elena Marcos, and Leonor Calvo. Environmental drivers of re severity in extreme re events that a ect Mediterranean pine forest ecosystems. Forest Ecology and Management, 433(October 2018):24{32, 2019. ISSN 03781127. doi: 10.1016/j.foreco.2018.10.051. URL https://doi.org/10.1016/j.foreco.2018.10.051. Stuart Geman, Elie Bienenstock, and Ren e Doursat. Neural Networks and the Bias/Variance Dilemma. Neural Computation, 4(1):1{58, jan 1992. ISSN 0899-7667. doi: 10.1162/neco.1992.4.1.1. Andre Gensler, Janosch Henze, Bernhard Sick, and Nils Raabe. Deep Learning for solar power forecasting - An approach using AutoEncoder and LSTM Neural Networks. In 2016 IEEE International Conference on Systems, Man, and Cybernetics, SMC 2016 - Conference Proceedings, pages 2858{2865. Institute of Electrical and Electronics Engineers Inc., feb 2017. ISBN 9781509018970. doi: 10.1109/SMC.2016. Omid Ghorbanzadeh, Thomas Blaschke, Khalil Gholamnia, and Jagannath Aryal. Forest Fire Susceptibility and Risk Mapping Using Social/Infrastructural Vulnerability and Environmental Variables. Fire, 2(3): 50, sep 2019a. ISSN 2571-6255. doi: 10.3390/ re2030050. URL https://www.mdpi.com/2571-6255/2/ 3/50. 46 Omid Ghorbanzadeh, Khalil Valizadeh Kamran, Thomas Blaschke, Jagannath Aryal, Amin Naboureh, Jamshid Einali, and Jinhu Bian. Spatial Prediction of Wild re Susceptibility Using Field Survey GPS Data and Machine Learning Approaches. Fire, 2(3):43, jul 2019b. ISSN 2571-6255. doi: 10.3390/ re2030043. URL https://www.mdpi.com/2571-6255/2/3/43. Louis Giglio, Luigi Boschetti, David P. Roy, Michael L. Humber, and Christopher O. Justice. The Collection 6 MODIS burned area mapping algorithm and product. Remote Sensing of Environment, 217:72{85, nov 2018. ISSN 00344257. doi: 10.1016/j.rse.2018.08.005. Ljubomir Gigovi c, Hamid Reza Pourghasemi, Sini sa Drobnjak, and Shibiao Bai. Testing a New Ensemble Model Based on SVM and Random Forest in Forest Fire Susceptibility Assessment and Its Mapping in Serbia's Tara National Park. Forests, 10(5):408, may 2019. ISSN 1999-4907. doi: 10.3390/f10050408. URL https://www.mdpi.com/1999-4907/10/5/408. Y.J. Goldarag, Ali Mohammadzadeh, and A. S. Ardakani. Fire Risk Assessment Using Neural Network and Logistic Regression. Journal of the Indian Society of Remote Sensing, 44(6):885{894, 2016. ISSN 09743006. doi: 10.1007/s12524-016-0557-6. URL http://dx.doi.org/10.1007/s12524-016-0557-6. Carla Gomes. Computational Sustainability: Computational Methods for a Sustainable Environment, Economy, and Society. The Bridge, National Academy of Engineering, 39(4), 2009. Israel G omez and M. Pilar Mart n. Prototyping an arti cial neural network for burned area mapping on a regional scale in Mediterranean areas using MODIS images. International Journal of Applied Earth Observation and Geoinformation, 13(5):741{752, 2011. ISSN 15698432. doi: 10.1016/j.jag.2011.05.002. Noel Gorelick, Matt Hancher, Mike Dixon, Simon Ilyushchenko, David Thau, and Rebecca Moore. Google Earth Engine: Planetary-scale geospatial analysis for everyone. Remote Sensing of Environment, 2017. doi: 10.1016/j.rse.2017.06.031. URL https://doi.org/10.1016/j.rse.2017.06.031. Futao Guo, Guangyu Wang, Zhangwen Su, Huiling Liang, Wenhui Wang, Fangfang Lin, and Aiqin Liu. What drives forest re in Fujian, China? Evidence from logistic regression and Random Forests. In- ternational Journal of Wildland Fire, 25(5):505, may 2016a. ISSN 1049-8001. doi: 10.1071/WF15121. URL http://www.publish.csiro.au/?paper=WF15121. Futao Guo, Lianjun Zhang, Sen Jin, Mulualem Tigabu, Zhangwen Su, Wenhui Wang, Futao Guo, Lianjun Zhang, Sen Jin, Mulualem Tigabu, Zhangwen Su, and Wenhui Wang. Modeling Anthropogenic Fire Occurrence in the Boreal Forest of China Using Logistic Regression and Random Forests. Forests, 7(12): 250, oct 2016b. ISSN 1999-4907. doi: 10.3390/f7110250. URL http://www.mdpi.com/1999-4907/7/ 11/250. Dale Hamilton, Barry Myers, and Jonathan Branham. Evaluation Of Texture As An Input Of Spatial Context For Machine Learning Mapping Of Wildland Fire E ects. An International Journal (SIPIJ), 8 (5), 2017. doi: 10.5121/sipij.2017.8501. Jie Han, Zehao Shen, Lingxiao Ying, Guixiang Li, and Anping Chen. Early post- re regeneration of a re-prone subtropical mixed Yunnan pine forest in Southwest China: E ects of pre- re vegetation, re severity and topographic factors. Forest Ecology and Management, 356(2015):31{40, 2015. ISSN 03781127. doi: 10.1016/j.foreco.2015.06.016. URL http://dx.doi.org/10.1016/j.foreco.2015.06. Lucas Harris and Alan H. Taylor. Previous burns and topography limit and reinforce re severity in a large wild re. Ecosphere, 8(11), 2017. ISSN 21508925. doi: 10.1002/ecs2.2019. Trevor Hastie, Jerome Friedman, and Robert Tibshirani. The Elements of Statistical Learning: Data Mining, Inference, and Prediction. Springer, New York, NY, 2009. URL https://doi.org/10.1007/ 978-0-387-21606-5. 47 Todd J. Hawbaker, Melanie K. Vanderhoof, Yen Ju Beal, Joshua D. Takacs, Gail L. Schmidt, Je T. Falgout, Brad Williams, Nicole M. Fairaux, Megan K. Caldwell, Joshua J. Picotte, Stephen M. Howard, Susan Stitt, and John L. Dwyer. Mapping burned areas using dense time-series of Landsat data. Remote Sensing of Environment, 198:504{522, sep 2017. ISSN 00344257. doi: 10.1016/j.rse.2017.06.027. Marti A. Hearst, Susan T Dumais, Edgar Osuna, John Platt, and Bernhard Scholkopf. Support vector machines. IEEE Intelligent Systems and their applications, 13(4):18{28, 1998. Robert Hecht-Nielsen. Theory of the backpropagation neural network. In Neural networks for perception, pages 65{93. Elsevier, 1992. Risto K. Heikkinen, Mathieu Marmion, and Miska Luoto. Does the interpolation accuracy of species distribution models come at the expense of transferability? Ecography, 35(3):276{288, mar 2012. ISSN 09067590. doi: 10.1111/j.1600-0587.2011.06999.x. URL http://doi.wiley.com/10.1111/j. 1600-0587.2011.06999.x. Txomin Hermosilla, Michael A. Wulder, Joanne C. White, Nicholas C. Coops, and Geordie W. Hobart. Regional detection, characterization, and attribution of annual forest change from 1984 to 2012 using Landsat-derived time-series metrics. Remote Sensing of Environment, 170:121{132, 2015. ISSN 00344257. doi: 10.1016/j.rse.2015.09.004. URL http://dx.doi.org/10.1016/j.rse.2015.09.004. Sepp Hochreiter and Jurgen  Schmidhuber. Long short-term memory. Neural computation, 9(8):1735{1780, Jonathan L. Hodges and Brian Y. Lattimer. Wildland Fire Spread Modeling Using Convolutional Neural Networks. Fire Technology, nov 2019. ISSN 15728099. doi: 10.1007/s10694-019-00846-4. C. M. Ho man, J. Can eld, R. R. Linn, W. Mell, C. H. Sieg, F. Pimont, and J. Ziegler. Evaluating Crown Fire Rate of Spread Predictions from Physics-Based Models. Fire Technology, 52(1):221{237, jan 2016. ISSN 15728099. doi: 10.1007/s10694-015-0500-3. Zachary A. Holden, Penelope Morgan, and Je rey S. Evans. A predictive model of burn severity based on 20-year satellite-inferred burn severity data in a large southwestern US wilderness area. Forest Ecology and Management, 258(11):2399{2406, 2009. ISSN 03781127. doi: 10.1016/j.foreco.2009.08.017. Baisravan Homchaudhuri, Sheng Zhao, Kelly Cohen, and Manish Kumar. Generation of optimal re-line for ghting wildland res using genetic algorithms. Proceedings of the ASME Dynamic Systems and Control Conference 2009, DSCC2009, (PART A):111{118, 2010. doi: 10.1115/DSCC2009-2707. Haoyuan Hong, Paraskevas Tsangaratos, Ioanna Ilia, Junzhi Liu, A-Xing Zhu, and Chong Xu. Applying genetic algorithms to set the optimal combination of forest re related variables and model forest re susceptibility based on data mining models. The case of Dayu County, China. Science of The Total Environment, 630:1044{1056, jul 2018. ISSN 0048-9697. doi: 10.1016/J.SCITOTENV.2018.02.278. URL https://www.sciencedirect.com/science/article/pii/S004896971830679X. F M Anim Hossain, Youmin Zhang, Chi Yuan, and Chun-Yi Su. Wild re Flame and Smoke Detection Using Static Image Features and Arti cial Neural Network. In 2019 1st International Conference on Industrial Arti cial Intelligence (IAI), pages 1{6. Institute of Electrical and Electronics Engineers (IEEE), oct 2019. doi: 10.1109/iciai.2019.8850811. Bronwyn A. Hradsky, Trent D. Penman, Dan Ababei, Anca Hanea, Euan G. Ritchie, Alan York, and Julian Di Stefano. Bayesian networks elucidate interactions between re and other drivers of terrestrial fauna distributions. Ecosphere, 8(8):e01926, aug 2017. ISSN 21508925. doi: 10.1002/ecs2.1926. URL http://doi.wiley.com/10.1002/ecs2.1926. 48 Carolynne Hultquist, Gang Chen, and Kaiguang Zhao. A comparison of Gaussian process regression, random forests and support vector regression for burn severity assessment in diseased forests. Re- mote Sensing Letters, 5(8):723{732, aug 2014. ISSN 2150-704X. doi: 10.1080/2150704X.2014.963733. URL http://dx.doi.org/10.1080/2150704X.2014.963733http://www.tandfonline.com/doi/abs/ 10.1080/2150704X.2014.963733. L.S. Iliadis. A decision support system applying an integrated fuzzy model for long-term forest re risk estimation. Environmental Modelling & Software, 20(5):613{621, may 2005. ISSN 1364-8152. doi: 10.1016/J.ENVSOFT.2004.03.006. URL https://www.sciencedirect.com/science/article/ pii/S1364815204000933. Abolfazl Jaafari. Factors In uencing Regional-Scale Wild re Probability in Iran: An Applica- tion of Random Forest and Support Vector Machine. Spatial Modeling in GIS and R for Earth and Environmental Sciences, pages 607{619, jan 2019. doi: 10.1016/B978-0-12-815226-3. 00028-4. URL https://www-sciencedirect-com.login.ezproxy.library.ualberta.ca/science/ article/pii/B9780128152263000284. Abolfazl Jaafari, Eric K. Zenner, and Binh Thai Pham. Wild re spatial pattern analysis in the Za- gros Mountains, Iran: A comparative study of decision tree based classi ers. Ecological Informat- ics, 43:200{211, jan 2018. ISSN 1574-9541. doi: 10.1016/J.ECOINF.2017.12.006. URL https: //www.sciencedirect.com/science/article/pii/S157495411730167X. Abolfazl Jaafari, Eric K. Zenner, Mahdi Panahi, and Himan Shahabi. Hybrid arti cial intelligence mod- els based on a neuro-fuzzy system and metaheuristic optimization algorithms for spatial prediction of wild re probability. Agricultural and Forest Meteorology, 266-267:198{207, mar 2019. ISSN 01681923. doi: 10.1016/j.agrformet.2018.12.015. URL https://www.sciencedirect.com/science/article/pii/ S0168192318304088. Jacek Jakubowski, Maciej Solarczyk, and Micha l Wi snios. Smoke detection in a digital image with the use of convolutional network. page 14. SPIE-Intl Soc Optical Eng, mar 2019. ISBN 9781510627857. doi: 10.1117/12.2524560. J. . R. Jang. An s: adaptive-network-based fuzzy inference system. IEEE Transactions on Systems, Man, and Cybernetics, 23(3):665{685, May 1993. doi: 10.1109/21.256541. Maria Jo~ ao Sousa, Alexandra Moutinho, and Miguel Almeida. Wild re detection using transfer learning on augmented datasets. Expert Systems with Applications, page 112975, sep 2019. ISSN 09574174. doi: 10.1016/j.eswa.2019.112975. Torres Jo~ ao, Gon calves Jo~ ao, Marcos Bruno, and Honrado Jo~ ao. Indicator-based assessment of post- re re- covery dynamics using satellite NDVI time-series. Ecological Indicators, 89(January):199{212, 2018. ISSN 1470160X. doi: 10.1016/j.ecolind.2018.02.008. URL https://doi.org/10.1016/j.ecolind.2018.02. Jill F. Johnstone, Teresa N. Hollingsworth, F. Stuart Chapin, and Michelle C. Mack. Changes in re regime break the legacy lock on successional trajectories in Alaskan boreal forest. Global Change Biology, 16 (4):1281{1295, 2010. ISSN 13541013. doi: 10.1111/j.1365-2486.2009.02051.x. Kyle D. Julian and Mykel J. Kochenderfer. Autonomous distributed wild re surveillance using deep reinforcement learning. In AIAA Guidance, Navigation, and Control Conference, 2018, number 210039. American Institute of Aeronautics and Astronautics Inc, AIAA, jan 2018a. ISBN 9781624105265. doi: 10.2514/6.2018-1589. 49 Kyle D. Julian and Mykel J. Kochenderfer. Distributed Wild re Surveillance with Autonomous Aircraft using Deep Reinforcement Learning. Journal of Guidance, Control, and Dynamics, 42(8):1768{1778, oct 2018b. URL http://arxiv.org/abs/1810.04244. Martin Jung, Susanne Tautenhahn, Christian Wirth, and Jens Kattge. Estimating Basal Area of Spruce and Fir in Post- re Residual Stands in Central Siberia Using Quickbird, Feature Selection, and Random Forests. Procedia Computer Science, 18:2386{2395, jan 2013. ISSN 1877-0509. doi: 10.1016/J.PROCS. 2013.05.410. URL https://www.sciencedirect.com/science/article/pii/S187705091300553X. M. Njoki Kahiu and N. P. Hanan. Fire in sub-Saharan Africa: The fuel, cure and connectivity hypothesis. Global Ecology and Biogeography, 27(8):946{957, aug 2018. ISSN 1466822X. doi: 10.1111/geb.12753. URL http://doi.wiley.com/10.1111/geb.12753. Van R. Kane, C. Alina Cansler, Nicholas A. Povak, Jonathan T. Kane, Robert J. McGaughey, James A. Lutz, Derek J. Churchill, and Malcolm P. North. Mixed severity re e ects within the Rim re: Relative importance of local climate, re weather, topography, and forest structure. Forest Ecology and Management, 358:62{79, dec 2015. ISSN 0378-1127. doi: 10.1016/J.FORECO.2015.09.001. URL https://www.sciencedirect.com/science/article/pii/S0378112715004697. Anuj Karpatne, Imme Ebert-Upho , Sai Ravela, Hassan Ali Babaie, and Vipin Kumar. Machine Learning for the Geosciences: Challenges and Opportunities. nov 2017. URL http://arxiv.org/abs/1711. Robert E. Keane, Geo rey J. Cary, Ian D. Davies, Michael D. Flannigan, Robert H. Gardner, Sandra Lavorel, James M. Lenihan, Chao Li, and T. Scott Rupp. A classi cation of landscape re succession models: Spatial simulations of re and vegetation dynamics. Ecological Modelling, 179(1-2):3{27, nov 2004. ISSN 03043800. doi: 10.1016/j.ecolmodel.2004.03.015. Nima Khakzad. Modeling wild re spread in wildland-industrial interfaces using dynamic Bayesian network. Reliability Engineering & System Safety, 189:165{176, sep 2019. ISSN 0951-8320. doi: 10.1016/J.RESS. 2019.04.006. URL https://www.sciencedirect.com/science/article/pii/S0951832018313887. Sea Jin Kim, Chul-Hee Lim, Gang Sun Kim, Jongyeol Lee, Tobias Geiger, Omid Rahmati, Yowhan Son, and Woo-Kyun Lee. Multi-Temporal Analysis of Forest Fire Probability Using Socio-Economic and Environmental Variables. Remote Sensing, 11(1):86, jan 2019. ISSN 2072-4292. doi: 10.3390/rs11010086. URL https://www.mdpi.com/2072-4292/11/1/86. Seongchan Kim, Seungkyun Hong, Minsu Joh, and Sa-kwang Song. DeepRain: ConvLSTM Network for Precipitation Prediction using Multichannel Radar Data. nov 2017. URL http://arxiv.org/abs/ 1711.02316. ByoungChul Ko, Kwang-Ho Cheong, and Jae-Yeal Nam. Early re detection algorithm based on irreg- ular patterns of ames and hierarchical Bayesian Networks. Fire Safety Journal, 45(4):262{270, jun 2010. ISSN 0379-7112. doi: 10.1016/J.FIRESAF.2010.04.001. URL https://www.sciencedirect.com/ science/article/pii/S0379711210000378. Alexandru Korotcov, Valery Tkachenko, Daniel P. Russo, and Sean Ekins. Comparison of Deep Learning with Multiple Machine Learning Methods and Metrics Using Diverse Drug Discovery Data Sets. Molec- ular Pharmaceutics, 14(12):4462{4475, dec 2017. ISSN 15438392. doi: 10.1021/acs.molpharmaceut. 7b00578. Peter Kourtz. Arti cial intelligence: a new tool for forest management. Canadian Journal of For- est Research, 20(4):428{437, apr 1990. ISSN 0045-5067. doi: 10.1139/x90-060. URL http://www. nrcresearchpress.com/doi/10.1139/x90-060. 50 P.H. Kourtz. Arti cial intelligence applications in the next generation Canadian forest re control system, 1993. URL https://cfs.nrcan.gc.ca/publications?id=10775. V. I. Kozik, E. S. Nezhevenko, and A. S. Feoktistov. Adaptive prediction of forest re behavior on the basis of recurrent neural networks. Optoelectronics, Instrumentation and Data Processing, 49(3):250{259, may 2013. ISSN 8756-6990. doi: 10.3103/S8756699013030060. URL http://link.springer.com/10.3103/ S8756699013030060. V. I. Kozik, E. S. Nezhevenko, and A. S. Feoktistov. Studying the method of adaptive prediction of forest re evolution on the basis of recurrent neural networks. Optoelectronics, Instrumentation and Data Processing, 50(4):395{401, jul 2014. ISSN 8756-6990. doi: 10.3103/S8756699014040116. URL http://link.springer.com/10.3103/S8756699014040116. Max Kuhn and Kjell Johnson. Applied predictive modeling. Springer New York, jan 2013. ISBN 9781461468493. doi: 10.1007/978-1-4614-6849-3. Jan Kuka cka, Vladimir Golkov, and Daniel Cremers. Regularization for Deep Learning: A Taxonomy. oct 2017. URL http://arxiv.org/abs/1710.10686. Nataliia Kussul, Mykola Lavreniuk, Sergii Skakun, and Andrii Shelestov. Deep Learning Classi cation of Land Cover and Crop Types Using Remote Sensing Data. IEEE Geoscience and Remote Sensing Letters, 14(5):778{782, may 2017. ISSN 1545-598X. doi: 10.1109/LGRS.2017.2681128. URL http: //ieeexplore.ieee.org/document/7891032/. Ryan Lagerquist, Mike D. Flannigan, Xianli Wang, and Ginny A. Marshall. Automated prediction of extreme re weather from synoptic patterns in northern Alberta, Canada. Canadian Journal of Forest Research, 47(9):1175{1183, sep 2017. ISSN 0045-5067. doi: 10.1139/cjfr-2017-0063. URL http://www. nrcresearchpress.com/doi/10.1139/cjfr-2017-0063. Zachary Langford, Jitendra Kumar, and Forrest Ho man. Wild re mapping in interior alaska using deep neural networks on imbalanced datasets. In IEEE International Conference on Data Mining Workshops, ICDMW, volume 2018-Novem, pages 770{778. IEEE Computer Society, feb 2019. ISBN 9781538692882. doi: 10.1109/ICDMW.2018.00116. David J. Lary, Amir H. Alavi, Amir H. Gandomi, and Annette L. Walker. Machine learning in geosciences and remote sensing. Geoscience Frontiers, 7(1):3{10, jan 2016. ISSN 16749871. doi: 10.1016/j.gsf.2015. 07.003. URL https://www.sciencedirect.com/science/article/pii/S1674987115000821. Don J Latham. Arti cial Intelligence Applications To Fire Management. In Proceedings Of The Symposium On Wildland Fire, South Lake Tahoe, 1987. Christopher J. Lauer, Claire A. Montgomery, and Thomas G. Dietterich. Spatial interactions and opti- mal forest management on a re-threatened landscape. Forest Policy and Economics, 83:107{120, oct 2017. ISSN 1389-9341. doi: 10.1016/J.FORPOL.2017.07.006. URL https://www.sciencedirect.com/ science/article/pii/S1389934116304749. Yann LeCun, Yoshua Bengio, and Geo rey Hinton. Deep learning. nature, 521(7553):436, 2015. Michael Leuenberger, Joana Parente, Marj Tonini, M ario Gonzalez Pereira, and Mikhail Kanevski. Wild re susceptibility mapping: Deterministic vs. stochastic approaches. Environmental Modelling & Software, 101:194{203, mar 2018. ISSN 1364-8152. doi: 10.1016/J.ENVSOFT.2017.12.019. URL https://www. sciencedirect.com/science/article/pii/S1364815217303316. 51 Danielle Levac, Heather Colquhoun, and Kelly K O'Brien. Scoping studies: advanc- ing the methodology. Implementation science : IS, 5:69, sep 2010. ISSN 1748-5908. doi: 10.1186/1748-5908-5-69. URL http://www.ncbi.nlm.nih.gov/pubmed/20854677http://www. pubmedcentral.nih.gov/articlerender.fcgi?artid=PMC2954944. Berangere A. Leys, Julie L. Commerford, and Kendra K. McLauchlan. Reconstructing grassland re history using sedimentary charcoal: Considering count, size and shape. PLOS ONE, 12(4):e0176445, apr 2017. ISSN 1932-6203. doi: 10.1371/journal.pone.0176445. URL https://dx.plos.org/10.1371/journal. pone.0176445. Hanchao Li, Xiang Fei, and Chaobo He. Study on Most Important Factor and Most Vulnerable Location for a Forest Fire Case Using Various Machine Learning Techniques. 2018 Sixth International Conference on Advanced Cloud and Big Data (CBD), pages 298{303, 2018a. doi: 10.1109/CBD.2018.00060. URL https://ieeexplore.ieee.org/document/8530856/. Jin Li, Andrew D. Heap, Anna Potter, and James J. Daniell. Application of machine learning methods to spatial interpolation of environmental variables. Environmental Modelling and Software, 26(12):1647{ 1659, dec 2011. ISSN 13648152. doi: 10.1016/j.envsoft.2011.07.004. Li-Ming Li, Wei-Guo Song, Jian Ma, and Kohyu Satoh. Arti cial neural network approach for modeling the impact of population density and weather parameters on forest re risk. International Journal of Wildland Fire, 18(6):640, oct 2009. ISSN 1049-8001. doi: 10.1071/WF07136. URL http://www. publish.csiro.au/?paper=WF07136. Shufeng Li, Alice C. Hughes, Tao Su, Julie Lebreton Anberr ee, Alexei A. Oskolski, Mei Sun, David K. Ferguson, and Zhekun Zhou. Fire dynamics under monsoonal climate in Yunnan, SW China: past, present and future. Palaeogeography, Palaeoclimatology, Palaeoecology, 465:168{176, jan 2017. ISSN 0031-0182. doi: 10.1016/J.PALAEO.2016.10.028. URL https://www.sciencedirect.com/science/ article/pii/S0031018216306411. T. Li, E. Zhao, J. Zhang, and C. Hu. Detection of Wild re Smoke Images Based on a Densely Dilated Con- volutional Network. Electronics, 8(10):1131, oct 2019. ISSN 2079-9292. doi: 10.3390/electronics8101131. URL https://www.mdpi.com/2079-9292/8/10/1131. Xiaolian Li, Weiguo Song, Liping Lian, and Xiaoge Wei. Forest Fire Smoke Detection Using Back- Propagation Neural Network Based on MODIS Data. Remote Sensing, 7(4):4473{4498, apr 2015. ISSN 2072-4292. doi: 10.3390/rs70404473. URL http://www.mdpi.com/2072-4292/7/4/4473. Xiuqing Li, Zhenxue Chen, Q. M.J. Wu, and Chengyun Liu. 3D Parallel Fully Convolutional Networks for Real-time Video Wild re Smoke Detection, 2018b. ISSN 10518215. Hao Liang, Meng Zhang, and Hailan Wang. A Neural Network Model for Wild re Scale Prediction Using Meteorological Factors. IEEE Access, 7:176746{176755, 2019. ISSN 21693536. doi: 10.1109/ACCESS. 2019.2957837. Chul-Hee Lim, You Seung Kim, Myungsoo Won, Sea Jin Kim, and Woo-Kyun Lee. Can satellite-based data substitute for surveyed data to predict the spatial probability of forest re? A geostatistical approach to forest re in the Republic of Korea. Geomatics, Natural Hazards and Risk, 10(1):719{739, jan 2019. ISSN 1947-5705. doi: 10.1080/19475705.2018.1543210. URL https://www.tandfonline.com/doi/full/10. 1080/19475705.2018.1543210. Y Liu, Y Yang, C Liu, and Yu Gu. Forest Fire Detection Using Arti cial Neural Network Algorithm Imple- mented in Wireless Sensor Networks. ZTE Communications, jun 2015. URL http://wwwen.zte.com. cn/endata/magazine/ztecommunications/2015/2/articles/201507/t20150724{_}443252.html. 52 Zelin Liu, Changhui Peng, Timothy Work, Jean-Noel Candau, Annie DesRochers, and Daniel Kneeshaw. Application of machine-learning methods in forest ecology: recent progress and future challenges. En- vironmental Reviews, 26(4):339{350, dec 2018. ISSN 1181-8700. doi: 10.1139/er-2018-0034. URL http://www.nrcresearchpress.com/doi/10.1139/er-2018-0034. Zhihua Liu and Michael C. Wimberly. Climatic and Landscape In uences on Fire Regimes from 1984 to 2010 in the Western United States. PLOS ONE, 10(10):e0140839, oct 2015. ISSN 1932-6203. doi: 10.1371/journal.pone.0140839. URL http://dx.plos.org/10.1371/journal.pone.0140839. Zhihua Liu and Michael C. Wimberly. Direct and indirect e ects of climate change on projected future re regimes in the western United States. Science of The Total Environment, 542:65{75, jan 2016. ISSN 0048-9697. doi: 10.1016/J.SCITOTENV.2015.10.093. URL https://www.sciencedirect.com/ science/article/pii/S0048969715309098. Zhihua Liu and Jian Yang. Quantifying ecological drivers of ecosystem productivity of the early-successional boreal Larix gmelinii forest. Ecosphere, 5(7):art84, jul 2014. ISSN 2150-8925. doi: 10.1890/ES13-00372.1. URL http://doi.wiley.com/10.1890/ES13-00372.1. Zhihua Liu, Jian Yang, and Hong S. He. Identifying the Threshold of Dominant Controls on Fire Spread in a Boreal Forest Landscape of Northeast China. PLoS ONE, 8(1):e55618, jan 2013. ISSN 1932-6203. doi: 10.1371/journal.pone.0055618. URL https://dx.plos.org/10.1371/journal.pone.0055618. Pablito M. L opez-Serrano, Carlos A. L opez-S anchez, Juan G. Alvarez-Gonz alez, and Jorge Garc a- Guti errez. A Comparison of Machine Learning Techniques Applied to Landsat-5 TM Spectral Data for Biomass Estimation. Canadian Journal of Remote Sensing, 42(6):690{705, nov 2016. ISSN 0703- 8992. doi: 10.1080/07038992.2016.1217485. URL https://www.tandfonline.com/doi/full/10.1080/ 07038992.2016.1217485. F. Javier Lozano, S. Su arez-Seoane, M. Kelly, and E. Luis. A multi-scale approach for modeling re occurrence probability using satellite data and classi cation trees: A case study in a mountain- ous Mediterranean region. Remote Sensing of Environment, 112(3):708{719, mar 2008. ISSN 0034- 4257. doi: 10.1016/J.RSE.2007.06.006. URL https://www.sciencedirect.com/science/article/ pii/S003442570700243X. V Lozhkin, D Tarkhov, V Timofeev, O Lozhkina, and A Vasilyev. Di erential neural network ap- proach in information process for prediction of roadside air pollution by peat re. IOP Conference Series: Materials Science and Engineering, 158(1):012063, nov 2016. ISSN 1757-8981. doi: 10. 1088/1757-899X/158/1/012063. URL http://stacks.iop.org/1757-899X/158/i=1/a=012063?key= crossref.7abf8c3fd66f7ce48986b4554f7aecd5. Guilan Luo, Mei Zhang, Zizhong Yang, and Mingmei Song. Data mining of correlation between re disturbance habitat factors and spider communities. In 2017 4th International Conference on Systems and Informatics (ICSAI), pages 1471{1476. IEEE, nov 2017. ISBN 978-1-5386-1107-4. doi: 10.1109/ ICSAI.2017.8248518. URL http://ieeexplore.ieee.org/document/8248518/. Ruisen Luo, Yingying Dong, Muye Gan, Dejun Li, Shuli Niu, Amy Oliver, Ke Wang, and Yiqi Luo. Global Analysis of In uencing Forces of Fire Activity: the Threshold Relationships between Vegetation and Fire. Life Science Journal, 10(2):15{24, 2013. ISSN 0300-9165. Duncan C Lutes, Robert E Keane, John F Caratti, Carl H Key, Nathan C Benson, Steve Sutherland, and Larry J Gangi. FIREMON: Fire e ects monitoring and inventory system. Gen. Tech. Rep. RMRS- GTR-164. Fort Collins, CO: US Department of Agriculture, Forest Service, Rocky Mountain Research Station. 1 CD., 164, 2006. 53 Jamie M. Lydersen, Malcolm P. North, and Brandon M. Collins. Severity of an uncharacteristically large wild re, the Rim Fire, in forests with relatively restored frequent re regimes. Forest Ecology and Management, 328:326{334, sep 2014. ISSN 0378-1127. doi: 10.1016/J.FORECO.2014.06.005. URL https://www.sciencedirect.com/science/article/pii/S0378112714003661. Jamie M. Lydersen, Brandon M. Collins, Matthew L. Brooks, John R. Matchett, Kristen L. Shive, Nicholas A. Povak, Van R. Kane, and Douglas F. Smith. Evidence of fuels management and re weather in uencing re severity in an extreme re event. Ecological Applications, 27(7):2013{2030, oct 2017. ISSN 10510761. doi: 10.1002/eap.1586. URL http://doi.wiley.com/10.1002/eap.1586. James MacQueen et al. Some methods for classi cation and analysis of multivariate observations. In Proceedings of the fth Berkeley symposium on mathematical statistics and probability, volume 1, pages 281{297. Oakland, CA, USA, 1967. Eduardo Eiji Maeda, Antonio Roberto Formaggio, Yosio Edemir Shimabukuro, Gustavo Felipe Balu e Arcoverde, and Matthew C. Hansen. Predicting forest re in the Brazilian Amazon using MODIS imagery and arti cial neural networks. International Journal of Applied Earth Observation and Geoinformation, 11(4):265{272, aug 2009. ISSN 0303-2434. doi: 10.1016/J.JAG.2009.03.003. URL https://www.sciencedirect.com/science/article/pii/S0303243409000233. Nyasha Magadzire, Helen M. Klerk, Karen J. Esler, and Jasper A. Slingsby. Fire and life history a ect the distribution of plant species in a biodiversity hotspot. Diversity and Distributions, 25(7):1012{1023, jul 2019. ISSN 1366-9516. doi: 10.1111/ddi.12921. URL https://onlinelibrary.wiley.com/doi/abs/ 10.1111/ddi.12921. G. Mallinis, F. Maris, I. Kalinderis, and N. Koutsias. Assessment of Post- re Soil Erosion Risk in Fire- A ected Watersheds Using Remote Sensing and GIS. GIScience & Remote Sensing, 46(4):388{410, oct 2009. ISSN 1548-1603. doi: 10.2747/1548-1603.46.4.388. URL https://www.tandfonline.com/doi/ full/10.2747/1548-1603.46.4.388. Nicolas Mansuy, Carol Miller, Marc-Andr e Parisien, Sean A Parks, Enric Batllori, and Max A Moritz. Con- trasting human in uences and macro-environmental factors on re activity inside and outside protected areas of North America. Environmental Research Letters, 14(6):064007, may 2019. ISSN 1748-9326. doi: 10.1088/1748-9326/ab1bc5. URL https://iopscience.iop.org/article/10.1088/1748-9326/ ab1bc5. Natasha Markuzon and Stephan Kolitz. Data driven approach to estimating re danger from satellite images and weather information. In 2009 IEEE Applied Imagery Pattern Recognition Workshop (AIPR 2009), pages 1{7. IEEE, oct 2009. ISBN 978-1-4244-5146-3. doi: 10.1109/AIPR.2009.5466309. URL http://ieeexplore.ieee.org/document/5466309/. David L. Martell. A review of recent forest and wildland re management decision support systems research, jun 2015. ISSN 21986436. Yago Mart n, Mar a Zu niga-An ~ t on, and Marcos Rodrigues Mimbrero. Modelling temporal variation of re-occurrence towards the dynamic prediction of human wild re ignition danger in northeast Spain. Geomatics, Natural Hazards and Risk, 10(1):385{411, jan 2019. ISSN 1947-5705. doi: 10.1080/19475705. 2018.1526219. URL https://www.tandfonline.com/doi/full/10.1080/19475705.2018.1526219. S. Mart n-Alc on and L Coll. Unraveling the relative importance of factors driving post- re regenera- tion trajectories in non-serotinous Pinus nigra forests. Forest Ecology and Management, 361:13{22, feb 2016. ISSN 0378-1127. doi: 10.1016/J.FORECO.2015.11.006. URL https://www-sciencedirect-com. login.ezproxy.library.ualberta.ca/science/article/pii/S037811271500612X. 54 Arif Masrur, Andrey N. Petrov, and John DeGroote. Circumpolar spatio-temporal patterns and contribut- ing climatic factors of wild re activity in the Arctic tundra from 2001-2015. Environmental Research Letters, 13(1), jan 2018. ISSN 17489326. doi: 10.1088/1748-9326/aa9a76. Robert Mavsar, Armando Gonz alez Cab an, and Elsa Varela. The state of development of re management decision support systems in America and Europe. Forest Policy and Economics, 29:45{55, apr 2013. ISSN 13899341. doi: 10.1016/j.forpol.2012.11.009. R. Stockton Maxwell, Alan H. Taylor, Carl N. Skinner, Hugh D. Sa ord, Rachel E. Isaacs, Catherine Airey, and Amanda B. Young. Landscape-scale modeling of reference period forest conditions and re behavior on heavily logged lands. Ecosphere, 5(3):art32, mar 2014. ISSN 2150-8925. doi: 10.1890/ES13-00294.1. URL http://doi.wiley.com/10.1890/ES13-00294.1. M.J. Mayr, K.A. Vanselow, and C. Samimi. Fire regimes at the arid fringe: A 16-year remote sensing perspective (2000{2016) on the controls of re activity in Namibia from spatial predictive models. Eco- logical Indicators, 91:324{337, aug 2018. ISSN 1470-160X. doi: 10.1016/J.ECOLIND.2018.04.022. URL https://www.sciencedirect.com/science/article/pii/S1470160X18302759. Ronald J Mccormick, Thomas A Brandner, and Timothy F H Allen. TOWARD A THEORY OF MESO- SCALE WILDFIRE MODELING-A COMPLEX SYSTEMS APPROACH USING ARTIFICIAL NEU- RAL NETWORKS. In The Joint Fire Science Conference And Workshop, Boise, Idaho, 1999. Amy McGovern, Kimberly L. Elmore, David John Gagne, Sue Ellen Haupt, Christopher D. Karstens, Ryan Lagerquist, Travis Smith, John K. Williams, Amy McGovern, Kimberly L. Elmore, David John Gagne II, Sue Ellen Haupt, Christopher D. Karstens, Ryan Lagerquist, Travis Smith, and John K. Williams. Using Arti cial Intelligence to Improve Real-Time Decision-Making for High-Impact Weather. Bulletin of the American Meteorological Society, 98(10):2073{2090, oct 2017. ISSN 0003-0007. doi: 10.1175/ BAMS-D-16-0123.1. URL http://journals.ametsoc.org/doi/10.1175/BAMS-D-16-0123.1. Amy McGovern, Ryan Lagerquist, David John Gagne, G. Eli Jergensen, Kimberly L. Elmore, Cameron R. Homeyer, and Travis Smith. Making the Black Box More Transparent: Understanding the Physical Implications of Machine Learning. Bulletin of the American Meteorological Society, 100(11):2175{2199, nov 2019. ISSN 0003-0007. doi: 10.1175/BAMS-D-18-0195.1. URL http://journals.ametsoc.org/ doi/10.1175/BAMS-D-18-0195.1. Sean McGregor, Rachel Houtman, Hailey Buckingham, Claire Montgomery, Ronald Metoyer, and Thomas G Dietterich. Fast simulation for computational sustainability sequential decision making prob- lems. In Proceedings of the 4th International Conference on Computational Sustainability, pages 5{7, Sean McGregor, Rachel Houtman, Claire Montgomery, Ronald Metoyer, and Thomas G. Dietterich. Fast Optimization of Wild re Suppression Policies with SMAC. arXiv preprint, mar 2017. URL http: //arxiv.org/abs/1703.09391. James P. Minas, John W. Hearne, and John W. Handmer. A review of operations research methods applicable to wild re management. International Journal of Wildland Fire, 21(3):189, may 2012. ISSN 1049-8001. doi: 10.1071/WF10129. URL http://www.publish.csiro.au/?paper=WF10129. Yosune Miquelajauregui, Steven G. Cumming, and Sylvie Gauthier. Modelling Variable Fire Severity in Boreal Forests: E ects of Fire Intensity and Stand Structure. PLOS ONE, 11(2):e0150073, feb 2016. ISSN 1932-6203. doi: 10.1371/journal.pone.0150073. URL https://dx.plos.org/10.1371/journal. pone.0150073. Melanie Mitchell. An introduction to genetic algorithms. MIT Press, 1996. ISBN 9780262133166. 55 T.M. Mitchell. Machine Learning. McGraw-HIll, 1997. ISBN 0071154671. Varun Mithal, Guruprasad Nayak, Ankush Khandelwal, Vipin Kumar, Ramakrishna Nemani, and Nikunj Oza. Mapping burned areas in tropical forests using a novel machine learning framework. Remote Sensing, 10(1):69, 2018. Nikolaos E. Mitrakis, Giorgos Mallinis, Nikos Koutsias, and John B. Theocharis. Burned area mapping in Mediterranean environment using medium-resolution multi-spectral data and a neuro-fuzzy classi er. International Journal of Image and Data Fusion, 3(4):299{318, dec 2012. ISSN 1947-9832. doi: 10. 1080/19479832.2011.635604. URL http://www.tandfonline.com/doi/abs/10.1080/19479832.2011. Ioannis Mitsopoulos and Giorgos Mallinis. A data-driven approach to assess large re size generation in Greece. Natural Hazards, 88(3):1591{1607, sep 2017. ISSN 0921-030X. doi: 10.1007/s11069-017-2934-z. URL http://link.springer.com/10.1007/s11069-017-2934-z. J. R. Molina, A. Lora, C. Prades, and F. Rodr guez y Silva. Roadside vegetation planning and con- servation: New approach to prevent and mitigate wild res based on re ignition potential. Forest Ecology and Management, 444:163{173, jul 2019. ISSN 0378-1127. doi: 10.1016/J.FORECO.2019. 04.034. URL https://www-sciencedirect-com.login.ezproxy.library.ualberta.ca/science/ article/pii/S0378112719301501. Max A. Moritz, Marc-Andr e Parisien, Enric Batllori, Meg A. Krawchuk, Je Van Dorn, David J. Ganz, and Katharine Hayhoe. Climate change and disruptions to global re activity. Ecosphere, 3(6):art49, jun 2012. ISSN 2150-8925. doi: 10.1890/ES11-00345.1. URL http://doi.wiley.com/10.1890/ES11-00345.1. Amir Mosavi, Pinar Ozturk, Kwok-wing Chau, Amir Mosavi, Pinar Ozturk, and Kwok-wing Chau. Flood Prediction Using Machine Learning Models: Literature Review. Water, 10(11):1536, oct 2018. ISSN 2073-4441. doi: 10.3390/w10111536. URL http://www.mdpi.com/2073-4441/10/11/1536. Mohsen MOSTAFA, Shaban SHATAEE JOUIBARY, Majid LOTFALIAN, and Amir SADODDIN. WA- TERSHED ROAD NETWORK ANALYSIS WITH AN EMPHASIS ON FIRE FIGHTING MANAGE- MENT. Journal of Environmental Engineering and Landscape Management, 25(4):342{353, dec 2017. ISSN 1648-6897. doi: 10.3846/16486897.2017.1281816. URL http://journals.vgtu.lt/index.php/ JEELM/article/view/1712. Kudzai S. Mpakairi, Paradzayi Tagwireyi, Henry Ndaimani, and Hilary T. Madiri. Distribution of wild- land res and possible hotspots for the Zimbabwean component of Kavango-Zambezi Transfrontier Con- servation Area. South African Geographical Journal, 101(1):110{120, jan 2019. ISSN 0373-6245. doi: 10.1080/03736245.2018.1541023. URL https://www.tandfonline.com/doi/full/10.1080/03736245. 2018.1541023. Khan Muhammad, Jamil Ahmad, and Sung Wook Baik. Early re detection using convolutional neural networks during surveillance for e ective disaster management. Neurocomputing, 288:30{42, may 2018. ISSN 18728286. doi: 10.1016/j.neucom.2017.04.083. Kevin Murphy. Machine Learning: A Probabilistic Perspective. MIT Press, 2012. URL http://www. amazon.com/Machine-Learning-Probabilistic-Perspective-Computation/dp/0262018020. Khurram Nadeem, S. W. Taylor, Douglas G. Woolford, and C. B. Dean. Mesoscale spatiotemporal pre- dictive models of daily human- and lightning-caused wildland re occurrence in British Columbia. In- ternational Journal of Wildland Fire, 29(1):11, 2020. ISSN 1049-8001. doi: 10.1071/WF19058. URL http://www.publish.csiro.au/?paper=WF19058. 56 Hariharan Naganathan, Sudarshan P Seshasayee, Jonghoon Kim, Wai K Chong, and Jui-Sheng Chou. Wild re Predictions: Determining Reliable Models using Fused Dataset. Global Journal of Computer Science and Technology, 16(4C), 2016. URL https://computerresearch.org/index.php/computer/ article/view/1437. Nicholas J. Nauslar, Benjamin J. Hatchett, Timothy J. Brown, Michael L. Kaplan, and John F. Mejia. Impact of the North American monsoon on wild re activity in the southwest United States. International Journal of Climatology, 39(3):1539{1554, mar 2019. ISSN 08998418. doi: 10.1002/joc.5899. URL http://doi.wiley.com/10.1002/joc.5899. Trisalyn A. Nelson, Wiebe Nijland, Mathieu L. Bourbonnais, and Michael A. Wulder. Regression Tree Modeling of Spatial Pattern and Process Interactions. In Mapping Forest Landscape Patterns, pages 187{212. Springer New York, New York, NY, 2017. doi: 10.1007/978-1-4939-7331-6 5. URL http: //link.springer.com/10.1007/978-1-4939-7331-6{_}5. Nguyen Ngoc Thach, Dang Bao-Toan Ngo, Pham Xuan-Canh, Nguyen Hong-Thi, Bui Hang Thi, Hoang Nhat-Duc, and Tien Bui Dieu. Spatial pattern assessment of tropical forest re danger at Thuan Chau area (Vietnam) using GIS-based advanced machine learning algorithms: A comparative study. Ecological Informatics, 46:74{85, jul 2018. ISSN 1574-9541. doi: 10.1016/J.ECOINF.2018.05.009. URL https: //www.sciencedirect.com/science/article/pii/S1574954118300852. I. Nitze, G. Grosse, B. M. Jones, V. E. Romanovsky, and J. Boike. Remote sensing quanti es widespread abundance of permafrost region disturbances across the Arctic and Subarctic. Nature Communications, 9(1), dec 2018. ISSN 20411723. doi: 10.1038/s41467-018-07663-3. Christopher D. O' O'Connor, David E. Calkin, Matthew P. Thompson, Christopher D.O. O'Connor, David E. Calkin, and Matthew P. Thompson. An empirical machine learning method for predict- ing potential re control locations for pre- re planning and operational re management. Interna- tional Journal of Wildland Fire, 26(7):587, jul 2017. ISSN 1049-8001. doi: 10.1071/WF16135. URL http://www.publish.csiro.au/?paper=WF16135. Julian D. Olden, Joshua J. Lawler, and N. LeRoy Po . Machine Learning Methods Without Tears: A Primer for Ecologists. The Quarterly Review of Biology, 83(2):171{193, jun 2008. ISSN 0033-5770. doi: 10.1086/587826. URL https://www.journals.uchicago.edu/doi/10.1086/587826. Sandra Oliveira, Friderike Oehler, Jesus  San-Miguel-Ayanz, Andrea Camia, and Jos e M.C. Pereira. Modeling spatial patterns of re occurrence in Mediterranean Europe using Multiple Regression and Random Forest. Forest Ecology and Management, 275:117{129, jul 2012. ISSN 0378-1127. doi: 10.1016/J.FORECO.2012.03.003. URL https://www.sciencedirect.com/science/article/ pii/S0378112712001272. A. Murat Ozbayo glu and Recep Bozer. Estimation of the Burned Area in Forest Fires Using Compu- tational Intelligence Techniques. Procedia Computer Science, 12:282{287, jan 2012. ISSN 1877-0509. doi: 10.1016/J.PROCS.2012.09.070. URL https://www.sciencedirect.com/science/article/pii/ S1877050912006618. P. Papakosta, G. Xanthopoulos, and D. Straub. Probabilistic prediction of wild re economic losses to housing in Cyprus using Bayesian network analysis. International Journal of Wildland Fire, 26(1):10, feb 2017. ISSN 1049-8001. doi: 10.1071/WF15113. URL http://www.publish.csiro.au/?paper=WF15113. Marc-Andr e Parisien and Max A. Moritz. Environmental controls on the distribution of wild re at multiple spatial scales. Ecological Monographs, 79(1):127{154, feb 2009. ISSN 0012-9615. doi: 10.1890/07-1289.1. URL http://doi.wiley.com/10.1890/07-1289.1. 57 Marc-Andr e Parisien, Sean A. Parks, Meg A. Krawchuk, John M. Little, Mike D. Flannigan, Lynn M. Gowman, and Max A. Moritz. An analysis of controls on re activity in boreal Canada: comparing models built with di erent temporal resolutions. Ecological Applications, 24(6):1341{1356, sep 2014. ISSN 1051-0761. doi: 10.1890/13-1477.1. URL http://doi.wiley.com/10.1890/13-1477.1. Marc And r Parisien, Susan Snetsinger, Jonathan A. Greenberg, Cara R. Nelson, Tania Schoennagel, Solomon Z. Dobrowski, and Max A. Moritz. Spatial variability in wild re probability across the west- ern United States. International Journal of Wildland Fire, 21(4):313{327, 2012. ISSN 10498001. doi: 10.1071/WF11044. Sean A. Parks, Carol Miller, Marc-Andr e Parisien, Lisa M. Holsinger, Solomon Z. Dobrowski, and John Abatzoglou. Wildland re de cit and surplus in the western United States, 1984{2012. Ecosphere, 6 (12):art275, dec 2015. ISSN 2150-8925. doi: 10.1890/ES15-00294.1. URL http://doi.wiley.com/10. 1890/ES15-00294.1. Sean A Parks, Carol Miller, John T Abatzoglou, Lisa M Holsinger, Marc-Andr e Parisien, and Solomon Z Dobrowski. How will climate change a ect wildland re severity in the west- ern US? Environmental Research Letters, 11(3):035002, mar 2016. ISSN 1748-9326. doi: 10.1088/1748-9326/11/3/035002. URL http://stacks.iop.org/1748-9326/11/i=3/a=035002?key= crossref.4d33abcb068f5458baf3b94828ca073e. Sean A Parks, Lisa M Holsinger, Matthew H Panunto, W Matt Jolly, Solomon Z Dobrowski, and Gregory K Dillon. High-severity re: evaluating its key drivers and mapping its probability across western US forests. Environmental Research Letters, 13(4):044037, apr 2018. ISSN 1748-9326. doi: 10.1088/1748-9326/aab791. URL http://stacks.iop.org/1748-9326/13/i=4/a=044037?key= crossref.5c2b6b1d5870d4a9269af3badf873e81. Judea Pearl. Probabilistic Reasoning in Intelligent Systems:Networks of Plausible Inference. Morgan Kauf- mann, San Mateo, California, 1988. T. D. Penman, O. Price, and R. A. Bradstock. Bayes Nets as a method for analysing the in uence of management actions in re planning. International Journal of Wildland Fire, 20(8):909{920, 2011. ISSN 10498001. doi: 10.1071/WF10076. T. D. Penman, R. A. Bradstock, and O. F. Price. Reducing wild re risk to urban developments: Simulation of cost-e ective fuel treatment solutions in south eastern Australia. Environmental Mod- elling and Software, 52:166{175, 2014. ISSN 13648152. doi: 10.1016/j.envsoft.2013.09.030. URL http://dx.doi.org/10.1016/j.envsoft.2013.09.030. T. D. Penman, A. E. Nicholson, R. A. Bradstock, L. Collins, S. H. Penman, and O. F. Price. Reducing the risk of house loss due to wild res. Environmental Modelling and Software, 67:12{25, 2015. ISSN 13648152. doi: 10.1016/j.envsoft.2014.12.020. Allan Pereira, Jos e Pereira, Renata Libonati, Duarte Oom, Alberto Setzer, Fabiano Morelli, Fausto Machado-Silva, Luis de Carvalho, Allan A. Pereira, Jos e M. C. Pereira, Renata Libonati, Duarte Oom, Alberto W. Setzer, Fabiano Morelli, Fausto Machado-Silva, and Luis Marcelo Tavares de Carvalho. Burned Area Mapping in the Brazilian Savanna Using a One-Class Support Vector Machine Trained by Active Fires. Remote Sensing, 9(11):1161, nov 2017. ISSN 2072-4292. doi: 10.3390/rs9111161. URL http://www.mdpi.com/2072-4292/9/11/1161. George L. W. Perry, Janet M. Wilmshurst, Matt S. McGlone, and Aaron Napier. Reconstructing spatial vulnerability to forest loss by re in pre-historic New Zealand. Global Ecology and Biogeography, 21 (10):1029{1041, oct 2012. ISSN 1466822X. doi: 10.1111/j.1466-8238.2011.00745.x. URL http://doi. wiley.com/10.1111/j.1466-8238.2011.00745.x. 58 Matthew P. Peters and Louis R. Iverson. Incorporating ne-scale drought information into an eastern US wild re hazard model. International Journal of Wildland Fire, 26(5):393, may 2017. ISSN 1049-8001. doi: 10.1071/WF16130. URL http://www.publish.csiro.au/?paper=WF16130. Matthew P. Peters, Louis R. Iverson, Stephen N. Matthews, and Anantha M. Prasad. Wild re hazard mapping: exploring site conditions in eastern US wildland{urban interfaces. International Journal of Wildland Fire, 22(5):567, aug 2013. ISSN 1049-8001. doi: 10.1071/WF12177. URL http://www. publish.csiro.au/?paper=WF12177. G. P. Petropoulos, W. Knorr, M. Scholze, L. Boschetti, and G. Karantounias. Combining ASTER multispectral imagery analysis and support vector machines for rapid and cost-e ective post- re as- sessment: a case study from the Greek wildland res of 2007. Natural Hazards and Earth Sys- tem Science, 10(2):305{317, feb 2010. ISSN 1684-9981. doi: 10.5194/nhess-10-305-2010. URL http://www.nat-hazards-earth-syst-sci.net/10/305/2010/. George P. Petropoulos, Charalambos Kontoes, and Iphigenia Keramitsoglou. Burnt area delineation from a uni-temporal perspective based on Landsat TM imagery classi cation using Support Vector Ma- chines. International Journal of Applied Earth Observation and Geoinformation, 13(1):70{80, feb 2011. ISSN 0303-2434. doi: 10.1016/J.JAG.2010.06.008. URL https://www.sciencedirect.com/science/ article/pii/S0303243410000784. M.T. Pham, A. Raji c, J.D. Greig, J.M. Sargeant, A. Papadopoulos, and S.A. Mcewen. A scoping review of scoping reviews: Advancing the approach and enhancing the consistency. Research Synthesis Methods, 5(4):371{385, dec 2014. ISSN 17592887. doi: 10.1002/jrsm.1123. Thanh Cong Phan and Thanh Tam Nguyen. Remote Sensing meets Deep Learning: Exploiting Spatio- Temporal-Spectral Satellite Images for Early Wild re Detection. Technical report, 2019. URL https: //infoscience.epfl.ch/record/270339. Sharon B. Phillips, Viney P. Aneja, Daiwen Kang, and S. Pal Arya. Modelling and analysis of the atmo- spheric nitrogen deposition in North Carolina. In International Journal of Global Environmental Issues, volume 6, pages 231{252, 2006. doi: 10.1016/j.ecolmodel.2005.03.026. Andrew D. Pierce, Calvin A. Farris, and Alan H. Taylor. Use of random forests for modeling and mapping forest canopy fuels for re behavior analysis in Lassen Volcanic National Park, California, USA. Forest Ecology and Management, 279:77{89, sep 2012. ISSN 0378-1127. doi: 10.1016/J.FORECO.2012.05.010. URL https://www.sciencedirect.com/science/article/pii/S0378112712002654. David L Poole and Alan K Mackworth. Arti cial Intelligence: foundations of computational agents. Cam- bridge University Press, 2010. Patrick Poon, Alicia Kinoshita, Patrick K. Poon, and Alicia M. Kinoshita. Estimating Evapotranspiration in a Post-Fire Environment Using Remote Sensing and Machine Learning. Remote Sensing, 10(11):1728, nov 2018. ISSN 2072-4292. doi: 10.3390/rs10111728. URL http://www.mdpi.com/2072-4292/10/11/ Zohre Sadat Pourtaghi, Hamid Reza Pourghasemi, Roberta Aretano, and Teodoro Semeraro. Investigation of general indicators in uencing on forest re and its susceptibility modeling using di erent data mining techniques. Ecological Indicators, 64:72{84, may 2016. ISSN 1470-160X. doi: 10.1016/J.ECOLIND. 2015.12.030. URL https://www.sciencedirect.com/science/article/pii/S1470160X15007359. Ruiliang Pu and Peng Gong. Determination of Burnt Scars Using Logistic Regression and Neural Network Techniques from a Single Post-Fire Landsat 7 ETM + Image. Photogram- metric Engineering & Remote Sensing, 70(7):841{850, jul 2004. ISSN 00991112. doi: 10. 59 14358/PERS.70.7.841. URL http://openurl.ingenta.com/content/xref?genre=article{&}issn= 0099-1112{&}volume=70{&}issue=7{&}spage=841. J Ross Quinlan. C 4.5: Programs for machine learning. The Morgan Kaufmann Series in Machine Learning, San Mateo, CA: Morgan Kaufmann,| c1993, 1993. Carmen Quintano, Alfonso Fern andez-Manso, Leonor Calvo, and Dar A. Roberts. Vegetation and Soil Fire Damage Analysis Based on Species Distribution Modeling Trained with Multispectral Satellite Data. Remote Sensing, 11(15):1832, aug 2019. ISSN 2072-4292. doi: 10.3390/rs11151832. URL https: //www.mdpi.com/2072-4292/11/15/1832. Natalia Quintero, Olga Viedma, Itziar R. Urbieta, and Jos e M. Moreno. Assessing Landscape Fire Hazard by Multitemporal Automatic Classi cation of Landsat Time Series Using the Google Earth Engine in West-Central Spain. Forests, 10(6):518, jun 2019. ISSN 1999-4907. doi: 10.3390/f10060518. URL https://www.mdpi.com/1999-4907/10/6/518. David Radke, Anna Hessler, and Dan Ellsworth. FireCast: Leveraging Deep Learning to Predict Wild re Spread. In Proceedings of the Twenty-Eighth International Joint Conference on Arti cial Intelligence (IJCAI-19), pages 4575{4581. International Joint Conferences on Arti cial Intelligence, jul 2019. doi: 10.24963/ijcai.2019/636. M. Raissi, P. Perdikaris, and G. E. Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial di erential equations. Journal of Computational Physics, 378:686{707, feb 2019. ISSN 10902716. doi: 10.1016/j.jcp.2018.10. Maziar Raissi and George Em Karniadakis. Hidden physics models: Machine learning of nonlinear partial di erential equations. Journal of Computational Physics, 357:125{141, mar 2018. ISSN 10902716. doi: 10.1016/j.jcp.2017.11.039. Carl Edward Rasmussen and Christopher KI Williams. Gaussian processes for machine learning, volume 1. MIT press Cambridge, 2006. Stephan Rasp and Sebastian Lerch. Neural Networks for Postprocessing Ensemble Weather Forecasts. Monthly Weather Review, 146(11):3885{3900, nov 2018. ISSN 0027-0644. doi: 10.1175/MWR-D-18-0187. 1. URL http://journals.ametsoc.org/doi/10.1175/MWR-D-18-0187.1. Markus Reichstein, Gustau Camps-Valls, Bjorn Stevens, Martin Jung, Joachim Denzler, Nuno Carvalhais, and Prabhat. Deep learning and process understanding for data-driven Earth system science. Nature, 566(7743):195{204, feb 2019. ISSN 0028-0836. doi: 10.1038/s41586-019-0912-1. URL http://www. nature.com/articles/s41586-019-0912-1. Colleen E. Reid, Michael Jerrett, Maya L. Petersen, Gabriele G. P ster, Philip E. More eld, Ira B. Tager, Sean M. Ra use, and John R. Balmes. Spatiotemporal Prediction of Fine Particulate Matter During the 2008 Northern California Wild res Using Machine Learning. Environmental Science & Technology, 49 (6):3887{3896, mar 2015. ISSN 0013-936X. doi: 10.1021/es505846r. URL http://pubs.acs.org/doi/ 10.1021/es505846r. Quentin Renard, Rapha el P elissier, B. R. Ramesh, and Narendran Kodandapani. Environmental sus- ceptibility model for predicting forest re occurrence in the Western Ghats of India. International Journal of Wildland Fire, 21(4):368, jul 2012. ISSN 1049-8001. doi: 10.1071/WF10109. URL http://www.publish.csiro.au/?paper=WF10109. 60 April E. Reside, Jeremy VanDerWal, Alex Kutt, Ian Watson, and Stephen Williams. Fire regime shifts af- fect bird species distributions. Diversity and Distributions, 18(3):213{225, mar 2012. ISSN 13669516. doi: 10.1111/j.1472-4642.2011.00818.x. URL http://doi.wiley.com/10.1111/j.1472-4642.2011.00818. x. D. Riano, ~ S. L. Ustin, L. Usero, and M. A. Patricio. Estimation of Fuel Moisture Content Using Neural Networks. pages 489{498. Springer, Berlin, Heidelberg, 2005. doi: 10.1007/11499305 50. URL http: //link.springer.com/10.1007/11499305{_}50. Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. Model-Agnostic Interpretability of Machine Learning. jun 2016. URL http://arxiv.org/abs/1606.05386. Rihan, Zhao, Zhang, Guo, Ying, Deng, and Li. Wild res on the Mongolian Plateau: Identifying Drivers and Spatial Distributions to Predict Wild re Probability. Remote Sensing, 11(20):2361, oct 2019. ISSN 2072-4292. doi: 10.3390/rs11202361. URL https://www.mdpi.com/2072-4292/11/20/2361. David R. Roberts, Volker Bahn, Simone Ciuti, Mark S. Boyce, Jane Elith, Gurutzeta Guillera-Arroita, Sev- erin Hauenstein, Jos e J. Lahoz-Monfort, Boris Schr oder, Wilfried Thuiller, David I. Warton, Brendan A. Wintle, Florian Hartig, and Carsten F. Dormann. Cross-validation strategies for data with temporal, spatial, hierarchical, or phylogenetic structure. Ecography, 40(8):913{929, aug 2017. ISSN 09067590. doi: 10.1111/ecog.02881. URL http://doi.wiley.com/10.1111/ecog.02881. Marcos Rodrigues and Juan De la Riva. An insight into machine-learning algorithms to model human- caused wild re occurrence. Environmental Modelling and Software, 57:192{201, 2014. ISSN 13648152. doi: 10.1016/j.envsoft.2014.03.003. URL http://dx.doi.org/10.1016/j.envsoft.2014.03.003. Marcos Rodrigues, Ferm n Alcasena, and Cristina Vega-Garc a. Modeling initial attack success of wild re suppression in Catalonia, Spain. Science of The Total Environment, 666:915{927, may 2019. ISSN 00489697. doi: 10.1016/j.scitotenv.2019.02.323. URL https://linkinghub.elsevier.com/retrieve/ pii/S0048969719308319. Roque Rodriguez, Ana Cort es, Tom as Margalef, and Emilio Luque. An Adaptive System for Forest Fire Behavior Prediction. In 2008 11th IEEE International Conference on Computational Science and Engineering, pages 275{282. IEEE, jul 2008. ISBN 978-0-7695-3193-9. doi: 10.1109/CSE.2008.15. URL http://ieeexplore.ieee.org/document/4578243/. Roque Rodr guez, Ana Cort es, and Tom as Margalef. Injecting Dynamic Real-Time Data into a DDDAS for Forest Fire Behavior Prediction. pages 489{499. Springer, Berlin, Heidelberg, 2009. doi: 10.1007/ 978-3-642-01973-9 55. URL http://link.springer.com/10.1007/978-3-642-01973-9{_}55. Yuji Roh, Geon Heo, and Steven Euijong Whang. A Survey on Data Collection for Machine Learning: a Big Data { AI Integration Perspective. IEEE Transactions on Knowledge and Data Engineering, pages 1{1, nov 2018. URL http://arxiv.org/abs/1811.03402. David Rolnick, Priya L. Donti, Lynn H. Kaack, Kelly Kochanski, Alexandre Lacoste, Kris Sankaran, An- drew Slavin Ross, Nikola Milojevic-Dupont, Natasha Jaques, Anna Waldman-Brown, Alexandra Luc- cioni, Tegan Maharaj, Evan D. Sherwin, S. Karthik Mukkavilli, Konrad P. Kording, Carla Gomes, Andrew Y. Ng, Demis Hassabis, John C. Platt, Felix Creutzig, Jennifer Chayes, and Yoshua Bengio. Tackling Climate Change with Machine Learning. jun 2019. URL http://arxiv.org/abs/1906.05433. J. Ru ault and F. Mouillot. How a new re-suppression policy can abruptly reshape the re-weather relationship. Ecosphere, 6(10):art199, oct 2015. ISSN 2150-8925. doi: 10.1890/ES15-00182.1. URL http://doi.wiley.com/10.1890/ES15-00182.1. 61 Jakob Runge, Sebastian Bathiany, Erik Bollt, Gustau Camps-Valls, Dim Coumou, Ethan Deyle, Clark Glymour, Marlene Kretschmer, Miguel D. Mahecha, Jordi Munoz-Mar ~  , Egbert H. van Nes, Jonas Peters, Rick Quax, Markus Reichstein, Marten Sche er, Bernhard Sch olkopf, Peter Spirtes, George Sugihara, Jie Sun, Kun Zhang, and Jakob Zscheischler. Inferring causation from time series in Earth system sciences. Nature Communications, 10(1):1{13, dec 2019. ISSN 20411723. doi: 10.1038/s41467-019-10105-3. A. C. L. S a, J. M. C. Pereira, M. J. P. Vasconcelos, J. M. N. Silva, N. Ribeiro, and A. Awasse. Assessing the feasibility of sub-pixel burned area mapping in miombo woodlands of northern Mozambique using MODIS imagery. International Journal of Remote Sensing, 24(8):1783{1796, jan 2003. ISSN 0143- 1161. doi: 10.1080/01431160210144750. URL https://www.tandfonline.com/doi/full/10.1080/ Shruti Sachdeva, Tarunpreet Bhatia, and A. K. Verma. GIS-based evolutionary optimized Gradient Boosted Decision Trees for forest re susceptibility mapping. Natural Hazards, 92(3):1399{1418, jul 2018. ISSN 0921-030X. doi: 10.1007/s11069-018-3256-5. URL http://link.springer.com/10.1007/ s11069-018-3256-5. Y. Sa and A. Bouroumi. Prediction of forest res using Arti cial neural networks. Applied Mathematical Sciences, 7(6):271{286, 2013. doi: 10.12988/ams.2013.13025. George E. Sakr, Imad H. Elhajj, George Mitri, and Uchechukwu C. Wejinya. Arti cial intelligence for forest re prediction. In IEEE/ASME International Conference on Advanced Intelligent Mechatronics, AIM, 2010. ISBN 9781424480319. doi: 10.1109/AIM.2010.5695809. George E. Sakr, Imad H. Elhajj, and George Mitri. Ecient forest re occurrence prediction for developing countries using two weather parameters. Engineering Applications of Arti cial Intelli- gence, 24(5):888{894, aug 2011. ISSN 0952-1976. doi: 10.1016/J.ENGAPPAI.2011.02.017. URL https://www.sciencedirect.com/science/article/abs/pii/S0952197611000418. Jess San-Miguel-Ayanz, Ernst Schulte, Guido Schmuck, Andrea Camia, Peter Strobl, Giorgio Liberta, Cris- tiano Giovando, Roberto Boca, Fernando Sedano, Pieter Kempeneers, Daniel McInerney, Ceri Withmore, Sandra Santos de Oliveira, Marcos Rodrigues, Tracy Durrant, Paolo Corti, Friderike Oehler, Lara Vilar, and Giuseppe Amatulli. Comprehensive Monitoring of Wild res in Europe: The European Forest Fire Information System (EFFIS). In Approaches to Managing Disaster - Assessing Hazards, Emergencies and Disaster Impacts. InTech, mar 2012. doi: 10.5772/28441. L.A. Sanabria, X. Qin, J. Li, R.P. Cechet, and C. Lucas. Spatial interpolation of McArthur's Forest Fire Danger Index across Australia: Observational study. Environmental Modelling & Software, 50:37{50, dec 2013. ISSN 1364-8152. doi: 10.1016/J.ENVSOFT.2013.08.012. URL https://www.sciencedirect. com/science/article/pii/S1364815213001916?via{%}3Dihub. Onur Satir, Suha Berberoglu, and Cenk Donmez. Mapping regional forest re probability using arti cial neural network model in a Mediterranean forest ecosystem. Geomatics, Natural Hazards and Risk, 7 (5):1645{1658, sep 2016. ISSN 1947-5705. doi: 10.1080/19475705.2015.1084541. URL http://www. tandfonline.com/doi/full/10.1080/19475705.2015.1084541. Younes Oulad Sayad, Hajar Mousannif, and Hassan Al Moatassime. Predictive modeling of wild res: A new dataset and machine learning approach. Fire Safety Journal, 104:130{146, mar 2019. ISSN 03797112. doi: 10.1016/j. resaf.2019.01.006. Daniel L. Schmoldt. Application of Arti cial Intelligence to Risk Analysis for Forested Ecosystems. pages 49{74. Springer, Dordrecht, 2001. doi: 10.1007/978-94-017-2905-5 3. URL http://link.springer. com/10.1007/978-94-017-2905-5{_}3. 62 Frederic Paik Schoenberg. A NOTE ON THE CONSISTENT ESTIMATION OF SPATIAL-TEMPORAL POINT PROCESS PARAMETERS, 2016. URL https://www.jstor.org/stable/24721302. Gavin Shaddick and James V. Zidek. A case study in preferential sampling: Long term monitoring of air pollution in the UK. Spatial Statistics, 9:51{65, aug 2014. ISSN 2211-6753. doi: 10.1016/J.SPASTA.2014. 03.008. URL https://www.sciencedirect.com/science/article/abs/pii/S2211675314000219. Chaopeng Shen. A Transdisciplinary Review of Deep Learning Research and Its Relevance for Water Resources Scientists. Water Resources Research, 54(11):8558{8593, nov 2018. ISSN 0043-1397. doi: 10.1029/2018WR022643. URL https://onlinelibrary.wiley.com/doi/abs/10.1029/2018WR022643. Kirk R. Sherrill and William H. Romme. Spatial Variation in Post re Cheatgrass: Dinosaur National Monument, USA. Fire Ecology, 8(2):38{56, aug 2012. ISSN 19339747. doi: 10.4996/ reecology.0802038. URL http://fireecologyjournal.org/journal/abstract/?abstract=162. Mengyun Shi, Fengying Xie, Yue Zi, and Jihao Yin. Cloud detection of remote sensing images by deep learning. In International Geoscience and Remote Sensing Symposium (IGARSS), volume 2016-Novem, pages 701{704. Institute of Electrical and Electronics Engineers Inc., nov 2016. ISBN 9781509033324. doi: 10.1109/IGARSS.2016.7729176. Guruh Fajar Shidik and Khabib Mustofa. Predicting Size of Forest Fire Using Hybrid Model. pages 316{327. Springer, Berlin, Heidelberg, 2014. doi: 10.1007/978-3-642-55032-4 31. URL http://link. springer.com/10.1007/978-3-642-55032-4{_}31. Albert J. Simard. Fire severity, changing scales, and how things hang together. International Journal of Wildland Fire, 1(1):23{34, 1991. ISSN 10498001. doi: 10.1071/WF9910023. Imas Sukaesih Sitanggang and Mohd Hasmadi Ismail. Classi cation model for hotspot occurrences using a decision tree method. Geomatics, Natural Hazards and Risk, 2(2):111{121, jun 2011. ISSN 19475705. doi: 10.1080/19475705.2011.565807. I.S. Sitanggang, R. Yaakob, N. Mustapha, and A.N. Ainuddin. Predictive Models for Hotspots Occurrence using Decision Tree Algorithms and Logistic Regression. Journal of Applied Sciences, 13(2):252{261, feb 2013. ISSN 18125654. doi: 10.3923/jas.2013.252.261. URL http://www.scialert.net/abstract/ ?doi=jas.2013.252.261. Natasa Ski c and Jennifer Francis. Self-organizing maps: a powerful tool for the atmospheric sciences. Applications of Self-Organizing Maps, pages 251{268, 2012. W. R. Skinner, M. D. Flannigan, B. J. Stocks, D. L. Martell, B. M. Wotton, J. B. Todd, J. A. Mason, K. A. Logan, and E. M. Bosch. A 500 hPa synoptic wildland re climatology for large Canadian forest res, 1959-1996. Theoretical and Applied Climatology, 71(3-4):157{169, 2002. ISSN 0177798X. doi: 10.1007/s007040200002. Hamdy Soliman, Komal Sudan, and Ashish Mishra. A smart forest- re early detection sensory system: Another approach of utilizing wireless sensor and neural networks. In 2010 IEEE Sensors, pages 1900{ 1904. IEEE, nov 2010. ISBN 978-1-4244-8170-5. doi: 10.1109/ICSENS.2010.5690033. URL http: //ieeexplore.ieee.org/document/5690033/. Chao Song, Mei-Po Kwan, Weiguo Song, and Jiping Zhu. A Comparison between Spatial Econometric Models and Random Forest for Modeling Fire Occurrence. Sustainability, 9(5):819, may 2017. ISSN 2071-1050. doi: 10.3390/su9050819. URL http://www.mdpi.com/2071-1050/9/5/819. Kalli Srinivasa, Nageswara Prasad, and S Ramakrishna. An Autonomous Forest Fire Detection System Based On Spatial Data Mining and Fuzzy Logic. Technical Report 12, 2008. 63 B. J. Stocks and David L. Martell. Forest re management expenditures in Canada: 1970-2013. Forestry Chronicle, 92(3):298{306, jun 2016. ISSN 00157546. doi: 10.5558/tfc2016-056. Daniela Stojanova, Andrej Kobler, Sa so D zeroski, and Katerina Ta skova. LEARNING TO PREDICT FOREST FIRES WITH DIFFERENT DATA MINING TECHNIQUES. In Conference on data mining and data warehouses (SiKDD 2006), pages 255{258, 2006. URL http://www.academia.edu/download/ 30570649/10.1.1.116.2555.pdf. Daniela Stojanova, Andrej Kobler, Peter Ogrinc, Bernard Zenko, and Sa so D zeroski. Estimating the risk of re outbreaks in the natural environment. Data Mining and Knowledge Discovery, 24(2):411{442, mar 2012. ISSN 1384-5810. doi: 10.1007/s10618-011-0213-2. URL http://link.springer.com/10.1007/ s10618-011-0213-2. Jeremy Storer and Robert Green. PSO trained Neural Networks for predicting forest re size: A comparison of implementation and performance. In Proceedings of the International Joint Conference on Neural Networks, volume 2016-Octob, pages 676{683. Institute of Electrical and Electronics Engineers Inc., oct 2016. ISBN 9781509006199. doi: 10.1109/IJCNN.2016.7727265. Diana Stralberg, Xianli Wang, Marc-Andr e Parisien, Fran cois-Nicolas Robinne, P eter S olymos, C. Lisa Mahon, Scott E. Nielsen, and Erin M. Bayne. Wild re-mediated vegetation change in boreal forests of Alberta, Canada. Ecosphere, 9(3):e02156, 2018. ISSN 2150-8925. doi: 10.1002/ecs2.2156. Carolin Strobl, Anne-Laure Boulesteix, Achim Zeileis, and Torsten Hothorn. Bias in random forest variable importance measures: Illustrations, sources and a solution. BMC Bioinformatics, 8(1):25, dec 2007. ISSN 1471-2105. doi: 10.1186/1471-2105-8-25. URL https://bmcbioinformatics.biomedcentral. com/articles/10.1186/1471-2105-8-25. Esther D. Stroh, Matthew A. Struckho , Michael C. Stambaugh, and Richard P. Guyette. Fire and Climate Suitability for Woody Vegetation Communities in the South Central United States. Fire Ecology 2018 14:1, 14(1):106{124, feb 2018. ISSN 1933-9747. doi: 10.4996/FIREECOLOGY.140110612. URL https://fireecology.springeropen.com/articles/10.4996/fireecology.140110612. Sriram Ganapathi Subramanian and Mark Crowley. Learning Forest Wild re Dynamics from Satellite Images Using Reinforcement Learning. In Conference on Reinforcement Learning and Decision Making, page 5, 2017. URL http://www.ospo.noaa.gov/Products/atmosphere/wind.html. Sriram Ganapathi Subramanian and Mark Crowley. Using Spatial Reinforcement Learning to Build Forest Wild re Dynamics Models From Satellite Images. Frontiers in ICT, 5:6, apr 2018. ISSN 2297-198X. doi: 10.3389/ ct.2018.00006. URL http://journal.frontiersin.org/article/10.3389/fict.2018. 00006/full. AL Sullivan. A review of wildland re spread modelling, 1990-present 3: Mathematical analogues and simulation models. arXiv preprint arXiv:0706.4130, 2007. Andrew L. Sullivan. Wildland surface re spread modelling, 1990 - 2007. 3: Simulation and mathematical analogue models. International Journal of Wildland Fire, 18(4):387, 2009a. ISSN 1049-8001. doi: 10.1071/wf06144. Andrew L. Sullivan. Wildland surface re spread modelling, 1990 - 2007. 2: Empirical and quasi-empirical models. International Journal of Wildland Fire, 18(4):369, 2009b. ISSN 1049-8001. doi: 10.1071/ wf06142. Andrew L. Sullivan. Wildland surface re spread modelling, 1990 - 2007. 1: Physical and quasi-physical models. International Journal of Wildland Fire, 18(4):349, 2009c. ISSN 1049-8001. doi: 10.1071/wf06143. 64 Brian L. Sullivan, Jocelyn L. Aycrigg, Jessie H. Barry, Rick E. Bonney, Nicholas Bruns, Caren B. Cooper, Theo Damoulas, Andr e A. Dhondt, Tom Dietterich, Andrew Farnsworth, Daniel Fink, John W. Fitz- patrick, Thomas Fredericks, Je Gerbracht, Carla Gomes, Wesley M. Hochachka, Marshall J. Ili , Carl Lagoze, Frank A. La Sorte, Matthew Merri eld, Will Morris, Tina B. Phillips, Mark Reynolds, Amanda D. Rodewald, Kenneth V. Rosenberg, Nancy M. Trautmann, Andrea Wiggins, David W. Win- kler, Weng Keen Wong, Christopher L. Wood, Jun Yu, and Steve Kelling. The eBird enterprise: An integrated approach to development and application of citizen science, jan 2014. ISSN 00063207. Alexander Y. Sun and Bridget R Scanlon. How can big data and machine learning bene t environment and water management: A survey of methods, applications, and future directions. Environmental Research Letters, apr 2019. ISSN 1748-9326. doi: 10.1088/1748-9326/ab1b7d. URL http://iopscience.iop. org/article/10.1088/1748-9326/ab1b7d. F. Sunar and C. Ozkan. Forest re analysis with remote sensing data. International Journal of Remote Sensing, 22(12):2265{2277, jan 2001. ISSN 0143-1161. doi: 10.1080/01431160118510. URL https: //www.tandfonline.com/doi/full/10.1080/01431160118510. Richard S Sutton and Andrew G Barto. Introduction to reinforcement learning, volume 135. MIT press Cambridge, 1998. Richard S Sutton and Andrew G Barto. Reinforcement learning: An introduction. MIT press, 2018. Alexandra D. Syphard, Jon E. Keeley, Avi Bar Massada, Teresa J. Brennan, and Volker C. Radelo . Housing Arrangement and Location Determine the Likelihood of Housing Loss Due to Wild re. PLoS ONE, 7(3):e33954, mar 2012. ISSN 1932-6203. doi: 10.1371/journal.pone.0033954. URL http://dx. plos.org/10.1371/journal.pone.0033954. Alexandra D. Syphard, Avi Bar Massada, Van Butsic, and Jon E. Keeley. Land Use Planning and Wild re: Development Policies In uence Future Probability of Housing Loss. PLoS ONE, 8(8):e71708, aug 2013. ISSN 1932-6203. doi: 10.1371/journal.pone.0071708. URL https://dx.plos.org/10.1371/journal. pone.0071708. Alexandra D. Syphard, Van Butsic, Avi Bar-Massada, Jon E. Keeley, Je A. Tracey, and Robert N. Fisher. Setting priorities for private land conservation in re-prone landscapes: Are re risk reduction and biodiversity conservation competing or compatible objectives? Ecology and Society, 21(3):art2, jul 2016. ISSN 1708-3087. doi: 10.5751/ES-08410-210302. URL http://www.ecologyandsociety.org/ vol21/iss3/art2/. Alexandra D. Syphard, Heather Rustigian-Romsos, Michael Mann, Erin Conlisk, Max A. Moritz, and David Ackerly. The relative in uence of climate and housing development on current and projected future re patterns and structure loss across three California landscapes. Global Environmental Change, 56:41{55, may 2019. ISSN 0959-3780. doi: 10.1016/J.GLOENVCHA.2019.03.007. URL https://www. sciencedirect.com/science/article/pii/S0959378018313293. S. W. Taylor, Douglas G. Woolford, C. B. Dean, and David L. Martell. Wild re Prediction to Inform Fire Management: Statistical Science Challenges. Statistical Science, 28(4):586{615, 2013. ISSN 0883-4237. doi: 10.1214/13-STS451. S.W. Taylor. Atmospheric cascades shape wild re re management decision spaces - a theory unifying re weather and re management. Submitted, 2020. Mahyat Shafapour Tehrany, Simon Jones, Farzin Shabani, Francisco Mart nez-Alvarez, and Dieu Tien Bui. A novel ensemble modeling approach for the spatial prediction of tropical forest re susceptibility using LogitBoost machine learning classi er and multi-source geospatial data. Theoretical and Applied 65 Climatology, pages 1{17, sep 2018. ISSN 0177-798X. doi: 10.1007/s00704-018-2628-9. URL http: //link.springer.com/10.1007/s00704-018-2628-9. Jonathan R. Thompson and Thomas A. Spies. Factors associated with crown damage following recurring mixed-severity wild res and post- re management in southwestern Oregon. Landscape Ecology, 25(5): 775{789, may 2010. ISSN 0921-2973. doi: 10.1007/s10980-010-9456-3. URL http://link.springer. com/10.1007/s10980-010-9456-3. Matthew P. Thompson and Dave E. Calkin. Uncertainty and risk in wildland re management: A re- view. Journal of Environmental Management, 92(8):1895{1909, aug 2011. ISSN 0301-4797. doi: 10.1016/J.JENVMAN.2011.03.015. URL https://www.sciencedirect.com/science/article/pii/ S0301479711000818. Dieu Tien Bui, Quang-Thanh Bui, Quoc-Phi Nguyen, Biswajeet Pradhan, Haleh Nampak, and Phan Trong Trinh. A hybrid arti cial intelligence approach using GIS-based neural-fuzzy inference system and particle swarm optimization for forest re susceptibility modeling at a tropical area. Agricultural and Forest Meteorology, 233:32{44, feb 2017. ISSN 0168-1923. doi: 10.1016/J.AGRFORMET.2016.11.002. URL https://www.sciencedirect.com/science/article/pii/S0168192316304269. Ahmed Toujani, Hammadi Achour, and Sami Fa z. Estimating Forest Fire Losses Using Stochas- tic Approach: Case Study of the Kroumiria Mountains (Northwestern Tunisia). Applied Arti cial Intelligence, pages 1{25, sep 2018. ISSN 0883-9514. doi: 10.1080/08839514.2018.1514808. URL https://www.tandfonline.com/doi/full/10.1080/08839514.2018.1514808. James L. Tracy, Antonio Trabucco, A. Michelle Lawing, J. Tomasz Giermakowski, Maria Tchakerian, Gail M. Drus, and Robert N. Coulson. Random subset feature selection for ecological niche models of wild re activity in Western North America. Ecological Modelling, 383:52{68, sep 2018. ISSN 0304- 3800. doi: 10.1016/J.ECOLMODEL.2018.05.019. URL https://www.sciencedirect.com/science/ article/pii/S0304380018301868. Cordy Tymstra, Brian J. Stocks, Xinli Cai, and Mike D. Flannigan. Wild re management in Canada: Review, challenges and opportunities. Progress in Disaster Science, page 100045, 2019. ISSN 25900617. doi: 10.1016/j.pdisas.2019.100045. URL https://linkinghub.elsevier.com/retrieve/ pii/S2590061719300456. A.B. Utkin, Armando M Fernandes, Fernando Sim~ oes, and R. Vilar. Forest- re detection by means of lidar. Proceedings of IV International Conference on Forest Fire Research, (1993):1{14, 2002. Giorgio Vacchiano, Cristiano Foderi, Roberta Berretti, Enrico Marchi, and Renzo Motta. Modeling anthropogenic and natural re ignitions in an inner-alpine valley. Natural Hazards and Earth Sys- tem Sciences, 18(3):935{948, mar 2018. ISSN 1684-9981. doi: 10.5194/nhess-18-935-2018. URL https://www.nat-hazards-earth-syst-sci.net/18/935/2018/. D. Vakalis, H. Sarimveis, C. Kiranoudis, A. Alexandridis, and G. Bafas. A GIS based operational system for wildland re crisis management I. Mathematical modelling and simulation. Applied Mathematical Modelling, 28(4):389{410, 2004. ISSN 0307904X. doi: 10.1016/j.apm.2003.10.005. Miguel Conrado Valdez, Kang-Tsung Chang, Chi-Farn Chen, Shou-Hao Chiang, and Jorge Luis Santos. Modelling the spatial variability of wild re susceptibility in Honduras using remote sensing and geo- graphical information systems. Geomatics, Natural Hazards and Risk, 8(2):876{892, dec 2017. ISSN 1947-5705. doi: 10.1080/19475705.2016.1278404. URL https://www.tandfonline.com/doi/full/10. 1080/19475705.2016.1278404. 66 Ashley E. Van Beusekom, William A. Gould, A. Carolina Monmany, Azad Henareh Khalyani, Maya Quinones, ~ Stephen J. Fain, Maria Jos e Andrade-Nu nez, ~ and Grizelle Gonz alez. Fire weather and likelihood: characterizing climate space for re occurrence and extent in Puerto Rico. Climatic Change, 146(1-2):117{131, jan 2018. ISSN 0165-0009. doi: 10.1007/s10584-017-2045-6. URL http: //link.springer.com/10.1007/s10584-017-2045-6. P. van Breugel, I. Friis, Sebsebe Demissew, Jens-Peter Barnekow Lilles, and Roeland Kindt. Current and Future Fire Regimes and Their In uence on Natural Vegetation in Ethiopia. Ecosystems, 19(2):369{386, mar 2016. ISSN 1432-9840. doi: 10.1007/s10021-015-9938-x. URL http://link.springer.com/10. 1007/s10021-015-9938-x. Ce Van Wagner. Development and Structure of the Canadian Forest Fire Weather Index System. 1987. ISBN 0662151984. doi: 19927. Thomas Vandal, Evan Kodra, and Auroop R. Ganguly. Intercomparison of machine learning methods for statistical downscaling: the case of daily and extreme precipitation. Theoretical and Applied Climatology, pages 1{14, sep 2018. ISSN 0177-798X. doi: 10.1007/s00704-018-2613-3. URL http://link.springer. com/10.1007/s00704-018-2613-3. Christos Vasilakos, Kostas Kalabokidis, John Hatzopoulos, George Kallos, and Yiannis Matsinos. Integrat- ing new methods and tools in re danger rating. International Journal of Wildland Fire, 16(3):306{316, 2007. ISSN 10498001. doi: 10.1071/WF05091. Christos Vasilakos, Kostas Kalabokidis, John Hatzopoulos, and Ioannis Matsinos. Identifying wildland re ignition factors through sensitivity analysis of a neural network. Natural Hazards, 50(1):125{143, jul 2009. ISSN 0921-030X. doi: 10.1007/s11069-008-9326-3. URL http://link.springer.com/10.1007/ s11069-008-9326-3. Daniel Vec n-Arias, Fernando Castedo-Dorado, Celestino Ord onez, ~ and Jos e Ram on Rodr guez-P erez. Bio- physical and lightning characteristics drive lightning-induced re occurrence in the central plateau of the Iberian Peninsula. Agricultural and Forest Meteorology, 225:36{47, sep 2016. ISSN 0168- 1923. doi: 10.1016/J.AGRFORMET.2016.05.003. URL https://www.sciencedirect.com/science/ article/pii/S0168192316302593. C. Vega-Garcia, B.S. Lee, P.M. Woodard, and S.J. Titus. Applying neural network technology to human- caused wild re occurrence prediction. AI Applications, pages 9{18, 1996. URL https://cfs.nrcan. gc.ca/publications?id=18949. O Viedma, L A Arroyo, R Mateo, A De Santis, and J M Moreno. EFFECTS OF ENVIRONMEN- TAL PROPERTIES, BURNING CONDITIONS AND HUMAN-RELATED VARIABLES ON FIRE SEVERITY DERIVED FROM LANDSAT TM IMAGES FOR A LARGE FIRE IN CENTRAL SPAIN. In Advances in Remote Sensing and GIS applications in Forest Fire Management From local to global assessments, page 157. 2011. Olga Viedma, Juan Quesada, Iv an Torres, Angela De Santis, and Jos e M. Moreno. Fire Severity in a Large Fire in a Pinus pinaster Forest is Highly Predictable from Burning Conditions, Stand Structure, and Topography. Ecosystems, 18(2):237{250, mar 2015. ISSN 1432-9840. doi: 10.1007/s10021-014-9824-y. URL http://link.springer.com/10.1007/s10021-014-9824-y. Domingos Xavier Viegas, Karin L. Riley, Isaac C. Grenfell, Mark A. Finney, and Nicholas L. Crookston. Utilizing random forests imputation of forest plot data for landscape-level wild- re analyses. Imprensa da Universidade de Coimbra, Coimbra, 2014. ISBN 978-989-26-0884- 6 (PDF). doi: 10.14195/978-989-26-0884-6 67. URL https://digitalis.uc.pt/en/livro/ utilizing{_}random{_}forests{_}imputation{_}forest{_}plot{_}data{_}landscape{_}level{_}wildfire{_}analyses. 67 Dinesh Babu Irulappa Pillai Vijayakumar, Fr ed eric Raulier, Pierre Y. Bernier, Sylvie Gauthier, Yves Bergeron, and David Pothier. Lengthening the historical records of re history over large areas of boreal forest in eastern Canada using empirical relationships. Forest Ecology and Management, 347:30{39, jul 2015. ISSN 0378-1127. doi: 10.1016/J.FORECO.2015.03.011. URL https://www.sciencedirect.com/ science/article/pii/S0378112715001310{#}b0140. Dinesh Babu Irulappa Pillai Vijayakumar, Fr ed eric Raulier, Pierre Bernier, David Par e, Sylvie Gauthier, Yves Bergeron, and David Pothier. Cover density recovery after re disturbance controls landscape aboveground biomass carbon in the boreal forest of eastern Canada. Forest Ecology and Management, 360:170{180, jan 2016. ISSN 0378-1127. doi: 10.1016/J.FORECO.2015.10.035. URL https://www. sciencedirect.com/science/article/pii/S0378112715005927?via{%}3Dihub. Lara Vilar, Israel G omez, Javier Mart nez-Vega, Pilar Echavarr a, David Rian ~o, and M. Pilar Mart n. Multitemporal Modelling of Socio-Economic Wild re Drivers in Central Spain between the 1980s and the 2000s: Comparing Generalized Linear Models to Machine Learning Algorithms. PLOS ONE, 11(8): e0161344, aug 2016. ISSN 1932-6203. doi: 10.1371/journal.pone.0161344. URL http://dx.plos.org/ 10.1371/journal.pone.0161344. Yan Wang, Chunyu Yu, Ran Tu, and Yongming Zhang. Fire detection model in Tibet based on grey-fuzzy neural network algorithm. Expert Systems with Applications, 38(8):9580{9586, aug 2011. ISSN 0957- 4174. doi: 10.1016/J.ESWA.2011.01.163. URL https://www.sciencedirect.com/science/article/ pii/S0957417411001965. Yuanbin Wang, Langfei Dang, and Jieying Ren. Forest re image recognition based on convolutional neural network. Journal of Algorithms & Computational Technology, 13:174830261988768, jan 2019. ISSN 1748-3026. doi: 10.1177/1748302619887689. Gregory L. Watson, Donatello Telesca, Colleen E. Reid, Gabriele G. P ster, and Michael Jerrett. Machine learning models accurately predict ozone exposure during wild re events. Environmental Pollution, 254: 112792, nov 2019. ISSN 18736424. doi: 10.1016/j.envpol.2019.06.088. David H Wolpert. The lack of a priori distinctions between learning algorithms. Neural computation, 8(7): 1341{1390, 1996. Zhiwei Wu, Hong S. He, Jian Yang, Zhihua Liu, and Yu Liang. Relative e ects of climatic and local factors on re occurrence in boreal forest landscapes of northeastern China. Science of The Total Environment, 493:472{480, sep 2014. ISSN 0048-9697. doi: 10.1016/J.SCITOTENV.2014.06.011. URL https://www.sciencedirect.com/science/article/pii/S0048969714008547. Zhiwei Wu, Hong S. He, Jian Yang, and Yu Liang. De ning re environment zones in the boreal forests of northeastern China. Science of The Total Environment, 518-519:106{116, jun 2015. ISSN 0048-9697. doi: 10.1016/J.SCITOTENV.2015.02.063. URL https://www.sciencedirect.com/science/article/ pii/S0048969715002065. Dexen D Z Xi, Stephen W Taylor, Douglas G Woolford, and C B Dean. Statistical Models of Key Components of Wild re Risk. 2019. doi: 10.1146/annurev-statistics-031017-100450. URL www.annualreviews.org. Dao Wen Xie and Shi Liang Shi. Prediction for Burned Area of Forest Fires Based on SVM Model. Applied Mechanics and Materials, 513-517:4084{4089, feb 2014. ISSN 1662-7482. doi: 10.4028/www.scienti c. net/AMM.513-517.4084. URL https://www.scientific.net/AMM.513-517.4084. Ying Xie and Minggang Peng. Forest re forecasting using ensemble learning approaches. Neural Computing and Applications, 31(9):4541{4550, sep 2019. ISSN 0941-0643. doi: 10.1007/s00521-018-3515-0. URL http://link.springer.com/10.1007/s00521-018-3515-0. 68 Jiayun Yao, Michael Brauer, Sean Ra use, and Sarah B. Henderson. Machine Learning Approach To Estimate Hourly Exposure to Fine Particulate Matter for Urban, Rural, and Remote Populations during Wild re Seasons. Environmental Science & Technology, 52(22):13239{13249, nov 2018a. ISSN 0013- 936X. doi: 10.1021/acs.est.8b01921. URL http://pubs.acs.org/doi/10.1021/acs.est.8b01921. Jiayun Yao, Sean M. Ra use, Michael Brauer, Grant J. Williamson, David M.J.S. Bowman, Fay H. Johnston, and Sarah B. Henderson. Predicting the minimum height of forest re smoke within the atmosphere using machine learning and data from the CALIPSO satellite. Remote Sensing of Environment, 206:98{106, mar 2018b. ISSN 0034-4257. doi: 10.1016/J.RSE.2017.12.027. URL https://www.sciencedirect.com/science/article/abs/pii/S003442571730603X. Lingxiao Ying, Jie Han, Yongsheng Du, and Zehao Shen. Forest re characteristics in China: Spatial pat- terns and determinants with thresholds. Forest Ecology and Management, 424:345{354, sep 2018. ISSN 0378-1127. doi: 10.1016/J.FORECO.2018.05.020. URL https://www.sciencedirect.com/science/ article/pii/S0378112717317668. Adam M. Young, Philip E. Higuera, Paul A. Du y, and Feng Sheng Hu. Climatic thresholds shape northern high-latitude re regimes and imply vulnerability to future climate change. Ecography, 40(5):606{617, may 2017. ISSN 16000587. doi: 10.1111/ecog.02205. Adam M. Young, Philip E. Higuera, John T. Abatzoglou, Paul A. Du y, and Feng Sheng Hu. Consequences of climatic thresholds for projecting re activity and ecological change. Global Ecology and Biogeography, 28(4):521{532, apr 2019. ISSN 14668238. doi: 10.1111/geb.12872. Bo Yu, Fang Chen, Bin Li, Li Wang, and Mingquan Wu. Fire Risk Prediction Using Remote Sensed Products: A Case of Cambodia. Photogrammetric Engineering & Remote Sensing, 83(1):19{25, jan 2017. ISSN 0099-1112. doi: 10.14358/PERS.83.1.19. URL http://www.ingentaconnect.com/content/10. 14358/PERS.83.1.19. Yong Poh Yu, Rosli Omar, Rhett D Harrison, Mohan Kumar Sammathuria, and Abdul Rahim Nik. Pattern clustering of forest res based on meteorological variables and its classi cation using hybrid data mining methods. Journal of Computational Biology and Bioinformatics Research, 3(4):47{52, 2011. URL http: //www.academicjournals.org/jcbbr. Jie Yuan, Lidong Wang, Peng Wu, Chao Gao, and Lingqing Sun. Detection of Wild res along Transmission Lines Using Deep Time and Space Features. Pattern Recognition and Image Analysis, 28(4):805{812, oct 2018. ISSN 15556212. doi: 10.1134/S1054661818040168. Bianca Zadrozny. Learning and evaluating classi ers under sample selection bias. In Twenty- rst interna- tional conference on Machine learning - ICML '04, page 114, New York, New York, USA, 2004. ACM Press. ISBN 1581138285. doi: 10.1145/1015330.1015425. URL http://portal.acm.org/citation. cfm?doid=1015330.1015425. Harold S.J. Zald and Christopher J. Dunn. Severe re weather and intensive forest management increase re severity in a multi-ownership landscape. Ecological Applications, 28(4):1068{1080, 2018. ISSN 19395582. doi: 10.1002/eap.1710. Olivier Zammit, Xavier Descombes, and Josiane Zerubia. Burnt area mapping using Support Vector Machines. Forest Ecology and Management, 234:S240, 2006. ISSN 03781127. doi: 10.1016/j.foreco.2006. 08.269. URL http://linkinghub.elsevier.com/retrieve/pii/S0378112706008097. Bin Zhang, Wei Wei, Bingqian He, and Chuanlei Guo. Early wild re smoke detection based on improved codebook model and convolutional neural networks. In Xudong Jiang and Jenq- Neng Hwang, editors, Tenth International Conference on Digital Image Processing (ICDIP 69 2018), page 120. SPIE, aug 2018a. ISBN 9781510621992. doi: 10.1117/12.2502974. URL https://www.spiedigitallibrary.org/conference-proceedings-of-spie/10806/2502974/ Early-wildfire-smoke-detection-based-on-improved-codebook-model-and/10.1117/12. 2502974.full. Guoli Zhang, Ming Wang, and Kai Liu. Forest Fire Susceptibility Modeling Using a Convolutional Neural Network for Yunnan Province of China. International Journal of Disaster Risk Science, 10(3):1{18, sep 2019. ISSN 2095-0055. doi: 10.1007/s13753-019-00233-1. URL http://link.springer.com/10.1007/ s13753-019-00233-1. Qi Xing Zhang, Gao Hua Lin, Yong Ming Zhang, Gao Xu, and Jin Jun Wang. Wildland Forest Fire Smoke Detection Based on Faster R-CNN using Synthetic Smoke Images. In Procedia Engineering, volume 211, pages 441{446. Elsevier Ltd, 2018b. doi: 10.1016/j.proeng.2017.12.034. Qingjie Zhang, Jiaolong Xu, Liang Xu, and Haifeng Guo. Deep Convolutional Neural Networks for Forest Fire Detection. Atlantis Press, 2016. doi: 10.2991/ifmeita-16.2016.105. Zhanqing Li, A. Khananian, R.H. Fraser, and J. Cihlar. Automatic detection of re smoke using arti cial neural networks and threshold approaches applied to AVHRR imagery. IEEE Transactions on Geoscience and Remote Sensing, 39(9):1859{1870, 2001. ISSN 01962892. doi: 10.1109/36.951076. URL http: //ieeexplore.ieee.org/document/951076/. Feng Zhao, Chengquan Huang, and Zhiliang Zhu. Use of Vegetation Change Tracker and Support Vector Machine to Map Disturbance Types in Greater Yellowstone Ecosystems in a 1984{2010 Landsat Time Series. IEEE Geoscience and Remote Sensing Letters, 12(8):1650{1654, aug 2015. ISSN 1545-598X. doi: 10.1109/LGRS.2015.2418159. URL http://ieeexplore.ieee.org/document/7088596/. Jianhui Zhao, Zhong Zhang, Shizhong Han, Chengzhang Qu, Zhiyong Yuan, and Dengyi Zhang. SVM based forest re detection using static and dynamic features. Computer Science and Information Systems, 8 (3):821{841, 2011. ISSN 1820-0214. doi: 10.2298/csis101012030z. Yi Zhao, Jiale Ma, Xiaohui Li, and Jie Zhang. Saliency Detection and Deep Learning-Based Wild re Identi cation in UAV Imagery. Sensors, 18(3):712, feb 2018. ISSN 1424-8220. doi: 10.3390/s18030712. URL http://www.mdpi.com/1424-8220/18/3/712. Zhong Zheng, Wei Huang, Songnian Li, and Yongnian Zeng. Forest re spread simulating model us- ing cellular automaton with extreme learning machine. Ecological Modelling, 348(May 2018):33{43, 2017. ISSN 03043800. doi: 10.1016/j.ecolmodel.2016.12.022. URL http://dx.doi.org/10.1016/j. ecolmodel.2016.12.022. Yufei Zou, Susan M. O'Neill, Narasimhan K. Larkin, Ernesto C. Alvarado, Robert Solomon, Cli ord Mass, Yang Liu, M. Talat Odman, and Huizhong Shen. Machine Learning-Based Integration of High-Resolution Wild re Smoke Simulations and Observations for Regional Health Impact Assessment. International Journal of Environmental Research and Public Health, 16(12):2137, jun 2019. ISSN 1660-4601. doi: 10.3390/ijerph16122137. URL https://www.mdpi.com/1660-4601/16/12/2137. K. Zwirglmaier, P. Papakosta, and D. Straub. Learning a Bayesian network model for predicting wild re behavior. In ICOSSAR 2013, 2013. 70 Supplementary Material This supplemental contains all papers identi ed in this review with ML applications for wild re science and management, organized by problem domains. Note that some papers are repeated in multiple problem domains. S.1. Fuels Characterization, Fire Detection And Mapping S.1.1 Fuels characterization Citation ML methods used Study Region Riano ~ et al. [2005] ANN Not speci ed Garc a et al. [2011] SVM Alto Tajo Natural Park, central Spain Pierce et al. [2012] RF Lassen Volcanic National Park, Cal- ifornia, USA Chirici et al. [2013] DT, RF, BRT Sicily, Italy Viegas et al. [2014] RF Eastern Oregon, USA L opez-Serrano et al. SVM, KNN, RF Sierra Madre Occidental, Mexico [2016] S.1.2 Fire detection Citation ML methods used Study Region Arrue et al. [2000] ANN Experiments at University of Seville Al-Rawi et al. [2001] ANN Eastern Spain Zhanqing Li et al. [2001] ANN Canada Utkin et al. [2002] ANN Not speci ed Cordoba et al. [2004] GA Experiments Fernandes et al. [2004a] ANN Not speci ed Fernandes et al. [2004b] ANN Not speci ed Srinivasa et al. [2008] KM Not speci ed Angayarkkani and Rad- ANN Not speci ed hakrishnan [2010] Ko et al. [2010] BN test images Soliman et al. [2010] ANN Laboratory experiments Angayarkkani and Rad- ANFIS not speci ed hakrishnan [2011] Wang et al. [2011] ANFIS Tibet Zhao et al. [2011] SVM, GMM Test images Li et al. [2015] ANN China, North East Asia, Russia, Canada, Australia Liu et al. [2015] ANN Laboratory experiments Zhang et al. [2016] CNN, SVM Test images Akhlou et al. [2018] CNN Corsica Li et al. [2018b] CNN Test images Continued on next page 71 Citation ML methods used Study Region Muhammad et al. [2018] CNN Test images Yuan et al. [2018] CNN Test images Zhang et al. [2018a] CNN Test images Zhang et al. [2018b] CNN Synthetic data Zhao et al. [2018] CNN Test images Alexandrov et al. [2019] CNN, Test images HAAR CASCADES, YOLO Ba et al. [2019] CNN Satellite test images Barmpoutis et al. [2019] CNN Test images Cao et al. [2019] CNN, LSTM Test images and video Hossain et al. [2019] ANN Test images Jakubowski et al. [2019] CNN Test images Jo~ ao Sousa et al. [2019] CNN Corsica Li et al. [2019] CNN Test images Phan and Nguyen CNN American Continent [2019] Sayad et al. [2019] ANN, SVM Canada Wang et al. [2019] CNN Test images S.1.3 Fire perimeter and severity mapping Citation ML methods used Study Region Al-Rawi et al. [2001] ANN Eastern Spain Brumby et al. [2001] GA Cerro Grande Fire, New Mexico, USA Sunar and Ozkan [2001] ANN, ISODATA south coast of Turkey Al-Rawi et al. [2002] ANN Valencia, Spain S a et al. [2003] DT, BAG Northern Mozambique Pu and Gong [2004] ANN Northern California, USA Zammit et al. [2006] SVM, KM, KNN Southern France Alonso-Benito et al. SVM Tenerife and Gran Canaria [2008] Cao et al. [2009] SVM Mongolia and China Celik [2010] GA Reno Lake Tahoe area, Nevada, USA Petropoulos et al. [2010] SVM Greece E. Dragozi, I. Z. Gi- SVM, KNN Stresa, Italy tas, D.G. Stavrakoudis [2011] G omez and Pilar ANN Iberian Peninsula Mart n [2011] Petropoulos et al. [2011] SVM Greece Mitrakis et al. [2012] NFM, ANN, SVM, AD- Greece ABOOST Dragozi et al. [2014] SVM Parnitha and Rhodes, Greece Continued on next page 72 Citation ML methods used Study Region Hultquist et al. [2014] GP, RF, SVM Big Sur, California, USA Zhao et al. [2015] SVM Greater Yellowstone Ecosystem, USA Hamilton et al. [2017] SVM Idaho, USA Hawbaker et al. [2017] BRT USA Pereira et al. [2017] SVM Cerrado savanna, Brazil Collins et al. [2018] RF Victoria, Australia Nitze et al. [2018] RF Alaska, Eastern Canada, Western Siberia and Ea... Crowley et al. [2019] BULC Elephant Hill Fire, British Columbia, Canada Langford et al. [2019] DNN Interior Alaska, USA Quintano et al. [2019] MAXENT La Cabrera, Spain Quintero et al. [2019] RF West Central Spain S.2. Fire Weather And Climate Change S.2.1 Fire weather prediction Citation ML methods used Study Region Skinner et al. [2002] KM Canada Crimmins [2006] SOM Southwest USA Sanabria et al. [2013] RF Australia Lagerquist et al. [2017] SOM Alberta, Canada Nauslar et al. [2019] SOM Southwest USA S.2.2 Lightning prediction Citation ML methods used Study Region Blouin et al. [2016] RF Alberta, Canada Bates et al. [2017] DT, RF Australia S.2.3 Climate change Citation ML methods used Study Region Moritz et al. [2012] MAXENT Global Amatulli et al. [2013] RF Mediterranean Europe Batllori et al. [2013] MAXENT Mediterranen ecosystems, Global Liu and Wimberly BRT, RF Western USA [2016] Parks et al. [2016] BRT Western USA van Breugel et al. [2016] RF, SVM, BRT, MAX- Ethiopia ENT, ANN, DT Continued on next page 73 Citation ML methods used Study Region Davis et al. [2017] MAXENT Paci c Northwest, USA Li et al. [2017] MAXENT Yunnan Province, Southwest China Young et al. [2017] BRT Alaska, USA Boulanger et al. [2018] RF, BRT, DT Canada Stralberg et al. [2018] RF Alberta, Canada Stroh et al. [2018] MAXENT South central USA Tracy et al. [2018] MAXENT Western North America Buckland et al. [2019] ANN Nebraska, USA Young et al. [2019] BRT Alaska, USA S.3. Fire Occurrence, Susceptibility and Risk S.3.1 Fire occurrence prediction Citation ML methods used Study Region Vega-Garcia et al. ANN Alberta, Canada [1996] Alonso-Betanzos et al. ANN Galicia, Northwest Spain [2002] Alonso-Betanzos et al. ANN Galicia, Northwest Spain [2003] Vasilakos et al. [2007] ANN Lesvos Island, Greece Sakr et al. [2010] SVM Lebanon Sakr et al. [2011] SVM, ANN Lebanon Stojanova et al. [2012] KNN, NB, DT, SVM, BN, Slovenia ADABOOST, BAG, RF Dutta et al. [2013] ANN, DNN Australia Chen et al. [2015] MAXENT Daxinganling Mountains, North- eastern China De Angelis et al. [2015] MAXENT Canton Ticino, Switzerland Dutta et al. [2016] DNN Australia Vec n-Arias et al. [2016] RF Central Iberian Peninsula, Spain Cao et al. [2017] ANN, RF Yunnan Province, China Yu et al. [2017] RF Cambodia Van Beusekom et al. RF Puerto Rico [2018] S.3.2 Landscape-scale Burned area prediction Citation ML methods used Study Region Cheng and Wang [2008] RNN Canada Archibald et al. [2009] RF Southern Africa Arnold et al. [2014] HCL, MAXENT Interior Western USA Mayr et al. [2018] DT, RF, SVM, KM Namibia Continued on next page 74 Citation ML methods used Study Region de Bem et al. [2018] ANN, GA Federal District, Brazil S.3.3 Fire Susceptibility Mapping Citation ML methods used Study Region Chuvieco et al. [1999] ANN Mediterranean Europe De Vasconcelos et al. ANN central Portugal [2001] Amatulli et al. [2006] DT Gargano Peninsula, Italy Amatulli and Camia DT Tuscany, Italy [2007] Lozano et al. [2008] DT Northwestern Spain Holden et al. [2009] RF Gila National Forest, New Mexico, USA Maeda et al. [2009] ANN Brazil Mallinis et al. [2009] DT, KM Greece Parisien and Moritz MAXENT, BRT USA [2009] Barrett et al. [2011] RF Alaska Dimuccio et al. [2011] ANN Central Portugal Dlamini [2011] BN Swaziland Bisquert et al. [2012] ANN Galicia, Northwest Spain Moritz et al. [2012] MAXENT Global Oliveira et al. [2012] RF Mediterranean Europe Parisien et al. [2012] MAXENT Western USA Renard et al. [2012] MAXENT Western Ghats, India Syphard et al. [2012] MAXENT Southern California, USA Bar Massada et al. RF, MAXENT Michigan, USA [2013] Luo et al. [2013] RF global Peters et al. [2013] MAXENT Northeast USA Syphard et al. [2013] MAXENT South coast ecoregion, San Diego County, USA Arpaci et al. [2014] RF, MAXENT Tyrol, European Alps Parisien et al. [2014] DT Canada Rodrigues and De la RF, BRT, SVM Peninsular Spain Riva [2014] Duane et al. [2015] MAXENT Catalonia, Spain Bashari et al. [2016] BN Isfahan province, Iran Curt et al. [2016] BRT Southeastern France Fonseca et al. [2016] MAXENT Brazil Goldarag et al. [2016] ANN Golestan province, Northern Iran Guo et al. [2016b] RF Daxing'an Mountains, Northeast China Guo et al. [2016a] RF Fujian province, China Pourtaghi et al. [2016] BRT, RF Golestan province, Northern Iran Continued on next page 75 Citation ML methods used Study Region Satir et al. [2016] ANN Upper Seyhan Basin, Turkey van Breugel et al. [2016] RF, SVM, BRT, MAX- Ethiopia ENT, ANN, DT Vilar et al. [2016] MAXENT Madrid region, Spain Adab [2017] ANN Northeast Iran Cao et al. [2017] ANN, RF Yunnan Province, China Davis et al. [2017] MAXENT Paci c Northwest, USA Ebrahimy et al. [2017] MAXENT Eastern Azerbaijan Li et al. [2017] MAXENT Yunnan Province, Southwest China MOSTAFA et al. [2017] SVM Golestan province, Northern Iran Peters and Iverson MAXENT Northeast USA [2017] Song et al. [2017] RF Hefei City, China Tien Bui et al. [2017] NFM, PSO, RF, SVM Lam Dong province, Vietnam Valdez et al. [2017] RF Honduras Adab et al. [2018] MAXENT Mazandaran province, Iran Hong et al. [2018] SVM, RF, GA Southwest Jiangxi Province Jaafari et al. [2018] DT, DT, NB Zagros Mountains, Iran Kahiu and Hanan [2018] BRT sub Saharan Africa Leuenberger et al. RF, ANN Dao, Lafoes, Portugal [2018] Ngoc Thach et al. [2018] SVM, RF, ANN Thuan Chau district, Vietnam Parks et al. [2018] BRT Western USA Sachdeva et al. [2018] BRT, ANN, RF, SVM, Northern India DT, NB, GA, PSO Tehrany et al. [2018] LB, SVM, RF Lao Cai Province, Vietnam Tracy et al. [2018] MAXENT Western North America Vacchiano et al. [2018] MAXENT Aosta Valley, Northern Italy Fernandez-Manso et al. MAXENT Valencia, Spain [2019] Ghorbanzadeh et al. ANN Mazandaran Province, Northern [2019a] Iran Ghorbanzadeh et al. ANN, SVM, RF Mazandaran Province, Northern [2019b] Iran Gigovi c et al. [2019] SVM, RF Tara National Park, Serbia Jaafari [2019] RF, SVM Zagros Mountains, Iran Jaafari et al. [2019] NFM, GA Hyrcanian ecoregion, Iran Kim et al. [2019] MAXENT, RF South Korea Lim et al. [2019] MAXENT South Korea Mart n et al. [2019] MAXENT Northeast Spain Mpakairi et al. [2019] MAXENT northwestern Zimbabwe Quintano et al. [2019] MAXENT La Cabrera, Spain Rihan et al. [2019] RF Mongolian Plateau Syphard et al. [2019] MAXENT California, USA Zhang et al. [2019] CNN, RF, SVM, ANN Yunnan Province, China S.3.4 Landscape controls on re 76 Citation ML methods used Study Region Amatulli et al. [2006] DT Gargano Peninsula, Italy Amatulli and Camia DT Tuscany, Italy [2007] Archibald et al. [2009] RF Southern Africa Holden et al. [2009] RF Gila National Forest, New Mexico, USA Li et al. [2009] ANN Japan Parisien and Moritz MAXENT, BRT USA [2009] Vasilakos et al. [2009] ANN Lesvos Island, Greece Dlamini [2010] BN Swaziland Aldersley et al. [2011] DT, RF global Dimuccio et al. [2011] ANN Central Portugal Sitanggang and Ismail DT Rokan Hilir District, Sumatra, In- [2011] donesia Viedma et al. [2011] BRT, DT Guadalajara province, Central Spain Bisquert et al. [2012] ANN Galicia, Northwest Spain Moritz et al. [2012] MAXENT Global Oliveira et al. [2012] RF Mediterranean Europe Parisien et al. [2012] MAXENT Western USA Renard et al. [2012] MAXENT Western Ghats, India Syphard et al. [2012] MAXENT Southern California, USA Bar Massada et al. RF, MAXENT Michigan, USA [2013] Batllori et al. [2013] MAXENT Mediterranen ecosystems, Global Liu et al. [2013] BRT Great Xing'an Mountains, North- eastern China Luo et al. [2013] RF global Peters et al. [2013] MAXENT Northeast USA Sitanggang et al. [2013] DT Rokan Hilir District, Sumatra, In- donesia Syphard et al. [2013] MAXENT South coast ecoregion, San Diego County, USA Arpaci et al. [2014] RF, MAXENT Tyrol, European Alps Lydersen et al. [2014] RF central Sierra Nevada, California, USA Maxwell et al. [2014] RF Parisien et al. [2014] DT Canada Rodrigues and De la RF, BRT, SVM Peninsular Spain Riva [2014] Wu et al. [2014] RF, DT Great Xing'an Mountains, China Arganaraz ~ et al. [2015] BRT Central Argentina Chen et al. [2015] MAXENT Daxinganling Mountains, North- eastern China Continued on next page 77 Citation ML methods used Study Region Chingono and Mbohwa MAXENT Southern Africa [2015] Curt et al. [2015] BRT New Caledonia Duane et al. [2015] MAXENT Catalonia, Spain Kane et al. [2015] RF Sierra Nevada, California, USA Liu and Wimberly BRT Western USA [2015] Parks et al. [2015] BRT Western USA Viedma et al. [2015] BRT Guadalajara province, Central Spain Vijayakumar et al. RF central Quebec, Canada [2015] Wu et al. [2015] KNN, DT Great Xing'an Mountains, China Bashari et al. [2016] BN Isfahan province, Iran Coppoletta et al. [2016] RF, DT Northern Sierra Nevada, California, USA Curt et al. [2016] BRT Southeastern France Fernandes et al. [2016] BRT Portugal Fonseca et al. [2016] MAXENT Brazil Goldarag et al. [2016] ANN Golestan province, Northern Iran Guo et al. [2016b] RF Daxing'an Mountains, Northeast China Guo et al. [2016a] RF Fujian province, China Miquelajauregui et al. RF, DT central Quebec, Canada [2016] Pourtaghi et al. [2016] BRT, RF Golestan province, Northern Iran Satir et al. [2016] ANN Upper Seyhan Basin, Turkey Syphard et al. [2016] MAXENT South coast ecoregion, San Diego county, USA Vilar et al. [2016] MAXENT Madrid region, Spain Adab [2017] ANN Northeast Iran Cao et al. [2017] ANN, RF Yunnan Province, China Davis et al. [2017] MAXENT Paci c Northwest, USA Dwomoh and Wimberly BRT Upper Guinean Region, West Africa [2017] Ebrahimy et al. [2017] MAXENT Eastern Azerbaijan Forkel et al. [2017] RF, GA global Harris and Taylor [2017] RF Sierra Nevada, California, USA Leys et al. [2017] RF Central Great Plains, USA Li et al. [2017] MAXENT Yunnan Province, Southwest China Lydersen et al. [2017] RF central Sierra Nevada, California, USA MOSTAFA et al. [2017] SVM Golestan province, Northern Iran Nelson et al. [2017] DT, BRT, RF British Columbia, Canada Peters and Iverson MAXENT Northeast USA [2017] Song et al. [2017] RF Hefei City, China Continued on next page 78 Citation ML methods used Study Region Tien Bui et al. [2017] NFM, PSO, RF, SVM Lam Dong province, Vietnam Valdez et al. [2017] RF Honduras Young et al. [2017] BRT Alaska, USA Adab et al. [2018] MAXENT Mazandaran province, Iran Fang et al. [2018] BRT Great Xing'an Mountains, China Hong et al. [2018] SVM, RF, GA Southwest Jiangxi Province Jaafari et al. [2018] DT, DT, NB Zagros Mountains, Iran Kahiu and Hanan [2018] BRT sub Saharan Africa Masrur et al. [2018] RF circumpolar arctic Mayr et al. [2018] DT, RF, SVM, KM Namibia Parks et al. [2018] BRT Western USA Tehrany et al. [2018] LB, SVM, RF Lao Cai Province, Vietnam Tracy et al. [2018] MAXENT Western North America Vacchiano et al. [2018] MAXENT Aosta Valley, Northern Italy Ying et al. [2018] RF China Clarke et al. [2019] MAXENT Southeast Australia Fernandez-Manso et al. MAXENT Valencia, Spain [2019] Forkel et al. [2019] RF global Garc a-Llamas et al. RF Northwest Spain [2019] Ghorbanzadeh et al. ANN Mazandaran Province, Northern [2019a] Iran Ghorbanzadeh et al. ANN, SVM, RF Mazandaran Province, Northern [2019b] Iran Gigovi c et al. [2019] SVM, RF Tara National Park, Serbia Jaafari [2019] RF, SVM Zagros Mountains, Iran Jaafari et al. [2019] NFM, GA Hyrcanian ecoregion, Iran Kim et al. [2019] MAXENT, RF South Korea Mansuy et al. [2019] MAXENT North America Molina et al. [2019] MAXENT Andalusia, southern Spain Mpakairi et al. [2019] MAXENT northwestern Zimbabwe Rihan et al. [2019] RF Mongolian Plateau Syphard et al. [2019] MAXENT California, USA Young et al. [2019] BRT Alaska, USA Zhang et al. [2019] CNN, RF, SVM, ANN Yunnan Province, China S.4. Fire Behaviour Prediction S.4.1 Fire Spread and Growth Citation ML methods used Study Region Vakalis et al. [2004] ANN Attica region, Greece Abdalhaq et al. [2005] GA not speci ed Rodriguez et al. [2008] GA Catalonia, Spain Continued on next page 79 Citation ML methods used Study Region Markuzon and Kolitz RF, BN, KNN Southwest USA [2009] Rodr guez et al. [2009] GA Catalonia, Spain Cencerrado et al. [2012] GA Ashley National Forest, Utah, USA Denham et al. [2012] GA Gestosa, Portugal Cencerrado et al. [2013] GA Catalonia, Spain Kozik et al. [2013] ANN not speci ed Art es et al. [2014] GA Northeast Spain Cencerrado et al. [2014] GA Catalonia, Spain Kozik et al. [2014] RNN not speci ed Ascoli et al. [2015] GA Southern Europe Chetehouna et al. [2015] ANN Experimental data Art es et al. [2016] GA Northeast Catalonia, Spain Carrillo et al. [2016] GA Arkadia region, Greece Art es et al. [2017] GA Northeast Catalonia, Spain Subramanian and RL Northern Alberta, Canada Crowley [2017] Zheng et al. [2017] ANN western USA Denham and Laneri GA Northern Patagonia Andean region [2018] Subramanian and RL Northern Alberta, Canada Crowley [2018] Hodges and Lattimer CNN California, USA [2019] Khakzad [2019] BN Canada Radke et al. [2019] CNN USA S.4.2 Burned area and re severity prediction Citation ML methods used Study Region Cortez and Morais DT, RF, ANN, SVM Montesinho natural park, Portugal [2007] Yu et al. [2011] SOM, ANN Montesinho natural park, Portugal Ozbayo glu and Bozer ANN, SVM Turkey [2012] Sa and Bouroumi ANN Montesinho natural park, Portugal [2013] Zwirglmaier et al. [2013] BN Cyprus Shidik and Mustofa ANN, NB, DT, RF, KNN, Montesinho natural park, Portugal [2014] SVM Xie and Shi [2014] SVM Guangzhou City area, China Alberg [2015] DT Montesinho natural park, Portugal Castelli et al. [2015] GA Montesinho natural park, Portugal Naganathan et al. SVM, DT, KNN USA [2016] Storer and Green [2016] ANN, PSO Montesinho natural park, Portugal Continued on next page 80 Citation ML methods used Study Region Mitsopoulos and Malli- BRT, RF Greece nis [2017] Al Janabi et al. [2018] ANN, SVM Montesinho natural park, Portugal Li et al. [2018a] DT Montesinho natural park, Portugal Toujani et al. [2018] HMM, SOM northwest Tunisia Zald and Dunn [2018] RF southwest Oregon, USA Coeld et al. [2019] DT, RF, ANN, KNN, BRT Alaska, USA Liang et al. [2019] ANN, RNN, LSTM Alberta, Canada Xie and Peng [2019] BRT, DNN, DT, SVM, Montesinho Natural Park, Portugal ANN, RF S.5. Fire E ects S.5.1 Soil erosion and deposits Citation ML methods used Study Region Mallinis et al. [2009] DT, KM Greece Buckland et al. [2019] ANN Nebraska, USA Quintano et al. [2019] MAXENT La Cabrera, Spain S.5.2 Smoke and particulate levels Citation ML methods used Study Region Reid et al. [2015] RF, BAG, BRT, SVM, GP, California, USA KNN Lozhkin et al. [2016] ANN Irkutsk Region, Russia Yao et al. [2018b] RF British Columbia, Canada Yao et al. [2018a] RF British Columbia, Canada Fuentes et al. [2019] ANN South Australia Watson et al. [2019] BRT, RF, SVM, KNN, Northern California, USA ANN Zou et al. [2019] BRT, RF Paci c Northwest, USA S.5.3 Post- re regeneration and ecology Citation ML methods used Study Region Johnstone et al. [2010] BRT Alaska, USA Thompson and Spies RF, DT northwest California, southwest [2010] Oregon, USA Barrett et al. [2011] RF Alaska Perry et al. [2012] BRT New Zealand Reside et al. [2012] MAXENT Australia Continued on next page 81 Citation ML methods used Study Region Sherrill and Romme BRT Dinosaur National Monument, USA [2012] Cai et al. [2013] BRT Huzhong National Reserve, China Debouk et al. [2013] ANN Catalonia, Spain Jung et al. [2013] GA, RF Central Siberian Plateau, Russia Chapin et al. [2014] RF Alaska, USA Liu and Yang [2014] BRT Huzhong Natural Reserve, China Han et al. [2015] RF Yunnan province, China Hermosilla et al. [2015] RF Saskatchewan, Canada Divya and Vijayalak- NB India shmi [2016] Mart n-Alc on and Coll RF Catalonia, Spain [2016] Vijayakumar et al. RF Quebec, Canada [2016] Fairman et al. [2017] RF Victoria, Australia Luo et al. [2017] DT Cangshan Mountain, China Papakosta et al. [2017] BN Cyprus Jo~ ao et al. [2018] RF northern Portugal Poon et al. [2018] SVM San Bernardino, California, USA Cardil et al. [2019] BRT southwestern Europe Magadzire et al. [2019] MAXENT Cape Floristic Region, South Africa S.5.4 Socioeconomic e ects Citation ML methods used Study Region Hradsky et al. [2017] BN Otway Ranges, Australia S.6. Fire Management S.6.1 Planning and policy Citation ML methods used Study Region Penman et al. [2011] BN Wollemi National Park, Australia Bao et al. [2015] GA Longdong Forest Park, Guangzhou, China Ru ault and Mouillot [2015] Bradley et al. [2016] RF western USA McGregor et al. [2016] MDP not speci ed McGregor et al. [2017] RF Deschutes National Forest, Oregon, USA S.6.2 Fuel treatment 82 Citation ML methods used Study Region Penman et al. [2014] BN Southeastern Australia Arca et al. [2015] GA Southern Sardinia, Italy Lauer et al. [2017] RL southwest Oregon S.6.3 Wild re preparedness and response Citation ML methods used Study Region Homchaudhuri et al. GA not speci ed [2010] Costafreda-Aumedes ANN Spain et al. [2015] Penman et al. [2015] BN Sydney Basin Bioregion, Australia O'Connor et al. [2017] BRT, MAXENT Southern Idaho, Northern Nevada, USA Julian and Kochender- RL, DL simulation fer [2018b] Rodrigues et al. [2019] BN Catalonia, Spain S.6.4 Social factors Citation ML methods used Study Region Delgado et al. [2018] BN Spain http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Statistics arXiv (Cornell University)

A review of machine learning applications in wildfire science and management

Loading next page...
 
/lp/arxiv-cornell-university/a-review-of-machine-learning-applications-in-wildfire-science-and-2fBoUrnVLi
ISSN
1181-8700
eISSN
ARCH-3347
DOI
10.1139/er-2020-0019
Publisher site
See Article on Publisher Site

Abstract

A review of machine learning applications in wild re science and management 1,2 2 y3 Piyush Jain , Sean C P Coogan , Sriram Ganapathi Subramanian , Mark Crowley z3 4 2 , Steve Taylor , and Mike D Flannigan Natural Resources Canada,Canadian Forest Service, Northern Forestry Centre, Edmonton, AB Canadian Partnership for Wildland Fire Science, University of Alberta, Renewable Resources, Edmonton, AB University of Waterloo, Electrical and Computer Engineering, Waterloo, ON Natural Resources Canada,Canadian Forest Service, Paci c Forestry Centre, Victoria, BC August 20, 2020 Abstract Arti cial intelligence has been applied in wild re science and management since the 1990s, with early applications including neural networks and expert systems. Since then the eld has rapidly progressed congruently with the wide adoption of machine learning (ML) methods in the environmental sciences. Here, we present a scoping review of ML applications in wild re science and management. Our overall objective is to improve awareness of ML methods among wild re researchers and managers, as well as illustrate the diverse and challenging range of problems in wild re science available to ML data scientists. To that end, we rst present an overview of popular ML approaches used in wild re science to date, and then review the use of ML in wild re science as broadly categorized into six problem domains, including: 1) fuels characterization, re detection, and mapping; 2) re weather and climate change; 3) re occurrence, susceptibility, and risk; 4) re behavior prediction; 5) re e ects; and 6) re management. Furthermore, we discuss the advantages and limitations of various ML approaches relating to data size, computational requirements, generalizability, and interpretability, as well as identify opportunities for future advances in the science and management of wild res within a data science context. In total, we identi ed 300 relevant publications up to the end of 2019, where the most frequently used ML methods across problem domains included random forests, MaxEnt, arti cial neural networks, decision trees, support vector machines, and genetic algorithms. As such, there exists opportunities to apply more current ML methods | including deep learning and agent based learning | in the wild re sciences, especially in instances involving very large multivariate datasets. We must recognize, however, that despite the ability of ML methods to learn on their own, expertise in wild re science is necessary to ensure realistic modelling of re processes across multiple scales, while the complexity of some ML methods, such as deep learning, requires a dedicated and sophisticated knowledge of their application. Finally, we stress that the wild re research and management communities play an active role in providing relevant, high quality, and freely available wild re data for use by practitioners of ML methods. Keywords: machine learning, wild re science, re management, wildland re, support vector ma- chine, arti cial neural network, decision trees, Bayesian networks, reinforcement learning, deep learning piyush.jain@canada.ca ss@uwaterloo.ca mcrowley@uwaterloo.ca arXiv:2003.00646v2 [cs.LG] 19 Aug 2020 1 Introduction Wildland re is a widespread and critical element of the earth system [Bond and Keeley, 2005], and is a continuous global feature that occurs in every month of the year. Presently, global annual area burned is estimated to be approximately 420 Mha [Giglio et al., 2018], which is greater in area than the country of India. Globally, most of the area burned by wild res occurs in grasslands and savannas. Humans are responsible for starting over 90% of wildland res, and lightning is responsible for almost all of the remaining ignitions. Wildland res can result in signi cant impacts to humans, either directly through loss of life and destruction to communities, or indirectly through smoke exposure. Moreover, as the climate warms we are seeing increasing impacts from wildland re [Coogan et al., 2019]. Consequently, billions of dollars are spent every year on re management activities aimed at mitigating or preventing wild res' negative e ects. Understanding and better predicting wild res is therefore crucial in several important areas of wild re management, including emergency response, ecosystem management, land-use planning, and climate adaptation to name a few. Wildland re itself is a complex process; its occurrence and behaviour are the product of several interrelated factors, including ignition source, fuel composition, weather, and topography. Furthermore, re activity can be examined viewed across a vast range of scales, from ignition and combustion processes that occur at a scale of centimeters over a period of seconds, to re spread and growth over minutes to days from meters to kilometers. At larger extents, measures of re frequency may be measured over years to millennia at regional, continental, and planetary scales (see Simard [1991] for a classi cation of re severity scales, and Taylor et al. [2013] for a review of numerical and statistical models that have been used to characterize and predict re activity at a range of scales). For example, combustion and re behavior are fundamentally physicochemical processes that can be usefully represented in mechanistic (i.e., physics- based) models at relatively ne scales [Coen, 2018]. However, such models are often limited both by the ability to resolve relevant physical processes, as well as the quality and availability of input data [Ho man et al., 2016]. Moreover, with the limitations associated with currently available computing power it is not feasible to apply physical models to inform re management and research across the larger and longer scales that are needed and in near real time. Thus, wild re science and management rely heavily on the development of empirical and statistical models for meso, synoptic, strategic, and global scale processes [Simard, 1991], the utility of which are dependent upon their ability to represent the often complex and non-linear relationships between the variables of interest, as well as by the quality and availability of data. While the complexities of wildland re often present challenges for modelling, signi cant advances have been made in wild re monitoring and observation primarily due to the increasing availability and capability of remote-sensing technologies. Several satellites (eg. NASA TERRA, AQUA and GOES), for instance, have onboard re detection sensors (e.g., Advanced Very High Resolution Radiometer (AVHRR), Moderate Resolution Imaging Spectroradiometer (MODIS), Visible Infrared Imaging Radiometer Suite (VIIRS)), and these sensors along with those on other satellites (e.g., LANDSAT series) routinely monitor vegetation distributions and changes. Additionally, improvements in numerical weather prediction and climate models are simultaneously o ering smaller spatial resolutions and longer lead forecast times [Bauer et al., 2015] which potentially o er improved predictability of extreme re weather events. Such developments make a data-centric approach to wild re modeling a natural evolution for many research problems given sucient data. Consequently, there has been a growing interest in the use of Machine Learning (ML) methodologies in wild re science and management in recent years. Although no formal de nition exists, we adopt the conventional interpretation of ML as the study of computer algorithms that can improve automatically through experience [Mitchell, 1997]. This approach is necessarily data-centric, with the performance of ML algorithms dependent on the quality and quantity of available data relevant to the task at hand. The eld of ML has undergone an explosion of new algorithmic advances in recent years and is deeply connected to the broader eld of Arti cial Intelligence (AI). AI researchers aim to understand and synthesize intelligent agents which can act appropriately to their situation and objectives, adapt to changing environments, and learn from experience [Poole and Mackworth, 2 2010]. The motivations for using AI for forested ecosystem related research, including disturbances due to wild re, insects, and disease, were discussed in an early paper [Schmoldt, 2001], while Olden et al. [2008] further argued for the use of ML methods to model complex problems in ecology. The use of ML models in the environmental sciences has seen a rapid uptake in the last decade, as is evidenced by recent reviews in the geosciences [Karpatne et al., 2017], forest ecology [Liu et al., 2018], extreme weather prediction [McGovern et al., 2017], ood forecasting [Mosavi et al., 2018], statistical downscaling [Vandal et al., 2018], remote sensing [Lary et al., 2016], and water resources [Shen, 2018, Sun and Scanlon, 2019]. Two recent perspectives have also made compelling arguments for the application of deep learning in earth system sciences [Reichstein et al., 2019] and for tackling climate change [Rolnick et al., 2019]. To date, however, no such paper has synthesized the diversity of ML approaches used in the various challenges facing wildland re science. In this paper, we review the current state of literature on ML applications in wild re science and management. Our overall objective is to improve awareness of ML methods among re researchers and managers, and illustrate the diverse and challenging problems in wild re open to data scientists. This paper is organized as follows. In Section 2, we discuss commonly used ML methods, focusing on those most commonly encountered in wild re science. In Section 3, we give an overview of the scoping review and literature search methodology employed in this paper. In this section we also highlight the results of our literature search and examine the uptake of ML methods in wild re science since the 1990s. In Section 4, we review the relevant literature within six broadly categorized wild re modeling domains: (i) Fuels characterization, re detection, and mapping; (ii) re weather and climate change; (iii) re probability and risk; (iv) re behavior prediction; (v) re e ects; and (vi) re management. In Section 5, we discuss our ndings and identify further opportunities for the application of ML methods in wild re science and management. Finally, in Section 6 we o er conclusions. Thus, this review will serve to guide and inform both researchers and practitioners in the wild re community looking to use ML methods, as well as provide ML researchers the opportunity to identify possible applications in wild re science and management. 2 Arti cial Intelligence and Machine Learning \De nition: Machine Learning - (Methods which) detect patterns in data, use the uncov- ered patterns to predict future data or other outcomes of interest" from Machine Learning: A Probabilistic Perspective, 2012 [Murphy, 2012]. ML itself can be seen as a branch of AI or statistics, depending who you ask, that focuses on building predictive, descriptive, or actionable models for a given problem by using collected data, or incoming data, speci c to that problem. ML methods learn directly from data and dispense with the need for a large number of expert rules or the need to model individual environmental variables with perfect accuracy. ML algorithms develop their own internal model of the underlying distributions when learning from data and thus need not be explicitly provided with physical properties of di erent parameters. Take for example, the task of modeling wildland re spread, the relevant physical properties which include fuel composition, local weather and topography. The current state of the art in wild re prediction includes physics-based simulators that re ghters and strategic planners rely on to take many critical decisions regarding allocation of scarce re ghting resources in the event of a wild re [Sullivan, 2007]. These physics- based simulators, however, have certain critical limitations; they normally render very low accuracies, have a prediction bias in regions where they are designed to be used, are often hard to design and implement due to the requirement of large number of expert rules. Furthermore, modelling many complex environmental variables is often dicult due to large resource requirements and complex or heterogeneous data formats. ML algorithms, however, learn their own mappings between parametric rules directly from data and do not require expert rules, which is particularly advantageous when the number of parameters are quite large and their physical properties quite complex, as in the case of wildland re. Therefore, a ML approach to wild re response may help to avoid many of the limitations of physics-based simulators. 3 A major goal of this review is to provide an overview of the various ML methods utilized in wild re sci- ence and management. Importantly, we also provide a generalized framework for guiding wild re scientists interested in applying ML methods to speci c problem domains in wildland re research. This conceptual framework, derived from the approach in [Murphy, 2012] and modi ed to show examples relevant to wild- land re and management is shown in Fig. 1. In general, ML methods can be identi ed as belonging to one of three types: supervised learning; unsupervised learning; or, agent based learning. We describe each of these below. Supervised Learning - In supervised ML all problems can be seen as one of learning a parametrized function, often called a \model", that maps inputs (i.e., predictor variables) to outputs (or \target vari- ables") both of which are known. The goal of supervised learning is to use an algorithm to learn the parameters of that function using available data. In fact, both linear and logistic regression can be seen as very simple forms of supervised learning. Most of the successful and popular ML methods fall into this category. Unsupervised Learning - If the target variables are not available, then ML problems are typically much harder to solve. In unsupervised learning, the canonical tasks are dimensionality reduction and clustering, where relationships or patterns are extracted from the data without any guidance as to the \right" answer. Extracting embedded dimensions which minimize variance, or assigning datapoints to (labelled) classes which maximize some notion of natural proximity or other measures of similarity are examples of unsupervised ML tasks. Agent Based Learning - Between supervised and unsupervised learning are a group of ML methods where learning happens by simulating behaviors and interactions of a single or a group of autonomous agents. These are general unsupervised methods which use incomplete information about the target vari- ables, (i.e., information is available for some instances but not others), requiring generalizable models to be learned. A speci c case in this space is Reinforcement Learning [Sutton and Barto, 1998], which is used to model decision making problems over time where critical parts of the environment can only be observed interactively through trial and error. This class of problems arises often in the real world and require ecient learning and careful de nition of values (or preferences) and exploration strategies. In the next section, we present a brief introduction to commonly used ML methods from the aforemen- tioned learning paradigms. We note that this list is not meant to be exhaustive, and that some methods can accommodate both supervised and unsupervised learning tasks. It should be noted that the classi - cation of a method as belonging to either ML or traditional statistics is often a question of taste. For the purpose of this review | and in the interests of economy | we have designated a number of methods as belonging to traditional statistics rather than ML. For a complete listing see tables 1 and 2. 2.1 Decision Trees Decision Trees (DT) [Breiman, 2017] belong to the class of supervised learning algorithms and are another example of a universal function approximator, although in their basic form such universality is dicult to achieve. DTs can be used for both classi cation and regression problems. A decision tree is a set of if-then- else rules with multiple branches joined by decision nodes and terminated by leaf nodes. The decision node is where the tree splits into di erent branches, with each branch corresponding to the particular decision being taken by the algorithm whereas leaf nodes represent the model output. This could be a label for a classi cation problem or a continuous value in case of a regression problem. A large set of decision nodes is used in this way to build the DT. The objective of DTs are to accurately capture the relationships between input and outputs using the smallest possible tree that avoids over tting. C4.5 [Quinlan, 1993] and Classi cation and Regression Trees (CART, [Breiman et al., 1984]) are examples of common single DT algorithms. Note that while the term CART is also used as an umbrella term for single tree methods, we use DT here to refer to all such methods. The majority of decision tree applications are ensemble decision tree (EDT) models that use multiple trees in parallel (ie. bootstrap aggregation or bagging) or sequentially (ie., boosting) to arrive at a nal model. In this way, EDTs make use of many weak learners to form a strong learner while being robust to over tting. EDTs are well described in many ML/AI textbooks and 4 Machine Learning Types Agent-based learning Supervised Learning Unsupervised Learning Map labelled input to known Understand patterns and Single or multiple agents output discover output interact with environment Continuous Categorical Reward based rather than Target Target Target variable not available target action Variable Variable Dimensionality Decision Regression Classification Clustering Optimization Reduction Making NB, DT, CART, SOM, DQN, A3C, ANN, DT, BRT, KM, SOM, GA, MCTS, A3C RF, DNN, GP autoencoders, MCTS RF, KNN, SVM, autoencoders, ANN, GA, RNN, t-SNE, RF, BRT, K-SVM, LR, LDA GMM, ISODATA, MAXENT, CA, MaxEnt, PCA, HMM, HC, PCA, MLR, GLM, GAM factor analysis DBSCAN - Fire susceptibility - Fuels - Fire Detection - Optimizing fire - Fuel treatment - Landscape - Fire Spread/Burn characterization - Fire mapping simulators - Planning and controls on fire area prediction - Fire detection - burned area - Fire spread and Policy - Fire susceptibility - Fire occurence - Fire mapping prediction growth - Wildfire response - Fire Spread/Burn - Fire severity - Fire weather area prediction - Smoke Prediction prediction - Climate Change Figure 1: A diagram showing the main machine learning types, types of data, and modeling tasks in relation to popular algorithms and potential applications in wild re science and management. Note that the algorithms shown bolded are core ML methods whereas non-bolded algorithms are often not considered ML. are widely available as implemented libraries. 2.1.1 Random Forests A Random Forest (RF) [Breiman, 2001] is an ensemble model composed of a many individually trained DTs, and is the most popular implementation of a bagged decision tree. Each component DT in a RF model makes a classi cation decision where the class with the maximum number of votes is determined to be the nal classi cation for the input data. RFs can also be used for regression where the nal output is determined by averaging over the individual tree outputs. The underlying principle of the RF algorithm is that a random subset of features is selected at each node of each tree; the samples for training each component tree are selected using bagging, which resamples (with replacement) the original set of datapoints. The high performance of this algorithm is achieved by minimizing correlation between trees while reducing model variance so that a large number of di erent trees provides greater accuracy than individual trees. However, this improved performance comes at the cost of an increase in bias and loss of interpretability (although variable importance can still be inferred through permutation tests). Possible Popular Task Data Approach Applications Algorithms Machine Learning Methods A3C Asynchronous Advantage Actor-Critic AdaBoost Adaptive Boosting ANFIS Adaptive Neuro Fuzzy Inference System ANN Arti cial Neural Networks ADP Approximate Dynamic Programming (a.k.a. reinforcement learning) Bag Bagged Decision Trees BN Bayesian Networks BRT Boosted Regression Trees (a.k.a. Gradient Boosted Machine) BULC Bayesian Updating of Land Cover CART Classi cation and Regression Tree CNN Convolutional Neural Network DNN Deep Neural Network DQN Deep Q-Network DT Decision Trees (incl. CART, J48, jRip) EDT Ensemble Decision Trees (incl. bagging and boosting) ELM Extreme Machine Learning (i.e., feedforward network) GA Genetic algorithms (a.k.a evolutionary algorithms) GBM Gradient Boosted Machine (a.k.a. Boosted Regression Trees, incl. XGBoost, AdaBoost, LogitBoost) GMM Gaussian Mixture Models GP Gaussian Processes HCL Hard Competitive Learning HMM Hidden Markov Models ISODATA Iterative Self-Organizing DATA algorithm KNN K Nearest Neighbor KM K-means Clustering LB LogitBoost (incl. AdaBoost) LSTM Long Short Term Memory MaxEnt Maximum Entropy MCMC Markov Chain Monte Carlo MCTS Monte Carlo Tree Search MLP Multilayer Perceptron MDP Markov Decision Process NB Naive Bayes NFM Neuro-Fuzzy models PSO Particle Swarm Optimization RF Random Forest RL Reinforcement Learning RNN Recurrent Neural Network SGB Stochastic Gradient Boosting SOM Self-organizing Maps SVM Support Vector Machines t-SNE T-distributed Stochastic Neighbor Embedding Table 1: Table of acronyms and de nitions for common machine learning algorithms referred to in text. 2.1.2 Boosted Ensembles Boosting describes a strategy where one combines a set of weak learners | usually decision trees | to make a strong learner using a sequential additive model. Each successive model improves on the previous 6 Non-machine learning methods DBSCAN Density-based spatial clustering of applications with noise GAM Generalized Additive Model GLM Generalized Linear Model KLR Kernel Logistic Regression LDA Linear Discriminant Analysis LR Logistic Regression MARS Multivariate Adaptive Regression Splines MLR Multiple Linear Regression PCA Principal Component Analysis SLR Simple Linear regression Table 2: Table of acronyms and de nitions for common data analysis algorithms usually considered as foundational to, or outside of, machine learning itself. by taking into account the model errors from the previous model, which can be done in more than one way. For example, the adaptive boosting algorithm, known as AdaBoost [Freund and Shapire, 1995], works by increasing the weight of observations that were previously misclassi ed. This can in principle reduce the classi cation error leading to a high level of precision [Hastie et al., 2009]. Another very popular implementation for ensemble boosted trees is Gradient Boosting Machine (GBMs), which makes use of the fact that each DT model represents a function that can be di erentiated with re- spect to its parameters, i.e., how much a change in the parameters will change the output of the function. GBMs sequentially build an ensemble of multiple weak learners by following a simple gradient which points in the opposite direction to weakest results of the current combined model [Friedman, 2001]. The details for the GBM algorithm are as follows. Denoting the target output as Y , and given a tree-based ensemble model, represented as a function T (X ) ! Y , after adding i weak learners already, the \perfect" function for the (i + 1)th weak learner would be h(x) = T (x) Y which exactly corrects the previous model (i.e., T (x) = T (x) + h(x) = Y ). In practice, we can only approach this perfect update (i+1) i by performing functional gradient descent where we use an approximation of the true residual (i.e., loss function) at each step. In our case this approximation is simply the sum of the residuals from each weak learner decision tree L(Y; T (X )) = Y T (X ). GBM explicitly uses the gradient r L(Y; T (X ) of the i T i loss function of each tree to t a new tree and add it to the ensemble. In a number of domains, and particularly in the context of ecological modeling GBM is often referred to as Boosted Regression Trees (BRTs) [Elith et al., 2008]. For consistency with the majority of literature reviewed in this paper we henceforth use the latter term. It should be noted that while deep neural networks (DNNs) and EDT methods are both universal function approximators, EDTs are more easily interpretable and faster to learn with less data than DNNs. However, there are fewer and fewer cases where trees-based methods can be shown to provide superior performance on any particular metric when DNNs are trained properly with enough data (see for example, Korotcov et al. [2017]). 2.2 Support Vector Machines Another category of supervised learning includes Support Vector Machines (SVM) [Hearst et al., 1998] and related kernel-based methods. SVM is a classi er that determines the hyper-plane (decision boundary) in an n-dimensional space separating the boundary of each class, for data in n dimensions. SVM nds the optimal hyper-plane in such a way that the distance between the nearest point of each class to the decision boundary is maximized. If the data can be separated by a line then the hyper-plane is de ned to be of the form w x + b = 0 where the w is the weight vector, x is the input vector and b is the bias. The distance of the hyper-plane to the closest data point d, called a support vector, is de ned as the margin of separation. The objective is to nd the optimal hyper-plane that minimizes the margin. If they are 7 not linearly separable, kernel SVM methods such as Radial Basis Functions (RBF) rst apply a set of transformations to the data to a higher dimensional space where nding this hyperplane would be easier. SVMs have been widely used for both classi cation and regression problems, although recently developed deep learning algorithms have proved to be more ecient than SVMs given a large amount of training data. However, for problems with limited training samples, SVMs might give better performances than deep learning based classi ers. 2.3 Arti cial Neural Networks and Deep Learning The basic unit of an Arti cial Neural Network (ANN) is a neuron (also called a perceptron or logistic unit). A neuron is inspired by the functioning of neurons in mammalian brains in that it can learn simple associations, but in reality it is much simpler than its biological counterpart. A neuron has a set of inputs which are combined linearly through multiplication with weights associated with the input. The nal weighted sum forms the output signal which is then passed through a (generally) non-linear activation function. Examples of activation functions include sigmoid, tanh, and the Recti ed Linear Unit (ReLU). This non-linearity is important for general learning since it creates an abrupt cuto (or threshold) between positive and negative signals. The weights on each connection represent the function parameters which are t using supervised learning by optimizing the threshold so that it reaches a maximally distinguishing value. In practice, even simple ANNs, often called Multi-Layered Perceptrons (MLP), combine many neuron units in parallel, each processing the same input with independent weights. In addition, a second layer of hidden neuron units can be added to allow more degrees of freedom to t general functions, see Figure 2(a). MLPs are capable of solving simple classi cation and regression problems. For instance, if the task is one of classi cation, then the output is the predicted class for the input data, whereas in the case of a regression task the output is the regressed value for the input data. Deep learning [LeCun et al., 2015] refers to using Deep Neural Networks (DNNs) which are ANNs with multiple hidden layers (nominally more than 3) and include Convolutional Neural Networks (CNNs) popularized in image analysis and Recurrent Neural Networks (RNNs) which can be used to model dynamic temporal phenomena. The architecture of DNNs can vary in connectivity between nodes, the number of layers employed, the types of activation functions used, and many other types of hyperparameters. Nodes within a single layer can be fully connected, or connected with some form of convolutional layer (e.g., CNNs), recurrent units (e.g., RNNs), or other sparse connectivity. The only requirement of all these connectivity structures and activation functions is that they are di erentiable. Regardless of the architecture, the most common process of training a ANN involves processing input data fed through the network layers and activation functions to produce an output. In the supervised setting, this output is then compared to the known true output (i.e., labelled training data) resulting in an error measurement (loss or cost function) used to evaluate model performance. The error for DNNs are commonly calculated as a cross entropy loss between the predicted output label and the true output label. Since every part of the network is mathematically di erentiable we can compute a gradient for the entire network. This gradient is used to calculate the proportional change in each network weight needed to produce an in nitesimal increase in the likelihood of the network producing the same output for the most recent output. The gradient is then weighted by the computed error, and thereafter all the weights are updated in sequence using a backpropagation algorithm [Hecht-Nielsen, 1992]. ANNs can also be con gured for unsupervised learning tasks. For example, self-organizing maps (SOMs) are a form of ANN adapted for dealing with spatial data and have therefore found widespread use in the atmospheric sciences [Ski c and Francis, 2012]. A SOM is a form of unsupervised learning that consists of a two-dimensional array of nodes as the input layer, representing say, a gridded atmospheric variable at a single time. The algorithm clusters similar atmospheric patterns together and results in a dimensionality reduction of the input data. More recently, unsupervised learning methods from deep learning, such as autoencoder networks, are starting to replace SOMs in the environmental sciences [Shen, 2018]. 8 Figure 2: Logistic regression can be seen as basic building block for neural networks, with no hidden layer and a sigmoid activation function. Classic shallow neural networks (also known as Multi-Layer Perceptrons) have at least one hidden layer and can have a variety of activation functions. Deep neural networks essentially have a much larger number of hidden layers as well as use additional regularization and optimization methods to enhance training. 2.4 Bayesian methods 2.4.1 Bayesian Networks Bayesian networks (Bayes net, belief network; BN) are a popular tool in many applied domains because they provide an intuitive graphical language for specifying the probabilistic relationships between variables as well as tools for calculating the resulting probabilities [Pearl, 1988]. The basis of BNs is Bayes' theorem, which relates the conditional and marginal probabilities of random variables. BNs can be treated as a ML task if one is trying to automatically t the parameters of the model from data, or even more challenging, to learn the best graphical structure that should be used to represent a dataset. BNs have close ties to causal reasoning, but it is important to remember that the relationships encoded in a BN are inherently correlational rather than causal. BNs are acyclic graphs, consisting of nodes and arrows (or arcs), de ning a probability distribution over variables U . The set of parents of a node (variable) X , denoted  , are all nodes with directed arcs going into X . BNs provide compact representation of conditional distributions since p(X jX ; : : : ; X ) = p(X j ) where X ; : : : ; X are arranged to be all of the ancestors of X i 1 i1 i X 1 i1 i other than its direct parents. Each node X is associated with a conditional probability table over X and its parents de ning p(Xj ). If a node has no parents, a prior distribution is speci ed for p(X ). The joint probability distribution of the network is then speci ed by the chain rule P (U ) = p(Xj ). X2U 2.4.2 Na ve Bayes A special case of a BN is the Na ve Bayes (NB) classi er, which assumes conditional independence between input features, which allows the likelihood function to be constructed by a simple multiplication of the conditional probability of each input variable conditional on the output. Therefore, while NB is fast 9 and straightforward to implement, prediction accuracy can be low for problems where the assumption of conditional independence does not hold. 2.4.3 Maximum Entropy Maximum Entropy (MaxEnt), originally introduced by Phillips et al. [2006], is a presence only framework that ts a spatial probability distribution by maximising entropy, consistent with existing knowledge. MaxEnt can be considered a Bayesian method since it is compatible with an application of Bayes Theorem as existing knowledge is equivalent to specifying a prior distribution. MaxEnt has found widespread use in landscape ecology species distribution modeling [Elith, Phillips, Hastie, Dud k, Chee, and Yates, 2011], where prior knowledge consists of occurrence observations for the species of interest. 2.5 Reward based methods 2.5.1 Genetic Algorithms Genetic algorithms (GA) are heuristic algorithms inspired by Darwin's theory of evolution (natural selec- tion) and belong to a more general class of evolutionary algorithms [Mitchell, 1996]. GAs are often used to generate solutions to search and optimization problems by using biologically motivated operators such as mutation, crossover, and selection. In general, GAs involve several steps. The rst step involves creating an initial population of potential solutions, with each solution encoded as a chromosome. Second a tness function appropriate to the problem is de ned, which returns a tness score determining how likely an individual is to be chosen for reproduction. The third step requires the selection of pairs of individuals, denoted as parents. In the fourth step, a new population of nite individuals are created by generating two new o spring from each set of parents using crossover, whereby a new chromosome is created by some random selection process from each parents chromosomes. In the nal step called mutation, a small sample of the new population is chosen and a small perturbation is made to the parameters to maintain diversity. The entire process is repeated many times until the desired results are satisfactory (based on the tness function), or some measure of convergence is reached. 2.5.2 Reinforcement Learning Reinforcement learning (RL) represents a very di erent learning paradigm to supervised or unsupervised learning. In RL, an agent (or actor) interacts with its environment and learns a desired behavior (set of actions) in order to maximize some reward. RL is a solution to a Markov Decision Process (MDP) where the transition probabilities are not explicitly known but need to be learned. This type of learning is well suited to problems of automated decision making, such as required for automated control (e.g., robotics) or for system optimization (e.g., management policies). Various RL algorithms include Monte Carlo Tree Search (MTCS), Q-Learning, and Actor-Critic algorithms. For an introduction to RL see Sutton and Barto [2018]. 2.6 Clustering methods Clustering is the process of splitting a set of points into groups where each point in a group is more similar to its own group than any other group. There are di erent ways in which clustering can be done, for example, the K-means (KM) clustering algorithm [MacQueen et al., 1967], based on a centroid model, is perhaps the most well-known clustering algorithm. In K-means, the notion of similarity is based on closeness to the centroid of each cluster. K-means is an iterative process in which the centroid of a group and points belonging to a group are updated at each step. The K-means algorithm consists of ve steps: (i) specify the number of clusters; (ii) each data point is randomly assigned to a cluster; (iii) the centroids of each cluster is calculated; (iv) the points are reassigned to the nearest centroids, and (v) cluster centroids are recomputed. Steps iv and v repeat until no further changes are possible. Although KM is the most widely 10 used clustering algorithm, several other clustering algorithms exist including, for example, agglomerative Hierarchical Clustering (HC), Gaussian Mixture Models (GMMs) and Iterative Self-Organizing DATA (ISODATA). 2.7 Other methods 2.7.1 K-Nearest Neighbor The K-Nearest Neighbors (KNN) algorithm is a simple but very e ective supervised classi cation algorithm which is based on the intuitive premise that similar data points are in close proximity according to some metric [Altman, 1992]. Speci cally, a KNN calculates the similarity of data points to each other using the Euclidean distance between the K nearest data points. The optimal value of K can be found experimentally over a range values using the classi cation error. KNN is widely used in applications where a search query is performed such that results should be similar to another pre-existing entity. Examples of this include nding similar images to a speci ed image and recommender systems. Another popular application of KNN is outlier (or anomaly) detection, whereby the points (in a multidimensional space) farthest away from their nearest neighbours may be classi ed as outliers. 2.7.2 Neuro-Fuzzy models Fuzzy logic is an approach for encoding expert human knowledge into a system by de ning logical rules about how di erent classes overlap and interact without being constrained to \all-or-nothing" notions of set inclusion or probability of occurrence. Although early implementations of fuzzy logic systems depended on setting rules manually, and therefore are not considered machine learning, using fuzzy rules as inputs or extracting them from ML methods are often described as \neuro-fuzzy" methods. For example, the Adaptive Neuro-Fuzzy Inference System (ANFIS) [Jang, 1993] fuses fuzzy logical rules with an ANN approach, while trying to maintain the bene ts of both. ANFIS is a universal function approximator like ANNs. However, since this algorithm originated in the 1990s, it precedes the recent deep learning revolution so is not necessarily appropriate for very large data problems with complex patterns arising in high-dimensional spaces. Alternatively, human acquired fuzzy rules can be integrated into ANNs learning; however, it is not guaranteed that the resulting trained neural network will still be interpretable. It should be noted that fuzzy rules and fuzzy logic are not a major direction of research within the core ML community. 3 Literature search and scoping review The combination of ML and wild re science and management comprises a diverse range of topics in a rela- tively nascent eld of multidisciplinary research. Thus, we employed a scoping review methodology [Arksey and O'Malley, 2005, Levac et al., 2010] for this paper. The goal of a scoping review is to characterize the existing literature in a particular eld of study, particularly when a topic has yet to be extensively reviewed and the related concepts are complex and heterogeneous [Pham, Raji c, Greig, Sargeant, Papadopoulos, and Mcewen, 2014]. Furthermore, scoping reviews can be particularly useful for summarizing and dissem- inating research ndings, and for identifying research gaps in the published literature. A critical review of methodological advances and limitations and comparison with other methods is left for future work. We performed a literature search using the Google Scholar and Scopus databases and the key words \wild- re" or \wildland re" or\forest re" or \bush re" in combination with \machine learning" or \random forest" or \decision trees" or \regression trees" or \support vector machine" or \maximum entropy" or \neural network" or \deep learning" or \reinforcement learning". We also used the Fire Research Institute online database (http://fireresearchinstitute.org) using the following search terms: \Arti cial In- telligence"; \Machine Learning"; \Random Forests"; \Expert Systems"; and \Support Vector Machines". 11 50 1996 2000 2005 2010 2015 2019 Year of publication Figure 3: Number of publications by year for 300 publications on topic of ML and wild re science and management as identi ed in this review. SVM RF Other MAXENT KNN GA DT 60 DL BRT BN ANN 1996 2000 2005 2010 2015 2019 Year of publication Figure 4: Number of ML applications by category and by year for 300 publications on topic of ML and wild re science and management as identi ed in this review. Furthermore, we obtained papers from references cited within papers we had obtained using literature databases. After performing our literature search, we identi ed a total of 300 publications relevant to the topic of ML applications in wild re science and management (see supplementary material for a full bibliography). Furthermore, a search of the Scopus database revealed a dramatic increase in the number of wild re and Number of applications Number of publications ML articles published in recent years (see Fig. 3). After identifying publications for review, we further applied the following criteria to exclude non-relevant or unsuitable publications, including: (i) conference submissions where a journal publication describing the same work was available; (ii) conference posters; (iii) articles in which the methodology and results were not adequately described to conduct an assessment of the study; (iv) articles not available to as either by open access or by subscription; and (v) studies that did not present new methodologies or results. 4 Wild re applications In summary, we found a total of 300 journal papers or conference proceedings on the topic of ML applica- tions in wild re science and management. We found the problem domains with the highest application of ML methods was Fire Occurrence , Susceptibility and Risk (127 papers) followed by Fuels Characterization , Fire Detection And Mapping (66 papers), Fire Behaviour Prediction (43 papers), Fire E ects (35 papers), Fire Weather and Climate Change (20 papers), and Fire Management (16 papers). Within Fire Occur- rence, Susceptibility and Risk, the subdomains with the most papers were Fire Susceptibility Mapping (71 papers) and Landscape Controls on Fire (101 papers). Refer to table 3 and the supplementary material for a break-down of each problem subdomain and ML methods used, as well as study areas considered. 4.1 Fuels Characterization, Fire Detection, and Mapping 4.1.1 Fuels characterization Fires ignite in a few fuel particles; subsequent heat transfer between particles through conduction, radiation and convection, and the resulting re behavior (fuel consumption, spread rate, intensity) is in uenced by properties of the live and dead vegetative fuels, including moisture content, biomass, and vertical and horizontal distribution. Fuel properties are a required input in all re behavior models, whether it be a simple categorical vegetation type, as in the Canadian FBP System, or as physical quantities in a 3 dimensional space (eg. see FIRETEC model). Research to predict fuel properties has been carried out at two di erent scales 1) regression applications to predict quantities such as the crown biomass of single trees from more easily measured variables such as height and diameter, and 2) classi cation applications to map fuel type descriptors or fuel quantities over a landscape from visual interpretation of air photographs or by interpretation of the spectral properties of remote sensing imagery. However, relatively few studies have employed ML to wild re fuel prediction, leaving the potential for substantially more research in this area. In an early study, Riano ~ et al. [2005] used an ANN to predict and map the equivalent water thickness and dry matter content of wet and dry leaf samples from 49 species of broad leaf plants using re ectance and transmittance values in the Ispra region of Italy. Pierce et al. [2012] used RF to classify important canopy fuel variables (e.g. canopy cover, canopy height, canopy base height, and canopy bulk density) related to wildland re in Lassen Volcanic National Park, California, using eld measurements, topographic data, and NDVI to produce forest canopy fuel maps. Likewise, Viegas et al. [2014] used RF with Land re and biophysical variables to perform fuel classi cation and mapping in Eastern Oregon. The authors of the aforementioned study achieved relatively high overall modelling accuracy, for example, 97% for forest height, 86% for forest cover, and 84% for existing vegetation group (i.e. fuel type). L opez-Serrano et al. [2016] compared the performance of three common ML methods (i. SVM; ii. KNN; and iii. RF) and multiple linear regression in estimating above ground biomass in the Sierra Madre Occidental, Mexico. The authors reported the advantages and limitations of each method, concluding that that the non-parametric ML methods had an advantage over multiple linear regression for biomass estimation. Garc a et al. [2011] used SVM to classify LiDAR and multispectral data to map fuel types in Spain. Chirici et al. [2013] compared the use of CART, RF, and Stochastic Gradient Boosting SGB, an ensemble tree method that uses both boosting and bagging, for mapping forest fuel types in Italy, and found that SGB had the highest overall accuracy. 13 Section Domain NFM SVM KM GA BN BRT ANN DT RF KNN MAXENT DL NB Other 1.1 Fuels characterization - 2 - - - 1 1 1 4 1 - - - - 1.2 Fire detection 2 3 1 1 1 - 12 - - - - 18 - 3 1.3 Fire perimeter and severity 1 12 1 2 - 1 6 1 4 2 1 - - 6 mapping 2.1 Fire weather prediction - - 1 - - - - - 1 - - - - 3 2.2 Lightning prediction - - - - - - - 1 2 - - - - - 2.3 Climate change - 1 - - - 6 2 2 5 - 7 - - - 3.1 Fire occurrence prediction - 3 - - 1 - 7 1 5 1 2 - 1 4 3.2 Landscape-scale Burned area - 1 1 1 - - 1 1 2 - 1 1 - 1 prediction 3.3 Fire Susceptibility Mapping 2 12 1 3 2 8 16 9 26 - 27 1 2 3 3.4 Landscape controls on re 2 10 1 3 2 19 11 15 40 1 30 1 1 2 4.1 Fire Spread and Growth - - - 13 2 - 4 - 1 1 - 3 - 2 4.2 Burned area and re severity - 7 - 1 1 3 10 7 6 3 - 2 1 5 prediction 5.1 Soil erosion and deposits - - 1 - - - 1 1 - - 1 - - - 5.2 Smoke and particulate levels - 2 - - - 3 3 - 5 2 - - - 2 5.3 Post- re regeneration and - 1 - 1 1 6 1 2 10 - 2 - 1 - ecology 5.4 Socioeconomic e ects - - - - 1 - - - - - - - - - 6.1 Planning and policy - - - 1 1 - - - 2 - - - - 2 6.2 Fuel treatment - - - 1 1 - - - - - - - - 1 6.3 Wild re preparedness and re- - - - 1 2 1 1 - - - 1 1 - 1 sponse 6.4 Social factors - - - - 1 - - - - - - - - - Table 3: Summary of application of ML methods applied to di erent problem domains in wild re science and management. A table of acronyms for the ML methods are given in 1. Note that in some cases a paper may use more than one ML method and/or appear in multiple problem domains. 14 4.1.2 Fire detection Detecting wild res as soon as possible after they have ignited, and therefore while they are still relatively small, is critical to facilitating a quick and e ective response. Traditionally, res have mainly been detected by human observers, by distinguishing smoke in the eld of view directly from a re tower, or from a video feed from a tower, aircraft, or from the ground. All of these methods can be limited by spatial or temporal coverage, human error, the presence of smoke from other res and by hours of daylight. Automated detection of heat signatures or smoke in infra-red or optical images can extend the spatial and temporal coverage of detection, the detection eciency in smoky conditions, and remove bias associated with human observation. The analytical task is a classi cation problem that is quite well suited to ML methods. For example, Arrue et al. [2000] used ANNs for infrared (IR) image processing (in combination with visual imagery, meteorological and geographic data used in a decision function using fuzzy logic), to identify true wild res. Several researchers have similarly employed ANNs for re detection [Al-Rawi et al., 2001, Angayarkkani and Radhakrishnan, 2010, Fernandes et al., 2004a,b, Li et al., 2015, Soliman et al., 2010, Utkin et al., 2002, Sayad et al., 2019]. In addition, Liu et al. [2015] used ANNs on wireless sensor networks to build a re detection system, where multi-criteria detection was used on multiple attributes (e.g. ame, heat, light, and radiation) to detect and raise alarms. Other ML methods used in re detection systems include SVM to automatically detect wild res from videoframes [Zhao et al., 2011], GA for multi-objective optimization of a LiDAR-based re detection system [Cordoba et al., 2004], BN in a vision-based early re detection system [Ko et al., 2010], ANFIS [Angayarkkani and Radhakrishnan, 2011, Wang et al., 2011], and KM [Srinivasa et al., 2008]. CNNs (ie. deep learning), which are able to extract features and patterns from spatial images and are nding widespread use in object detection tasks, have recently been applied to the problem of re detection. Several of these applications trained the models on terrestrial based images of re and/or smoke [Zhang et al., 2016, 2018a,b, Yuan et al., 2018, Akhlou et al., 2018, Barmpoutis et al., 2019, Jakubowski et al., 2019, Jo~ ao Sousa et al., 2019, Li et al., 2018b, 2019, Muhammad et al., 2018, Wang et al., 2019]. Of particular note, Zhang et al. [2018b] found CNNs outperformed a SVM-based method and Barmpoutis et al. [2019] found a Faster region-based CNN outperformed another CNN based on YOLO (\you only look once"). Yuan et al. [2018] used CNN combined with optical ow to include time-dependent information. Li et al. [2018b] similarly used a 3D CNN to incorporate both spatial and temporal information and so were able to treat smoke detection as a segmentation problem for video images. Another approach by Cao et al. [2019] used convolutional layers as part of a Long Short Term Memory (LSTM) Neural network for smoke detection from a sequence of images (ie. video feed). They found the LSTM method achieved 97.8% accuracy, a 4.4% improvement over a single image-based deep learning method. Perhaps of greater utility for re management were re/smoke detection models trained on either unmanned aerial vehicle (UAV) images [Zhao et al., 2018, Alexandrov et al., 2019] or satellite imagery including GOES-16 [Phan and Nguyen, 2019] and MODIS [Ba et al., 2019]. Zhao et al. [2018] compared SVM, ANN and 3 CNN models and found their 15-layer CNN performed best with an accuracy of 98%. By comparison, the SVM based method, which was unable to extract spatial features, only had an accuracy of 43%. Alexandrov et al. [2019] found YOLO was both faster and more accurate than a region-based CNN method in contrast to Barmpoutis et al. [2019]. 4.1.3 Fire perimeter and severity mapping Fire maps have two management applications: 1) Accurate maps of the location of the active re perimeter are important for daily planning of suppression activities and/or evacuations, including modeling re growth 2) Maps of the nal burn perimeter and re severity are important for assessing and predicting the economic and ecological impacts of wildland re and for recovery planning. Historically, re perimeters were sketch-mapped from the air, from a ground or aerial GPS or other traverse, or by air-photo interpretation. Developing methods for mapping re perimeters and burn severity from remote sensing imagery has been an area of active research since the advent of remote sensing in the 1970s, and is mainly concerned with 15 classifying active re areas from inactive or non burned areas, burned from unburned areas (for extinguished res), or re severity measures such as the Normalized Burn Ratio [Lutes et al., 2006]. In early studies using ML methods for re mapping Al-Rawi et al. [2001] and Al-Rawi et al. [2002] used ANNs (speci cally, the supervised ART-II neural network) for burned scar mapping and re detection. Pu and Gong [2004] compared Logistic Regression (LR) with ANN for burned scar mapping using Landsat images; both methods achieved high accuracy (> 97%). Interestingly, however, the authors found that LR was more ecient for their relatively limited data set. The authors in Zammit et al. [2006] performed burned area mapping for two large res that occurred in France using satellite images and three ML algorithms, including SVM, K-nearest neighbour, and the K-means algorithm; overall SVM had the best performance. Likewise, E. Dragozi, I. Z. Gitas, D.G. Stavrakoudis [2011] compared the use of SVM against a nearest neighbour method for burned area mapping in Greece and found better performance with SVM. In fact, a number of studies [Alonso-Benito et al., 2008, Cao et al., 2009, Petropoulos et al., 2010, 2011, Zhao et al., 2015, Pereira et al., 2017, Branham et al., 2017, Hamilton et al., 2017] have successfully used SVM for burned scar mapping using satellite data. Mitrakis et al. [2012] performed burned area mapping in the Mediterranean region using a variety of ML algorithms, including a fuzzy neuron classi er (FNC), ANN, SVM, and AdaBoost, and found that, while all methods displayed similar accuracy, the FNC performed slightly better. Dragozi et al. [2014] applied SVM and a feature selection method (based on fuzzy logic) to IKONOS imagery for burned area mapping in Greece. Another approach to burned area mapping in the Mediterranean used an ANN and MODIS hotspot data [G omez and Pilar Mart n, 2011]. Pereira et al. [2017] used a one class SVM, which requires only positive training data (i.e. burned pixels), for burned scar mapping, which may o er a more sample ecient approach than general SVMs { the one class SVM approach may be useful in cases where good wild re training datasets are dicult to obtain. In Mithal et al. [2018], the authors developed a three-stage framework for burned area mapping using MODIS data and ANNs. Crowley et al. [2019] used Bayesian Updating of Landcover (BULC) to merge burned-area classi cations from three remote sensing sources (Landsat-8, Sentinel-2 and MODIS). Celik [2010] used GA for change detection in satellite images, while Sunar and Ozkan [2001] used the interactive Iterative Self-Organizing DATA algorithm (ISODATA) and ANN to map burned areas. In addition to burned area mapping, ML methods have been used for burn severity mapping, including GA [Brumby et al., 2001], MaxEnt [Quintano et al., 2019], bagged decision trees [S a et al., 2003], and others. For instance, Hultquist et al. [2014] used three popular ML approaches (Gaussian Process Regression (GPR) [Rasmussen and Williams, 2006], RF, and SVM) for burn severity assessment in the Big Sur ecoregion, California. RF gave the best overall performance and had lower sensitivity to di erent combinations of variables. All ML methods, however, performed better than conventional multiple regression techniques. Likewise, Hultquist et al. [2014] compared the use of GPR, RF, and SVM for burn severity assessment, and found that RF displayed the best performance. Another recent paper by Collins et al. [2018] investigated the applicability of RF for re severity mapping, and discussed the advantages and limitations of RF for di erent re and land conditions. One recent paper by Langford et al. [2019] used a 5-layer deep neural network (DNN) for mapping res in Interior Alaska with a number of MODIS derived variables (eg. NDVI and surface re ectance). They found that a validation-loss (VL) weight selection strategy for the unbalanced data set (i.e., the no- re class appeared much more frequently than re) allowed them to achieve better accuracy compared with a XGBoost method. However, without the VL approach, XGBoost outperformed the DNN, highlighting the need for methods to deal with unbalanced datasets in re mapping. 4.2 Fire Weather and Climate Change 4.2.1 Fire weather prediction Fire weather is a critical factor in determining whether a re will start, how fast it will spread, and where it will spread. Fire weather observations are commonly obtained from surface weather station networks operated by meteorological services or re management agencies. Weather observations may be interpolated 16 from these point locations to a grid over the domain of interest, which may include diverse topographical conditions; the interpolation task is a regression problem. Weather observations may subsequently be used in the calculation of meteorologically based re danger indices, such as the Canadian Fire Weather Index (FWI) System [Van Wagner, 1987]. Future re weather conditions and danger indices are commonly forecast using the output from numerical weather prediction (NWP) models (e.g., The European Forest Fire Information System [San-Miguel-Ayanz et al., 2012]). However, errors in the calculation of re danger indices that have a memory (such as the moisture indices of the FWI System) can accumulate in such projections. It is noteworthy that surface re danger measures may be correlated with large scale weather and climatic patterns. To date there has been relatively few papers that address re weather and danger prediction using ma- chine learning. The rst e ort [Crimmins, 2006] used self-organizing maps (SOMs) to explore the synoptic climatology of extreme re weather in the southwest USA. He found three key patterns representing south- westerly ow and large geopotential height gradients that were associated with over 80% of the extreme re weather days as determined by a re weather index. Nauslar et al. [2019] used SOMs to determine the timing of the North American Monsoon that plays a major role on the length of the active re season in the southwest USA. Lagerquist et al. [2017] also used SOMs to predict extreme re weather in northern Alberta, Canada. Extreme re weather was de ned by using extreme values of the Fine Fuel Moisture Code (FFMC), Initial Spread Index (ISI) and the Fire Weather Index (FWI), all components of the Cana- dian Fire Weather Index (FWI) System [Van Wagner, 1987]. Good performance was achieved with the FFMC and the ISI and this approach has the potential to be used in near real time, allowing input into re management decision systems. Other e orts have used a combination of conventional and machine learning approaches to interpolate meteorological re danger in Australia [Sanabria et al., 2013]. 4.2.2 Lightning prediction Lightning is second most common cause of wild res (behind human causes); thus predicting the location and timing of future storms/strikes is of great importance to predicting re occurrence. Electronic lightning detection systems have been deployed in many parts of the world for several decades and have accrued rich strike location/time datasets. Lightning prediction models have employed these data to derive regression relationships with atmospheric conditions and stability indices that can be forecast with NWP. Ensemble forecasts of lightning using RF is a viable modelling approach for Alberta, Canada [Blouin et al., 2016]. Bates et al. [2017] used two machine learning methods (CART and RF) and three statistical methods to classify wet and dry thunderstorms (lightning associated with dry thunderstorms are more likely to start res) in Australia. 4.2.3 Climate Change Transfer modeling, whereby a model produced for one study region and/or distribution of environmental conditions is applied to other cases [Phillips et al., 2006], is a common approach in climate change science. Model transferability should be considered when using ML methods to estimated projected quantities due to climate change or other environmental changes. With regards to climate change, transfer modeling is essentially an extrapolation task. Previous studies in the context of species distribution modeling have indicated ML approaches may be suitable for transfer modeling under future climate scenarios. For exam- ple, Heikkinen et al. [2012] indicated MaxEnt and generalized boosting methods (GBM) have the better transferability than either ANN and RF, and that the relatively poor transferability of RF may be due to over tting. There are several publications on wild res and climate change that use ML approaches. Amatulli et al. [2013] found that Multivariate Adaptive Regression Splines (MARS) were better predictors of future monthly area burned for 5 European countries as compared to Multiple Linear Regression and RF. [Parks et al., 2016] projected re severity for future time periods in Western USA using BRT. Young et al. [2017] similarly used BRT to project future re intervals in Alaska and found up to a fourfold increase in (30 17 year) re occurrence probability by 2100. Several authors used MaxEnt to project future re probability globally [Moritz et al., 2012], for Mediterranean ecosystems [Batllori et al., 2013], in Southwest China [Li et al., 2017], the paci c northwestern USA [Davis et al., 2017], and for south central USA [Stroh et al., 2018]. An alternative approach for projecting future potential burn probability was employed by Stralberg et al. [2018] who used RF to determine future vegetation distributions as inputs to ensemble Burn-P3 simulations. Another interesting paper of note was by Boulanger et al. [2018] who built a consensus model with 2 di erent predictor datasets and 5 di erent regression methods (generalised linear models, RF, BRT, CART and MARS) to make projections of future area burned in Canada. The consensus model can be used to quantify uncertainty in future area burned estimates. The authors noted that model uncertainty for future periods (> 200%) can be higher than that of di erent climate models under di erent carbon forcing scenarios. This highlights the need for further work in the application of ML methods for projecting future re danger under climate change. 4.3 Fire Occurrence, Susceptibility and Risk Papers in this domain include prediction of re occurrence and area burned (at a landscape or seasonal scales), mapping of re susceptibility (or similar de nitions of risk) and analysis of landscape or environ- mental controls on re. 4.3.1 Fire occurrence prediction Predictions of the number and location of re starts in the upcoming day(s) are important to preparedness planning | that is, the acquisition of resources, including the relocation of mobile resources and readiness for expected re activity. The origins of re occurrence prediction (FOP) models go back almost 100 years [Nadeem et al., 2020]. FOP models typically use regression methods to relate the response variable ( re reports or hotspots) to weather, lightning, and other covariates for a geographic unit, or as a spatial probability. The seminal work of Brillinger and others in developing the spatio-temporal FOP framework is reviewed in Taylor et al. [2013] The most commonly used ML method in studies predicting re occurrence were ANNs. As early as 1996, Vega-Garcia et al. [1996] used an ANN for human-caused wild re prediction in Alberta, Canada, correctly predicting 85% of no- re observations and 78% of re observations. Not long after, Alonso-Betanzos et al. [2002] and Alonso-Betanzos et al. [2003] used ANN to predict a daily re occurrence risk index using temperature, humidity, rainfall, and re history, as part of a larger system for real-time wild re management system in the Galicia region of Spain. Vasilakos et al. [2007] used separate ANNs for three di erent indices representing re weather (Fire Weather Index; FWI), hazard (Fire Hazard Index; FHI), and risk (Fire Risk Index) to create a composite re ignition index (FII) for estimating the probability of wild re occurrence on the Greek island of Lesvos. Sakr et al. [2010] used meteorological variables in a SVM to create a daily re risk index corresponding to the number of res that could potentially occur on a particular day. Sakr et al. [2011] then compared the use of SVM and ANN for re occurrence prediction based only on relative humidity and cumulative precipitation up to the speci c day. While Sakr et al. [2011] reported low errors for the number of res predicted by both the SVM and ANN models, ANN models outperformed SVM; however, the SVM performed better on binary classi cation of re/no re. It is important to note, however, that ANNs encompass a wide range of possible network architectures. In an Australian study, Dutta et al. [2013] compared the use of ten di erent types of ANN models for estimating monthly re occurrence from climate data, and found that an Elman RNN performed the best. After 2012, RF became the more popular method for predicting re occurrence among the papers reviewed here. Stojanova et al. [2012] evaluated several machine learning methods for predicting re outbreaks using geographical, remote sensed, and meteorological data in Slovenia, including single classi er methods (i.e., KNN, Naive Bayes, DT (using the J48 and jRIP algorithms), LR, SVM, and BN), and ensemble methods (AdaBoost, DT with bagging, and RF). The ensemble methods DT with bagging and RF displayed the best predictive performance with bagging having higher precision and RF having better 18 recall. Vec n-Arias et al. [2016] found that RF performed slightly better than LR for predicting lightning re occurrence in the Iberian Peninsula, based on topography, vegetation, meteorology, and lightning characteristics. Similarly, Cao et al. [2017] found that a cost-sensitive RF analysis outperformed GLM and ANN models for predicting wild re ignition susceptibility. In recent non-comparative studies, Yu et al. [2017] used RF to predict re risk ratings in Cambodia using publicly available remote sensed products, while Van Beusekom et al. [2018] used RF to predict re occurrence in Puerto Rico and found precipitation was found to be the most important predictor. The maximum entropy (MaxEnt) method has also been used for re occurrence prediction [De Angelis et al., 2015, Chen et al., 2015]. For example, De Angelis et al. [2015] used MaxEnt to evaluate di erent meteorological variables and re-indices (e.g. the Canadian Fire Weather Index, FWI) for daily re risk forecasting in the mountainous Canton Ticino region of Switzerland. The authors of that study found that combinations of such variables increased predictive power for identifying daily meteorological conditions for wild res. Dutta et al. [2016] use a two- stage machine learning approach (ensemble of unsupervised deep belief neural networks with conventional supervised ensemble machine learning) to predict bush- re hot spot incidence on a weekly time-scale. In the rst unsupervised deep learning phase, Dutta et al. [2016] used Deep Belief Networks (DBNet; an ensemble deep learning method) to generate simple features from environmental and climatic surfaces. In the second supervised ensemble classi cation stage, features extracted from the rst stage were fed as training inputs to ten ML classi ers (i.e., conventional supervised Binary Tree, Linear Discriminant Analyser, Na ve Bayes, KNN, Bagging Tree, AdaBoost, Gentle Boosting Tree, Random Under-Sampling Boosting Tree, Subspace Discriminant, and Subspace KNN) to establish the best classi er for bush re hotspot estimation. The authors found that bagging and the conventional KNN classi er were the two best classi ers with 94.5% and 91.8% accuracy, respectively. 4.3.2 Landscape scale burned area prediction The use of ML methods in studies of burned area prediction have only occurred relatively recently compared to other wild re domains, yet such studies have incorporated a variety of ML methods. For example, Cheng and Wang [2008] used an RNN to forecast annual average area burned in Canada, while Archibald et al. [2009] used RF to evaluate the relative importance of human and climatic drivers of burnt area in Southern Africa. Arnold et al. [2014] used Hard Competitive Learning (HCL) to identify clusters of unique pre- re antecedent climate conditions in the interior western US which they then used to construct re danger models based on MaxEnt. Mayr et al. [2018] evaluated ve common statistical and ML methods for predicting burned area and re occurrence in Namibia, including GLM, Multivariate Adaptive Regression Splines (MARS), Regres- sion Trees from Recursive Partitioning (RPART), RF, and SVMs for Regression (SVR). The RF model performed best for predicting burned area and re occurrence; however, adjusted R values were slightly higher for RPART and SVR in both cases. Likewise, de Bem et al. [2018] compared the use of LR and ANN for modelling burned area in Brazil. Both LR and ANN showed similar performance; however, the ANN had better accuracy values when identifying non-burned areas, but displayed lower accuracy when classifying burned areas. 4.3.3 Fire Susceptibility Mapping A considerable number of references (71) used various ML algorithms to map wild re susceptibility, cor- responding to either the spatial probability or density of re occurrence (or other measures of re risk such as burn severity) although other terms such as re vulnerability and risk have also been used. The general approach was to build a spatial re susceptibility model using either remote sensed or agency reported re data with some combination of landscape, climate, structural and anthropogenic variables as explanatory variables. In general, the various modeling approaches used either a presence only framework (e.g., MaxEnt) or a presence/absence framework (e.g., BRT or RF). 19 Early attempts at re susceptibility mapping used CART [Amatulli et al., 2006, Amatulli and Camia, 2007, Lozano et al., 2008]. Amatulli and Camia [2007] compared re density maps in central Italy using CART and multivariate adaptive regression splines (MARS) and found while CART was more accurate that MARS led to smoother density model. More recent work has used ensemble based classi ers, such as RF and BRT, or ANNs (see table S.3.3 in supplementary material for a full list) Several of these papers also compared ML and non-ML methods for re susceptibility mapping and in general found superior performance from the ML methods. Speci cally, Adab [2017] mapped re hazard in the Northeast of Iran, and found ANN performed better than binary logistic regression (BLR) with an AUC of 87% compared with 81% for BLR. Bisquert et al. [2012] found ANN outperformed logistic regression for mapping re risk in the North-west of Spain. Goldarag et al. [2016] also compared ANN and linear regression for re susceptibility mapping in Northern Iran and found ANN had much better accuracy (93.49%) than linear regression (65.76%). Guo et al. [2016b] and Guo et al. [2016a] compared RF and logistic regression for re susceptibility mapping in China and found RF led to better performance. Oliveira et al. [2012] compared RF and LR for re density mapping in Mediterranean Europe and found RF outperformed linear regression. De Vasconcelos et al. [2001] found ANN had better classi cation accuracy than logistic regression for ignition probability maps in parts of Portugal. Referring to table 3 and section S.3.3 of the supplementary material a frequently used ML method for re susceptibility mapping was Maximum Entropy (MaxEnt) which is extensively used in landscape ecology for species distribution modeling [Elith et al., 2011]. In particular, Vilar et al. [2016] found MaxEnt performed better than GLM for re susceptibility mapping in central Spain with respect to sensitivity (i.e., true positive rate) and commission error (i.e., false positive rate), even though the AUC was lower. Of further note, Duane et al. [2015] partitioned their re data into topography-driven, wind-driven and convection-driven res in Catalonia and mapped the re susceptibility for each re type. Other ML methods used for regional re susceptibility mapping include Bayesian networks [Bashari et al., 2016, Dlamini, 2011] and novel hybrid methods such as Neuro-Fuzzy systems [Jaafari et al., 2019, Tien Bui et al., 2017]. Bashari et al. [2016] noted that Bayesian networks may be useful because it allows probabilities to be updated when new observations become available. SVM was also used by a number of authors as a benchmark for other ML methods [Ghorbanzadeh et al., 2019b, Gigovi c et al., 2019, Hong et al., 2018, Jaafari, 2019, Ngoc Thach et al., 2018, Rodrigues and De la Riva, 2014, Sachdeva et al., 2018, Tehrany et al., 2018, Tien Bui et al., 2017, van Breugel et al., 2016, Zhang et al., 2019] but as we discuss below, it did not perform as well as other methods to which it was being compared. There were two applications of ML for mapping global re susceptibility including Moritz et al. [2012] who used MaxEnt and Luo et al. [2013] who used RF. Both of these papers found that at a global scale, precipitation was one of the most important predictors of re risk. The majority of papers considered thus far used the entire study period (typically 4 or more years) to map re susceptibility, therefore neglecting the temporal aspect of re risk. However, a few authors have considered various temporal factors to map re susceptibility. Mart n et al. [2019] included seasonality and holidays as explanatory variables for re probability in northeast Spain. Vacchiano et al. [2018] predicted re susceptibility separately for the winter and summer seasons. Several papers produced maps of re susceptibility in the Eastern US by month of year [Peters et al., 2013, Peters and Iverson, 2017]. Parisien et al. [2014] examined di erences in annual re susceptibility maps and a 31 year climatology for the USA, highlighting the role of climate variability as a driver of re occurrence. In particular, they found FWI90 (the 90th percentile of the Canadian Fire Weather Index) was the dominant factor for annual re risk but not for climatological re risk. Cao et al. [2017] considered a 10 day resolution (corresponding to the available re data) for re risk mapping, which makes their approach similar to re occurrence prediction. In addition to re susceptibility mapping, a few papers focused on other aspects of re risk including mapping probability of burn severity classes [Holden et al., 2009, Parks et al., 2018, Tracy et al., 2018]. Parks et al. [2018] additionally considered the role of fuel treatments on re probability which has obvious implications for re management. Additionally Ghorbanzadeh et al. [2019a] combined re susceptibility maps with vulnerability and infrastructure indicators to produce a re hazard map. 20 A number of papers directly compared three or more ML (and sometimes non-ML) methods for re susceptibility mapping. Here we highlight some of these papers, which elucidate the performance and advantages/disadvantages of various ML methods. Cao et al. [2017] found a cost-sensitive RF model outperformed a standard RF model, ANN as well as probit and logistic regression. Ghorbanzadeh et al. [2019b] compared ANN, SVM and RF and found the best performance with RF. Gigovi c et al. [2019] compared SVM and RF for re susceptibility mapping in combination with Bayesian averaging to generate ensemble models. They found the ensemble model led to marginal improvement (AUC = 0.848) over SVM (AUC=0.834) and RF (AUC=0.844). For mapping both wild re ignitions and potential natural vegetation in Ethiopia van Breugel et al. [2016] also considered ensemble models consisting of a weighted combination of ML methods (RF, SVM, BRT, MaxEnt, ANN, CART) and non-ML methods (GLM and MARS) and concluded the ensemble member performed best over a number of metrics. However, in this paper RF showed the best overall performance of all methods including the ensemble model. Jaafari et al. [2018] compared 5 decision tree based classi ers for wild re susceptibility mapping in Iran. Here, the Alternating Decision tree (ADT) classi er achieved the highest performance (accuracy 94.3%) in both training and validation sets. Ngoc Thach et al. [2018] compared SVM, RF and a Multilayer Perceptron (MLP) neural network for forest re danger mapping in the region of Tjuan chau in Vietnam. They found the performance of all models were comparable although MLP had the highest AUC values. Interestingly Pourtaghi et al. [2016] found that a generalized additive model (GAM) outperformed RF and BRT for re susceptibility mapping in the Golestan province in Iran. This was one of the few examples we found where a non-ML method outperformed ML methods. Rodrigues and De la Riva [2014] compared RF, BRT, SVM and logistic regression for re susceptibility mapping and found RF led to the highest accuracy as well as the most parsimonious model. Tehrany et al. [2018] compared a LogitBoost ensemble-based decision tree (LEDT) algorithm with SVM, RF and Kernel logistic regression (KLR) for re susceptibility mapping in Lao Cai region of Vietnam and found the best performance with LEDT, closely followed by RF. Finally, of particular note, Zhang et al. [2019] compared CNN, RF, SVM, ANN and KLR for re susceptibility mapping in the Yunnan Province of China. This was the only application of deep learning we could nd for re susceptibility mapping. The authors found that CNN outperformed the other algorithms with overall accuracy of 87.92% compared with RF (84.36%), SVM (80.04%), MLP (78.47%), KLR (81.23%). They noted that the bene t of CNN is that it incorporates spatial correlations so that it can learn spatial features. However, the downside is that deep learning models are not as easily interpretted as other ML methods (such as RF and BRT). 4.3.4 Landscape controls on re Many of the ML methods used in re susceptibility mapping have also been used to examine landscape controls { ie. the relative importance of weather, vegetation, topography, structural and anthropogenic variables { on re activity, which may facilitate hypothesis formation and testing or model building. From table 3 the most commonly used methods in this section were MaxEnt, RF, BRT and ANN. These methods all allow for the determination of variable importance (i.e. the relative in uence of predictor variables in a given model of a response variable). A commonly used method to ascertain variable importance is through the use of partial dependence plots [Hastie et al., 2009]. This method works by averaging over models that exclude the predictor variable of interest, with the resulting reduction in AUC (or other performance metrics) representing the marginal e ect of the variable on the response. Partial dependence plots have the advantage of being able to be applied to a wide range of ML methods. A related method for determining variable importance, often used for RFs, is a permutation test which involves random permutation of each predictor variable [Strobl et al., 2007]. Another model-dependent approach used for ANN is the use of partial derivatives (of the activation functions of hidden and output nodes) as outlined by Vasilakos et al. [2009]. It should be noted that while many other methods for model interpretation and variable dependence exist, a discussion of these methods is outside the scope of this paper. In general, the drivers of re occurrence or area burned varied greatly by the study area considered (including the size of area) and the methods used. Consistent with other work on \top down" and \bottom 21 up" drivers of re activity, at large scales climate variables were often determined to be the main drivers of re activity whereas at smaller scales anthropogenic or structural factors exerted a larger in uence. Here we discuss some of the papers that highlight the diversity of results for di erent study areas and spatial scales (global, country, ecoregion, urban) but refer the reader to section S.3.4 of the supplementary material for a full listing of papers in this section. Note that many of the papers listed under section S.3.4 also belong to the re susceptibility mapping section and have already been discussed there. Aldersley et al. [2011] considered drivers of monthly area burned at global and regional scales using both regression trees and RF. They found climate factors (high temperature, moderate precipitation, and dry spells) were the most important drivers at the global scale, although at the regional scale the models exhibited higher variability due to the in uence of anthropogenic factors. At a continental scale Mansuy et al. [2019] used MaxEnt to show that climate variables were the dominant controls (over landscape and human factors) on area burned for most ecoregions for both protected areas and outside these areas, although anthropogenic factors exerted a stronger in uence in some regions such as the Tropical Wet Forests ecoregion. [Masrur et al., 2018] used RF to investigate controls on circumpolar arctic re and found June surface temperature anomalies were the most important variable for determining the likelihood of wild re occurrence on an annual scale. Chingono and Mbohwa [2015] used MaxEnt to model re occurrences in Southern Africa where most res are human-caused and found vegetation (i.e., dry mass productivity and NDVI) were the main drivers of biomass burning. Curt et al. [2015] used BRT to examine drivers of re in New Caledonia. Interestingly, they found that human factors (such as distance to villages, cities or roads) were dominant in uences for predicting re ignitions whereas vegetation and weather factors were most important for area burned. Curt et al. [2016] modeled re probabilities by di erent re ignition causes (lightning, intentional, accidental, negligence professional and negligence personal) in Southeastern France. They found socioeconomic factors (eg. housing and road density) were the dominant factors for ignitions and area burned for human-caused res. Fernandes et al. [2016] used BRT to examine large res in Portugal and found high pyrodiversity (ie. spatial structure due to re recurrence) and low landscape fuel connectivity were important drivers of area burned. Curt et al. [2016] modeled re probabilities by di erent re ignition causes (lightning, intentional, accidental, negligence professional and negligence personal) in Southeastern France. They found socioeconomic factors (eg. housing and road density) were the dominant factors for ignitions and area burned for human-caused res. Leys et al. [2017] used RF to nd the drivers that determine sedimentary charcoal counts in order to reconstruct grass re history in the Great Plains, USA. Not surprisingly, they found re regime characteristics (eg. area burned and re frequency) were the most important variables and concluded that charcoal records can therefore be used to reconstruct re histories. Li et al. [2009] used ANNs to show that wild re probability was strongly in uenced by population density in Japan, with a peak determined by the interplay of positive and negative e ects of human presence. This relationship, however, becomes more complex when weather parameters and forest cover percentage are added to the model. Liu et al. [2013] used BRT to study factors in uencing re size in the Great Xingan Mountains in Northeastern China. Their method included a \moving window" resampling technique that allowed them to look at the relative in uence of variables at di erent spatial scales. They showed that the most dominant factors in uencing re size were fuel and topography for small res, but re weather became the dominant factor for larger res. For regions of high population density, anthropogenic or structural factors are often dominant for re susceptibility. For example Molina et al. [2019] used MaxEnt to show distance to roads, settlements or powerlines were the dominant factors for re occurrence probability in the Andalusia region in southern Spain. MaxEnt has also been used for estimating spatial re probability under di erent scenarios such as future projections of housing development and private land conservation [Syphard et al., 2016]. One study in China using RF found mean spring temperature was the most important variable for re occurrence whereas forest stock was most important for area burned [Ying et al., 2018]. Some authors examined controls on re severity using high resolution data for a single large re. For example, several authors used RF to examine controls on burn severity for the 2013 Rim re in the Sierra Nevada [Lydersen et al., 2014, Kane et al., 2015, Lydersen et al., 2017]. At smaller spatial scales re 22 weather was the most important variable for re severity, whereas fuel treatments were most important at larger spatial scales [Lydersen et al., 2017]. A similar study by Harris and Taylor [2017] showed that previous re severity was an important factor in uencing re severity for the Rim re. For the 2005 Riba de Saelices re, Viedma et al. [2015] looked at factors contributing to burn severity using a BRT model and found burning conditions (including re weather variables) were more important compared than stand structure and topography. For burn severity these papers all used the Relativized di erenced Normalized Burn Ratio (RdNBR) metric, derived from Landsat satellite images, which allowed spatial modeling at high resolutions (eg. 30m by 30m). In addition to the more commonly used ML methods one paper by Wu et al. [2015] used KNN to identify spatially homogeneous re environment zones by clustering climate, vegetation, topography, and human activity related variables. They then used CART to examine variable importance for each of three re environment zones in south-eastern China. For landscape controls on re there were few studies comparing multiple ML methods. One such study by Nelson et al. [2017] compared CART, BRT and RF for classifying di erent re size classes in British Columbia, Canada. For both central and periphery regions they found the best performing model was BRT followed by CART and RF. For example, in the central region BRT achieved a classi cation accuracy of 88% compared with 82.9% and 49.6% for the CART and RF models respectively. It is not clear from the study why RF performed poorly, although it was noted that variable importance di ers appreciably between the three models. 4.4 Fire Behavior Prediction In general, re behavior includes physical processes and characteristics at a variety of scales including combustion rate, aming, smouldering residence time fuel consumption, ame height, and ame depth. However, the papers in this section deal mainly with larger scale processes and characteristics such as the prediction of re spread rates, re growth, burned area, and re severity, conditional on the occurrence (ignition) of one, or more, wild res. Here, our emphasis is on prognostic applications, in contrast to the Fuels Characterization, Fire Detection and Mapping problem domain, in which we focused on diagnostic applications. 4.4.1 Fire spread and growth Predicting the spread of a wildland re is an important task for re management agencies, particularly to aid in the deployment of suppression resources or to anticipate evacuations one or more days in advance. Thus, a large number of models have been developed using di erent approaches. In a series of reviews Sullivan [2009a,b,c] described re spread models he classi ed as being of physical or quasi-physical nature, or empirical or quasi-empirical nature, as well as mathematical analogues and simulation models. Many re growth simulation models convert one dimensional empirical or quasi-empirical spread rate models to two dimensions and then propagate a re perimeter across a modelled landscape. A wide range of ML methods have been applied to predict re growth. For example, Markuzon and Kolitz [2009] tested several classi ers (RF, BNs, and KNN) to estimate if a re would become large either one or two days following its observation; they found each of the tested methods performed similarly with RF correctly classifying large res at a rate over 75%, albeit with a number of false positives. Vakalis et al. [2004] used a ANN in combination with a fuzzy logic model to estimate the rate of spread in the mountainous region of Attica in Greece. A number of papers used genetic algorithms (GAs) to optimize input parameters to a physics or empirically based re simulator in order to improve re spread predictions [Abdalhaq et al., 2005, Rodriguez et al., 2008, Rodr guez et al., 2009, Art es et al., 2014, 2016, Carrillo et al., 2016, Denham et al., 2012, Cencerrado et al., 2012, 2013, 2014, Art es et al., 2017, Denham and Laneri, 2018]. For example, Cencerrado et al. [2014] developed a framework based on GAs to shorten the time needed to run deterministic re spread simulations. They tested the framework using the FARSITE [Finney, 2004] re spread simulator with di erent input scenarios sampled from distributions of vegetation models, wind speed/direction, and dead/live fuel moisture content. The algorithm used a tness function which discarded the most time-intensive simulations, but did not lead to an appreciable decrease in the 23 accuracy of the simulations. Such an approach is potentially useful for re management where it is desirable to predict re behavior as far in advance as possible so that the information can be enacted upon. This approach may greatly reduce overall simulation time by reducing the input parameter space as also noted by Art es et al. [2016] and Denham et al. [2012], or through parallelization of simulation runs for stochastic approaches [Art es et al., 2017, Denham and Laneri, 2018]. A di erent goal was considered by Ascoli et al. [2015] who used a GA to optimize fuel models in Southern Europe by calibrating the model with respect to rate of spread observations. Kozik et al. [2013] presented a re spread model that used a novel ANN implementation that incorpo- rated a Kalman lter for data assimilation that could potentially be run in real-time, the resulting model more closely resembling that of complex cellular automata than a traditional ANN. The same authors later implemented this model and simulated re growth under various scenarios with di erent wind speeds and directions, or both, although a direct comparison with real re data was not possible [Kozik et al., 2014]. Zheng et al. [2017] simulated re spread by integrating a cellular automata (CA) model with an Extreme Learning Machine (ELM; a type of feedforward ANN). Transition rules for the CA were determined by the ELM trained with data from historical res, as well as vegetation, topographic, and meteorological data. Likewise, Chetehouna et al. [2015] used ANNs to predict re behavior, including rate of spread, and ame height and angle. In contrast, Subramanian and Crowley [2017] formulated the problem of re spread prediction as a Markov Decision Process, where they proposed solutions based on both a classic reinforcement learning algorithm and a deep reinforcement learning algorithm { the authors found the deep learning approach improved on the traditional approach when tested on two large res in Alberta, Canada. The authors further developed this work to compare ve widely used reinforcement learning algorithms [Subramanian and Crowley, 2018], and found that the Asynchronous Advantage Actor-Critic (A3C) and Monte Carlo Tree Search (MCTS) algorithms achieved the best accuracy. Meanwhile, Khakzad [2019] developed a re spread model to predict the risk of re spread in Wildland-Industrial Interfaces, using Dynamic Bayesian Networks (DBN) in combination with a deterministic re spread model. The Canadian Fire Behavior Prediction (FBP) system, which uses meteorological and fuel conditions data as inputs, determined the re spread probabilities from one node to another in the aforementioned DBN. More recently Hodges and Lattimer [2019] trained a (deep learning) CNN to predict re spread using environmental variables (topography, weather and fuel related variables). Outputs of the CNN were spatial grids corresponding to the probability the burn map reached a pixel and the probability the burn map did not reach a pixel. Their method achieved a mean precision of 89% and mean sensitivity of 80% with reference 6 hourly burn maps computed using the physics-based FARSITE simulator. Radke et al. [2019] also used a similar approach to predict daily re spread for the 2016 Beaver Creek re in Colorado. 4.4.2 Burned area and re severity prediction There are a number of papers that focus on using ML approaches to directly predict the nal area burned from a wild re. Cortez and Morais [2007] compared multiple regression and four di erent ML methods (DT, RF, ANN, and SVM) to predict area burned using re and weather (i.e., temperature, precipitation, relative humidity and wind speed) data from the Montesinho natural park in northeastern Portugal, and found that SVM displayed the best performance. A number of publications subsequently used the data from Cortez and Morais [2007] to predict area burned using various ML methods, including ANN [Sa and Bouroumi, 2013, Storer and Green, 2016], genetic algorithms [Castelli et al., 2015], both ANN and SVM [Al Janabi et al., 2018], and decision trees [Alberg, 2015, Li et al., 2018a]. Notably, Castelli et al. [2015] found that a GA variant outperformed other ML methods including SVM. Xie and Shi [2014] used a similar set of input variables with SVM to predict burned area in for Guangzhou City in China. In addition to these studies, Toujani et al. [2018] used hidden Markov models (HMM) to predict burned area in the north- west of Tunisia, where the spatiotemporal factors used as inputs to the model were initially clustered using self-organizing maps (SOMs). Liang et al. [2019] compared back-propagation neural networks, recurrent neural networks (RNN) and Long Short Term Memory (LSTM) neural networks to predict wild re scale, a quantity related to area burned and re duration, in Alberta Canada. They found the highest accuracy 24 (90.9%) was achieved with LSTM. Most recently, Xie and Peng [2019] compared a number of machine learning methods for estimating area burned (regression) and binary classi cation of re sizes (> 5 Ha) in Montesinho natural park, Portugal. For the regression task, they found a tuned RF algorithm performed better than standard RF, tuned and standard gradient boosted machines, tuned and standard generalized linear models (GLMs) and deep learning. For the classi cation problem they found extreme gradient boosting and deep learning had a higher accuracy than CART, RF, SVM, ANN, and logistic regression. By attempting to predict membership of burned area size classes, a number of papers were able to recast the problem of burned area prediction as a classi cation problem. For example, Yu et al. [2011] used a combination of SOMs and back-propagation ANNs to classify forest res into size categories based on meteorological variables. This approach gave Yu et al. [2011] better accurary ( 90%) when compared with a rules-based method ( 82%). Ozbayo glu and Bozer [2012] estimated burned area size classes us- ing geographical and meteorological data using three di erent machine learning methods: i) Multilayer Perceptron (MLP); ii) Radial Basis Function Networks (RBFN); and iii) SVM. Overall, the best perform- ing method was MLP, which achieved a 65% success rate, using humidity and windspeed as predictors. Zwirglmaier et al. [2013] used a BN to predict area burned classes using historical re data, re weather data, re behaviour indices, land cover, and topographic data. Shidik and Mustofa [2014] used a hybrid model (Fuzzy C-Means and Back-Propagation ANN) to estimate re size classes using data from Cortez and Morais [2007], where the hybrid model performed best with an accuracy of 97.50% when compared with Naive Bayes (55.5%), DT (86.5%), RF (73.1%), KNN (85.5%) and SVM (90.3%). Mitsopoulos and Mallinis [2017] compared BRT, RF and Logistic Regression to predict 3 burned area classes for res in Greece. They found RF led to the best performance of the three tested methods and that re suppression and weather were the two most important explanatory variables. Coeld et al. [2019] compared CART, RF, ANN, KNN and gradient boosting to predict 3 burned area classes at time of ignition in Alaska. They found a parsimonious model using CART with Vapor Pressure De cit (VPD) provided the best performance of the models and variables considered. We found only one study that used ML to predict re behavior related to re severity, which is important in the context of re ecology, suggesting that there are opportunities to apply ML in this domain of wild re science. In that paper, Zald and Dunn [2018] used RF to determine that the most important predictor of re severity was daily re weather, followed by stand age and ownership, with less predictability given by topographic features. 4.5 Fire E ects Fire E ects prediction studies have largely used regression based approaches to relate costs, losses, or other impacts (e.g., soils, post- re ecology, wildlife, socioeconomic factors) to physical measures of re severity and exposure. Importantly, this category also includes wild re smoke and particulate modelling (but not smoke detection which was previously discussed in the re detection section). 4.5.1 Soil Erosion and Deposits Mallinis et al. [2009] modelled potential post- re soil erosion risk following a large intensive wild re in the Mediterranean area using CART and k-means algorithms. In that paper, before wild re, 55% of the study area was classi ed as having severe or heavy erosion potential, compared to 90% post- re, with an overall classi cation accuracy of 86%. Meanwhile, Buckland et al. [2019] used ANNs to examine the relationships between sand deposition in semi-arid grasslands and wild re occurrence, land use, and climatic conditions. The authors then predicted soil erosion levels in the future given climate change assumptions. 25 4.5.2 Smoke and Particulate Levels Smoke emitted from wild res can seriously lower air quality with adverse e ects on the health of both human and non-human animals, as well as other impacts. Thus, it is not surprising that ML methods have been used to understand the dynamics of smoke from wildland re. For example, Yao et al. [2018b] used RF to predict the minimum height of forest re smoke using data from the CALIPSO satellite. More commonly, ML methods have also been used to estimate population exposure to ne particulate matter (e.g., PM2.5: atmospheric particulate matter with diameter less than 2.5m), which can be useful for epidemiological studies and for informing public health actions. One such study by Yao et al. [2018a] also used RF to estimate hourly concentrations of PM2.5 in British Columbia, Canada. Zou et al. [2019] compared RF, BRT and MLR to estimate regional PM2.5 concentrations in the Paci c Northwest and found RF performed much better than the other algorithms. In another very broad study covering several datasets and ML methods, Reid et al. [2015] estimated spatial distributions of PM2.5 concentrations during the 2008 northern California wild res. The authors of the aforementioned study used 29 predictor variables and compared 11 di erent statistical models, including RF, BRT, SVM, and KNN. Overall, the BRT and RF models displayed the best performance. Emissions other than particulate matter have also been modelled using ML, as Lozhkin et al. [2016] used an ANN to predict carbon monoxide concentrations emitted from a peat re in Siberia, Russia. In another study, the authors used ten di erent statistical and ML methods and 21 covariates (including weather, geography, land-use, and atmospheric chemistry) to predict ozone exposures before and after wild re events [Watson et al., 2019]. Here, gradient boosting gave the best results with respect to both root mean square error and R values, followed by RF and SVM. In a di erent application related to smoke, Fuentes et al. [2019] used ANNs to detect smoke in several di erent grape varietals used for wine making. 4.5.3 Post- re regeneration, succession, and ecology The study of post- re regeneration is an important aspect of understanding forest and ecosystem responses and resilience to wild re disturbances, with important ecological and economic consequences. RF, for example, has been a popular ML method for understanding the important variables driving post- re regeneration [Jo~ ao et al., 2018, Vijayakumar et al., 2016]. Burn severity (a measure of above and below ground biomass loss due to re) is an important metric for understanding the impacts of wild re on vegetation and post- re regeneration, soils, and potential successional shifts in forest composition, and as such, has been included in many ML studies in this section, including [Barrett et al., 2011, Cai et al., 2013, Cardil et al., 2019, Chapin et al., 2014, Divya and Vijayalakshmi, 2016, Fairman et al., 2017, Han et al., 2015, Johnstone et al., 2010, Liu and Yang, 2014, Mart n-Alc on and Coll, 2016, Sherrill and Romme, 2012, Thompson and Spies, 2010]. For instance, Cardil et al. [2019] used BRT to demonstrate that remotely- sensed data (i.e., Relative Di erenced Normalized Burn Ratio index; RdNBR) can provide an acceptable assessment of re-induced impacts (i.e., burn severity) on forest vegetation, while [Fairman et al., 2017] used RF to identify the variables most important in explaining plot-level mortality and regeneration of Eucalyptus pauci ora in Victoria, Australia, a ected by high-severity wild res and subsequent re-burns. Debouk et al. [2013] assessed post- re vegetation regeneration status using eld measurements, a canopy height model, and Lidar (i.e., 3D laser scanning) data with a simple ANN. Post- re regeneration also has important implications for the successional trajectories of forested areas, and a few studies have examined this using ML approaches [Barrett et al., 2011, Cai et al., 2013, Johnstone et al., 2010]. For example, Barrett et al. [2011] used RF to model re severity, from which they made an assessment of the area susceptible to a shift from coniferous to deciduous forest cover in the Alaskan boreal forest, while Cai et al. [2013] used BRT to assess the in uence of environmental variables and burn severity on the composition and density of post- re tree recruitment, and thus the trajectory of succession, in northeastern China. In other studies not directly related to post- re regeneration, Hermosilla et al. [2015] used RF to attribute annual forest change to one of four categories, including wild re, in Saskatchewan, Canada, while [Jung et al., 2013] used GA and RF to estimate the basal area of post- re residual spruce (Picea obovate ) and r 26 (Abies sibirica ) stands in central Siberia using remotely sensed data. Magadzire et al. [2019] used MaxEnt to demonstrate that re return interval and species life history traits a ected the distribution of plant species in South Africa. ML has also been used to examine re e ects on the hydrological cycle, as Poon et al. [2018] used SVM to estimate both pre- and post-wild re evapotranspiration using remotely sensed variables. Considering the potential impacts of wild res on wildlife, it is perhaps surprising that relatively few of such studies have adopted ML approaches. However, ML methods have been used to predict the impacts of wild re and other drivers on species distributions and arthropod communities. Hradsky et al. [2017], for example, used non-parametric BNs to describe and quantify the drivers of faunal distributions in wild re- a ected landscapes in southeastern Australia. Similarly, Reside et al. [2012] used MaxEnt to model bird species distributions in response to re regime shifts in northern Australia, which is an important aspect of conservation planning in the region. ML has also been used to look at the e ects of wild re on fauna at the community level, as Luo et al. [2017] used DTs, Association Rule Mining, and AdaBoost to examine the e ects of re disturbance on spider communities in Cangshan Mountain, China. 4.5.4 Socioeconomic e ects ML methods have been little used to model socio-economic impacts of re to date. We found one study in which BNs were used to predict the economic impacts of wild res in Greece from 2006-2010 due to housing losses [Papakosta et al., 2017]. The authors did this by rst de ning a causal relationship between the participating variables, and then using BNs to estimate housing damages. It is worth noting that the problem of detecting these causal relationships from data is a dicult task and remains an active area of research in arti cial intelligence. 4.6 Fire management The goal of contemporary re management is to have the appropriate amount of re on the landscape, which may be accomplished through the management of vegetation including prescribed burning, the management of human activities (prevention), and re suppression. Fire management is a form of risk management that seeks to maximize re bene ts and minimize costs and losses [Finney, 2005]. Fire management decisions have a wide range of scales, including long-term strategic decisions about the acquisition and location of resources or the application of vegetation management in large regions, medium-term tactical decisions about the acquisition of additional resources, relocation, or release of resources during the re season, and short-term real time operational decisions about the deployment and utilization of resources on individual incidents. Fire preparedness and response is a supply chain with a hierarchical dependence. Taylor [2020] describes 20 common decision types in re management and maps the spatial-temporal dimensions of their decision spaces. Fire management models can be predictive, such as the probability of initial attack success, or pre- scriptive such as to maximize/minimize an objective function (e.g., optimal helicopter routing to minimize travel time in crew deployment). While advances have been made in the domain of wild re management using ML techniques, there have been relatively few studies in this area compared to other wild re problem domains. Thus, there appears to be great potential for ML to be applied to wild re management problems, which may lead to novel and innovative approaches in the future. 4.6.1 Planning and policy An important area of re management is planning and policy, where various ML methods have been applied to address pertinent challenges. For example, Bao et al. [2015] used GA, which are useful for solving multi-objective optimization problems, to optimize watchtower locations for forest re monitoring. Bradley et al. [2016] used RF to investigate the relationship between the protected status of forest in the western US and burn severity. Likewise, Ru ault and Mouillot [2015] also used BRTs to assess the impact 27 of re policy introduced in the 1980s on re activity in southern France and the relationships between re and weather, and Penman et al. [2011] used BNs to build a framework to simultaneously assess the relative merits of multiple management strategies in Wollemi National Park, NSW, Australia. McGregor et al. [2016] used Markov decision processes (MDP) and model free Monte Carlo method to create fast running simulations (based on the FARSITE simulator) to create interactive visualizations of forest futures over 100 years based on alternate high-level suppression policies. McGregor et al. [2017] demonstrated ways in which a variety of ML and optimization methods can be used to create an interactive approximate simulation tool for re managers. The authors of the aforementioned study utilized a modi ed version of the FARSITE re-spread simulator, which was augmented to run thousands of simulation trajectories while also including new models of lightning strike occurrences, re duration, and a forest vegetation simulator. McGregor et al. [2017] also clearly show how decision trees can be used to analyze a hierarchy of decision thresholds for deciding whether to suppress a re or not; their hierarchy splits on fuel levels, then intensity estimations, and nally weather predictors to arrive at a generalizable policy. 4.6.2 Fuel treatment ML methods have also been used to model the e ects of fuel treatments in order to mitigate wild re risk. For example, Penman et al. [2014] used a BN to examine the relative risk reduction of using prescribed burns on the landscape versus within the 500m interface zone adjacent to houses in the Sydney basin, Australia. Lauer et al. [2017] used approximate dynamic programming (also known as reinforcement learning) to determine the optimal timing and location of fuel treatments and timber harvest for a re-threatened landscape in Oregon, USA, with the objective of maximizing wealth through timber management. Similarly, Arca et al. [2015] used GA for multi-objective optimization of fuel treatments. 4.6.3 Wild re preparedness and response Wild re preparedness and response issues have also been examined using ML techniques. Costafreda- Aumedes et al. [2015] used ANNs to model the relationships between daily re load, re duration, re type, re size, and response time, as well as personnel and terrestrial/aerial units deployed for individual wild res in Spain. Most of the models in Costafreda-Aumedes et al. [2015] highlighted the positive correlation of burned area and re duration with the number of resources assigned to each re, and some highlighted the negative in uence of daily re load. In another study, Penman et al. [2015] used Bayesian Networks to assess the relative in uence of preventative and suppression management strategies on the probability of house loss in the Sydney basin, Australia. O'Connor et al. [2017] used BRT to develop a predictive model of re control locations in the Northern Rocky Mountains, USA, based on the likelihood of nal re perimeters, while Homchaudhuri et al. [2010] used GAs to optimize reline generation. Rodrigues et al. [2019] modelled the probability that wild re will escape initial attack using a RF model trained with re location, detection time, arrival time, weather, fuel types, and available resources data. Important variables in Rodrigues et al. [2019] included re weather and simultaneity of events. Julian and Kochenderfer [2018a] used two di erent RL algorithms to develop a system for autonomous control of one or more aircraft in order to monitor active wild res. 4.6.4 Social factors Recently, the use of ML in re management has grown to encompass more novel aspects of re management, even including the investigation of criminal motives related to arson, as Delgado et al. [2018] used BNs to characterize wild re arsonists in Spain thereby identifying ve motivational archetypes (i.e., slight negligence; gross negligence; impulsive; pro t; and revenge). 28 5 Discussion ML methods have seen a spectacular evolution in development, accuracy, computational eciency, and application in many elds since the 1990s. It is therefore not surprising that ML has been helpful in providing new insights into several critical sustainability and social challenges in the 21st century [Gomes, 2009, Sullivan et al., 2014, Butler, 2017]. The recent uptake and success of ML methods has been driven in large part by ongoing advances in computational power and technology. For example, the recent use of bandwidth optimized Graphics Processing Units (GPUs) takes advantage of parallel processing for simul- taneous execution of computationally expensive tasks, which has facilitated a wider use of computationally demanding but more accurate methods like DNNs. The advantages of powerful but ecient ML methods are therefore widely anticipated as being useful in wild re science and management. However, despite some early papers suggesting that data driven techniques would be useful in forest re management [Latham, 1987, Kourtz, 1990, 1993], our review has shown that there was relatively slow adoption of ML-based research in wild re science up to the 2000s compared with other elds, followed by a sharp increase in publication rate in the last decade. In the early 2000s, data mining techniques were quite popular and classic ML methods such as DTs, RF, and bagging and boosting techniques began to appear in the wild re science literature (e.g., Stojanova et al. [2006]). In fact, some researchers started using simple feed forward ANNs for small scale applications as early as the mid 1990s and early 2000s (e.g., Mccormick et al. [1999], Al-Rawi et al. [2002]). In the last three decades, almost all major ML methods have been used in some way in wild re applications, although some more computationally demanding methods, such as SOMs and cellular automatons, have only been actively experimented with in the last decade [Toujani et al., 2018, Zheng et al., 2017]. Furthermore, the recent development of DL algorithms, with a particular focus on extracting spatial features from images, has led to a sharp rise in the application of DL for wild re applications in the last decade. It is evident, however, from our review that while an increasing number of ML methodologies have been used across a variety of re research domains over the past 30 years, this research is unevenly distributed among ML algorithms, research domains and tasks, and has had limited application in re management. Many re science and management questions can be framed within a re risk context. Xi et al. [2019] discussed the advantages of adopting a risk framework with regard to statistical modeling of wild res. There the risk components of \hazard", \vulnerability" and \exposure" are replaced respectively by re probability, re behavior and re e ects. Most re management activities can be framed as risk controls to mitigate these components of risk. Traditionally, methods used in wild re re science to address these various questions have included physical modeling (e.g., Sullivan [2009a,b,c]), statistical methods (e.g., Taylor et al. [2013], Xi et al. [2019]), simulation modeling (e.g., Keane et al. [2004]), and operations research methods (Martell [2015], Minas et al. [2012]). In simple terms, any analytical study begins with one or more of four questions: \what happened?"; \why did it happen?"; \what will happen?"; or \what to do?" Corresponding data driven approaches to address these questions are respectively called descriptive, diagnostic, predictive, and prescriptive analyt- ics. The type of analytical approach adopted then circumscribes the types of methodological approaches (e.g., regression, classi cation, clustering, dimensionality reduction, decision making) and sets of possible algorithms appropriate to the analysis. In our review, we found that studies incorporating ML methods in wildland re science were predomi- nantly associated with descriptive or diagnostic analytics, re ecting the large body of work on re detection and mapping using classi cation methods, and on re susceptibility mapping and landscape controls on re using regression approaches. In many cases, the ML methods identi ed in our review are an alternative to statistical methods used for clustering and regression. While the aforementioned tasks are undoubtedly very important for understanding wildland re, we found much less work associated with predictive or pre- scriptive analytics, such as re occurrence prediction (predictive), re behaviour prediction (predictive), and re management (prescriptive). This may be because: a) particular domain knowledge is required to frame re management problems; b) re management data are often not publicly available, need a lot of 29 work to transform into an easily analyzable form, or do not exist at the scale of the problem; and c) some re management problems are not suited or can't be fully addressed by ML approaches. We note that much of the work on re risk in the re susceptibility and mapping domain used historical re and environmental data to map re susceptibility; therefore, while that work aims to inform future re risk, it cannot be considered to be predictive analytics, except, for example, in cases where it was used in combination with climate change projections. It appears then that, in general, wild re science research is currently more closely aligned with descriptive and diagnostic analytics, whereas wild re management goals are aligned with predictive and prescriptive analytics. This fundamental di erence identi es new opportunities for research in re management, which we discuss later in this paper. In the remainder of the paper, we examine some considerations for the use of ML methods, including: data considerations, model selection and accuracy, implementation challenges, interpretation, opportuni- ties, and implications for re management. 5.1 Data considerations ML is a data-centric modeling paradigm concerned with nding patterns in data. Importantly, data scientists need to determine, often in collaboration with re managers or domain experts, whether there are suitable and sucient data for a given modeling task. Some of the criteria for suitable data include whether: a) the predictands and covariates are or can be wrangled into the same temporal and spatial scale; b) the observations are a representative sample of the full range of conditions that may occur in application of a model to future observations; and c) whether the data are at spatiotemporal scale appropriate to the re science or management question. The rst of these criteria can be relaxed in some ML models such as ANNs and DNNs, where inputs and outputs can be at di erent spatial or temporal scales for appropriately designed network architectures, although data normalization may still be required. The second criterion also addresses the important question of whether enough data exists for training a given algorithm for a given problem. In general, this question depends on the nature of the problem, complexity of the underlying model, data uncertainty and many other factors (see Roh et al. [2018] for a further discussion of data requirements for ML). In any case, many complex problems require a substantive data wrangling e ort, to acquire, perform quality assurance, and fuse data into sampling units at the appropriate spatiotemporal scale. An example of this in daily re occurrence prediction, where observations of a variety of features (e.g., continuous measures such as re arrival time and location, or lightning strike times and locations) are discretized into three-dimensional (e.g., longitude, latitude, and day) cells called voxels. Another important consideration for the collection and use of data in machine learning is selection bias. A form of spatial selection bias called preferential sampling occurs when sampling occurs preferentially in locations where one expects a certain response [Diggle et al., 2010]. For example, preferential sampling may occur in air monitoring, because sensors may be placed in locations where poor air quality is expected [Shaddick and Zidek, 2014]. In general, preferential sampling or other selection biases may be avoided altogether by selecting an appropriate sampling strategy at the experimental design phase, or, where this is not possible, to take it into account in model evaluation [Zadrozny, 2004]. For the problem domain re detection and mapping, most applications of ML used some form of im- agery (e.g., remote sensed satellite images or terrestrial photographs). In particular, many papers used satellite data (e.g., Landsat, MODIS) to determine vegetation di erences before and after a re and so were able to map area burned. For re detection, many applications considered either remote sensed data for hotspot or smoke detection, or photographs of wild res (used as inputs to an image classi cation problem). For re weather and climate change, the three main sources of data were either weather station observa- tions, climate reanalyses (modelled data that include historical observations), or GCMs for future climate projections. Reanalyses and GCMs are typically highly dimensional large gridded spatiotemporal datasets which require careful feature selection and/or dimensional reduction for ML applications. Fire occurrence prediction, susceptibility, and risk applications used a large number of di erent environmental variables as predictors, but almost all used re locations and associated temporal information as predictands. Fire data itself is usually collated from re management agencies in the form of georeferenced points or perimeter 30 data, along with reported dates, ignition cause, and other related variables. Care should be taken using such data because changes in reporting standards or accuracy may lead to data inhomogeneity. As well as re locations and perimeters, re severity is an attribute of much interest to re scientists. Fire severity is often determined from remotely sensed data and represented using variables such as the Di erenced Nor- malized Burn Ratio (dNBR) and variants, or through eld sampling. However, remote sensed estimates of burn severity should be considered as proxies as they have low skill in some ecosystems. Other re ecology research historically relies on in situ eld, sampling although many of the ML applications attempt to resolve features of interest using remote sensed data. Smoke data can also be derived from remote sensed imagery or from air quality sensors (e.g., PM2.5, atmospheric particulate matter less than 2.5 m). Continued advances in remote sensing, as well as the quality and availability of remote sensed data prod- ucts, in weather and climate modeling have led to increased availability of large spatiotemporal datasets, which presents both an opportunity and challenge for the application of ML methods in wild re research and management. The era of \big data" has seen the development of cloud computing platforms to provide the computing and data storage facilities to deal with these large datasets. For example, in our review we found two papers [Crowley et al., 2019, Quintero et al., 2019] that used Google Earth Engine which inte- grates geospatial datasets with a coding environment [Gorelick et al., 2017]. In any case, data processing and management plays an important role in the use of large geospatial datasets. 5.2 Model selection and accuracy Given a wild re science question or management problem and available relevant data, a critical question to ask is what is the most appropriate modeling tool to address the problem? Is it a standard statistical model (e.g., linear regression or LR), a physical model (e.g., FIRETEC or other re simulator), a ML model, or a combination of approaches? Moreover, which speci c algorithm will yield the most accurate classi cation or regression. Given the heterogeneity of research questions, study areas, and datasets considered in the papers reviewed here, it is not possible to comprehensively answer these questions with respect to ML approaches. Even in the case where multiple studies used the same dataset [Cortez and Morais, 2007, Sa and Bouroumi, 2013, Storer and Green, 2016, Castelli et al., 2015, Al Janabi et al., 2018, Alberg, 2015, Li et al., 2018a, Castelli et al., 2015] the di erent research questions considered meant a direct comparison of ML methods was not possible between research studies. However, a number of individual studies did make comparisons between multiple ML methods, or between ML and statistical methods for a given wild re modeling problem and dataset. Here we highlight some of their ndings to provide some guidance with respect to model selection. In our review (see section 4 and the supplementary material), we found 29 papers comparing ML and statistical methods, where in the majority of these cases ML methods were found to be more accurate than traditional statistical methods (e.g., GLMs), or displayed similar performance [Pu and Gong, 2004, Bates et al., 2017, de Bem et al., 2018]. In only one study on climate change by Amatulli et al. [2013], MARS was found to be superior to RF for their analytical task. A sizable number of the comparative studies (14) involved classi cation problems that used LR as a benchmark method against ANN or ensemble tree methods. For studies comparing multiple ML methods, there was considerable variation in the choice of most accurate method; however, in general ensemble methods tended to outperform single classi er methods (e.g., Stojanova et al. [2012], Dutta et al. [2016], Mayr et al. [2018], Nelson et al. [2017], Reid et al. [2015], Watson et al. [2019]), except in one case where the most accurate model (CART) was also the most parsimonious [Coeld et al., 2019]. A few more recent papers also highlighted the advantages of DL over other methods. In particular, for re detection, Zhang et al. [2018b] compared CNNs with SVM and found that CNNs were more accurate, while Zhao et al. [2018] similarly found CNNs superior to SVMs and ANNs. For re susceptibility mapping, Zhang et al. [2019] found CNNs were more accurate than RF, SVMs, and ANNs. For time series forecasting problems, Liang et al. [2019] found LSTMs outperformed ANNs. Finally, Cao et al. [2019] found that using an LSTM combined with a CNN led to better re detection performance from video compared with CNNs alone. In any case, more rigorous inter-model comparisons are needed to reveal in which conditions, and in what sense particular methods are more accurate, as well as to establish procedures for evaluating accuracy. 31 ML methods are also prone to over tting, so it is important to evaluate models with robust test datasets using appropriate cross-validation strategies. For example, the na ve application of cross-validation to data that have spatial or spatio-temporal dependencies may lead to overly optimistic evaluations [Roberts et al., 2017]. In general, one also desires to minimise errors associated with either under-speci cation or over- speci cation of the model, a problem known as the bias-variance trade-o [Geman et al., 1992]. However, several recent advances have been made to reduce over tting in ML models, for instance, regularization techniques in DNNs [Kuka cka et al., 2017]. Moreover, when interpreting comparisons between ML and statistical methods, we should be cognizant that just as some ML methods require expert knowledge, the accuracy of statistical methods can also vary with the skill of the practitioner. Thompson and Calkin [2011] also emphasize the need for identifying sources of uncertainty in modeling so that they can better managed. 5.3 Implementation Challenges Beyond data and model selection, two important considerations for model speci cation are feature selection and spatial autocorrelation. Knowledge of the problem domain is extremely important in identifying a set of candidate features. However, while many ML methods are not limited by the number of features, more variables do not necessarily make for a more accurate, interpretable, or easily implemented model [Schoenberg, 2016, Breiman, 2001] and can lead to over tting and increased computational time. Two di erent ML methods to enable selection of a reduced and more optimal set of features include GAs and PSO. Sachdeva et al. [2018] used a GA to select input features for BRT and found this method gave the best accuracy compared with ANN, RF, SVM, SVM with PSO (PSO-SVM), DTs, logistic regression, and NB. Hong et al. [2018] employed a similar approach for re susceptibility mapping and found this led to improvements for both SVM and RF compared with their non-optimized counterparts. Tracy et al. [2018] used a novel random subset feature selection algorithm for feature selection, which they found led to higher AUC values and lower model complexity. Jaafari et al. [2019] used a NFM combined with the imperialist competitive algorithm (a variant of GA) for feature selection which led to very high model accuracy (0.99) in their study. Tien Bui et al. [2017] used PSO to choose inputs to a NFN and found this improved results. [Zhang et al., 2019] also considered the information gain ratio for feature selection. As noted in Moritz et al. [2012] and Mayr et al. [2018], one should also take spatial autocorrelation into account when modeling re probabilities spatially. In general, the presence of spatial autocorrelation violates the assumption of independence for parametric models, which can degrade model performance. One approach to deal with autocorrelation requires subsampling to remove any spatial autocorrelation Moritz et al. [2012]. It is also often necessary to subsample from non- re locations due to class imbalance between ignitions and non- ignitions (e.g., Cao et al. [2017], Zhang et al. [2019]). Song et al. [2017] considered spatial econometric models and found a spatial autocorrelation model worked better than RF, although Kim et al. [2019] note that RF may be robust to spatial autocorrelation with large samples. In contrast to many ML methods, a strength of CNNs is its ability to exploit spatial correlation in the data to enable the extraction of spatial features. 5.4 Interpretation A major obstacle for the adoption of ML methods to re modeling tasks is the perceived lack of inter- pretability or explainability of such methods, which are often considered to be \black box" models. Users (in this case re ghters and managers) need to trust ML model predictions, and so have the con dence and justi cation to apply these models, particularly in cases where proposed solutions are considered novel. Model intepretability should therefore be an important aspect of model development if models are to be selected and deployed in re management operations. Model interpretability varies signi cantly across the di erent types of ML. For example, conventional thinking is that tree-based methods are more inter- pretable than neural network methods. This is because a single decision tree classi er can be rendered as a ow chart corresponding to if-then-else statements, whereas an ANN represents a nonlinear function 32 approximated through a series of nonlinear activations. However, because they combine multiple trees in an optimized way, ensemble tree classi ers are less interpretable than single tree classi ers. On the other hand, BNs are one example of an ML technique where good explanations for results can be inferred due to their graphical representation; however, full Bayesian learning on large-scale data is very computation- ally expensive which may have limited early applications; however, as computational power has increased we have seen an increase in the popularity of BNs in wild re science and management applications (e.g., Penman et al. [2015], Papakosta et al. [2017]). DL-based architectures are widely considered to be among the least interpretable ML models, despite the fact that they can achieve very accurate function approximation [Chakraborty et al., 2017]. In fact, this is demonstrative of the well-known trade-o between prediction accuracy and interpretability (see Kuhn and Johnson [2013] for an in-depth discussion). The ML community, however, recognizes the problem of interpretability and work is underway to develop methods that allow for greater interpretability of ML methods, including methods for DL (see for example, McGovern et al. [2019]) or model-agnostic approaches [Ribeiro et al., 2016]. Runge et al. [2019] further argue that casual inference methods should be used in conjunction with predictive models to improve our understanding of physical systems. Finally, it is worth noting that assessing variable importance (see Sec. 4.3.4) for a given model can play a role in model interpretation. 5.5 Opportunities Our review highlights a number of potential opportunities in wild re science and management for ML applications where ML has not yet been applied or is under-utilized. Here we examine ML advances in other areas of environmental science that have analogous problems in wildland re science and which may be useful for identifying further ML applications. For instance, Li et al. [2011] compared ML algorithms for spatial interpolation and found that a RF model combined with geostatistical methods yielded good results; a similar method could be used to improve interpolation of re weather observations from weather stations, and so enhance re danger monitoring. Rasp and Lerch [2018] showed that ANNs could improve weather forecasts by post-processing ensemble forecasts, an approach which could similarly be applied to improve short-term forecasts of re weather. Belayneh et al. [2014] used ANNs and SVMs combined with wavelet transforms for long term drought forecasting in Ethiopia; such methods could also be useful for forecasting drought in the context of re danger potential. In the context of numerical weather prediction, Cohen et al. [2019] found better predictability using ML methods than dynamical models for subseasonal to seasonal weather forecasting, suggesting similar applications for long-term re weather forecasting. McGovern et al. [2017] discussed how AI techniques can be leveraged to improve decision making around high-impact weather. More recently, Reichstein et al. [2019] have further argued for the use of DL in the environmental sciences, citing its potential to extract spatiotemporal features from large geospatial datasets. Kussul et al. [2017] used CNNs to classify land cover and crop types and found that CNNs improved the results over standard ANN models; a similar approach could be used for fuels classi cation, which is an important input to re behaviour prediction models. Shi et al. [2016] also used CNNs to detect clouds in remote sensed imagery and were able to di erentiate between thin and thick cloud. A similar approach could be used for smoke detection, which is important for re detection, as well as in determining the presence of false negatives in hotspot data (due to smoke or cloud obscuration). Finally, recent proposals have called for hybrid models that combine process-based models and ML methods [Reichstein et al., 2019]. For example, ML models may replace user-speci ed parameterizations in numerical weather prediction models [Brenowitz and Bretherton, 2018]. Other recent approaches use ML methods to determine the solutions to nonlinear partial di erential equations Raissi and Karniadakis [2018], Raissi et al. [2019]. Such methods could nd future applications in improving re behaviour prediction models based on computationally expensive physics-based re simulators, in coupled re-atmosphere models, or in smoke dispersion modeling. In any case, the applications of ML that we have outlined are meant for illustrative purposes and are not meant to represent an exhaustive list of all possible applications. 33 5.6 Implications for re management We believe ML has been under-utilized in re management, particularly with respect to problems belonging to either predictive or prescriptive analytics. Fire management comprises a set of risk control measures, which are often cast in the framework of the emergency response phases: prevention; mitigation; prepared- ness; response; recovery; and review [Tymstra et al., 2019]. In terms of nancial expenditure, by far the largest percentage spent in the response phase [Stocks and Martell, 2016]. In practice, re management is largely determined by the need to manage resources in response to active or expected wild res, typically for lead times of days to weeks, or to manage vegetative fuels. This suggests the opportunity for increased research in areas of re weather prediction, re occurrence prediction, and re behaviour prediction, as well as optimizing re operations and fuel treatments. The identi cation of these areas, as well as the fact that wild re is both a spatial and temporal process, further reiterate the need for ML applications for time series forecasting. From this review, there were few papers that used time series ML methods for forecasting problems, suggesting an opportunity for further work in this area. In particular, recurrent neural networks (RNNs) were used for re behavior prediction [Cheng and Wang, 2008, Kozik et al., 2013, 2014] and re occurrence prediction [Dutta et al., 2013]. The most common variant of RNNs are Long Short Term Memory (LSTM) networks [Hochreiter and Schmidhuber, 1997], which have been used for burned area prediction [Liang et al., 2019] and re detection [Cao et al., 2019]. Because these methods implicitly model dynamical processes, they should lead to improve forecasting models compared with standard ANNs. For example Gensler et al. [2017] have used LSTMs to forecast solar power and Kim et al. [2017] used CNNs combined with LSTM for forecasting precipitation. We anticipate that these methods could also be employed for re weather, re occurrence, and re behaviour prediction. We note that there are a number of operational research and management science methods used in re management research including queuing, optimization, and simulation of complex system dynamics (e.g., Martell [2015]) where ML algorithms don't seem to provide an obvious alternative. For example, planning models to simulate the interactions between re management resource con gurations and re dynamics reviewed by [Mavsar et al., 2013]. From our review, a few papers used agent-based learning methods for re management. In particular, reinforcement learning was used for optimizing fuel treatments [Lauer et al., 2017] or for autonomous control of aircraft for re monitoring [Julian and Kochenderfer, 2018a]. GAs were used for generating optimal relines for active res [Homchaudhuri et al., 2010] and for reducing the time for re simulation [Cencerrado et al., 2014]. However, more work is needed to identify where ML methods could contribute to tactical, operational, or strategic re management decision making. An important challenge for the re research and management communities is enabling the transition of potentially useful ML models to re management operations. Although we identi ed several papers that emphasized their ML models could be deployed in re management operations [Art es et al., 2016, Alonso-Betanzos et al., 2002, Iliadis, 2005, Stojanova et al., 2012, Davis et al., 1989, 1986, Liu et al., 2015], it can be dicult to assess whether and how a study has been adopted by, or in uenced, re management agencies. This challenge is often exacerbated by a lack of resources and/or funding, as well as the di erent priorities and institutional cultures of researchers and re managers. One possible solution to this problem would be the formation of working groups dedicated to enabling this transition, preferably at the research proposal phase. In general, enabling operational ML methods will require tighter integration and greater collaboration between the research and management communities, particularly with regards to project design, data compilation and variable selection, implementation, and interpretation. However, it is worth noting that this is not a problem unique to ML, it is a long-standing and common issue in many areas of re research and other applied science disciplines, where continuous e ort is required to maintain communications and relationships between researchers and practitioners. Finally, we would like to stress that we believe the wild re research and management communities should play an active role in providing relevant, high quality, and freely available wild re data for use by practitioners of ML methods. For example, burned area and re weather data made available by Cortez and Morais [2007] was subsequently used by a number of authors in their work. It is imperative that the 34 quality of data collected by management agencies be as robust as possible, as the results of any modelling process are dependent upon the data used for analysis. It is worth considering how new data on, for example, hourly re growth or the daily use of re management resources, could be used in ML methods to yield better predictions or management recommendations | using new tools to answer new questions may require better or more complete data. Conversely, we must recognize that despite ML models being able to learn on their own, expertise in wild re science is necessary to ensure realistic modelling of wild re processes, while the complexity of some ML methods (e.g., DL) requires a dedicated and sophisticated knowledge of their application (we note that many of the most popular ML methods used in this study are fairly easy to implement, such as RF, MaxEnt, and DTs). The observation that no single ML algorithm is superior for all classes of problem, an idea encapsulated by the \no free lunch" theorem [Wolpert, 1996], further reinforces the need for domain-speci c knowledge. Thus, the proper implementation of ML in wild re science is a challenging endeavor, often requiring multidisciplinary teams and/or interdisciplinary specialists to e ectively produce meaningful results. 5.7 A word of caution ML holds tremendous potential for a number of wild re science and management problem domains. As indicated in this review, much work has already been undertaken in a number of areas, although further work is clearly needed for re management speci c problems. Despite this potential, ML should not be considered a panacea for all re research areas. ML is best suited to problems where there is sucient high- quality data, and this is not always the case. For example, for problems related to re management policy, data is needed at large spatiotemporal scales (i.e., ecosystem/administrative spatial units at timescales of decades or even centuries), and such data may simply not yet exist in current inventories. At the other extreme, data is needed at very ne spatiotemporal scales for re spread and behavior modeling, including high resolution fuel maps and surface weather variables which are often not available at the required scale and are dicult to acquire even in an experimental context. Another limitation of ML may occur when one attempts make predictions where no analog exists in the observed data, such as may be the case with climate change prediction. 6 Conclusions Our review shows that the application of ML methods in wild re science and management has been steadily increasing since their rst use in the 1990s, across core problem domains using a wide range ML methods. The bulk of work undertaken thus far has used traditional methods such as RF, BRT, MaxEnt, SVM and ANNs, partly due to the ease of application and partly due to their simple interpretability in many cases. However, problem domains associated with predictive (e.g., predicted re behavior) or prescriptive analytics (e.g. optimizing re management decisions) have seen much less work with ML methods. We therefore suggest opportunities exist for both the wild re community and ML practitioners to apply ML methods in these areas. Moreover, the increasing availability of large spatio-temporal datasets, from climate models or remote sensing for example, may be amenable to the use of deep learning methods, which can eciently extract spatial or temporal features from data. Another major opportunity is the application of agent based learning to re management operations, although many other opportunities exist. However, we must recognize that despite ML models being able to learn on their own, expertise in wild re science is necessary to ensure realistic modelling of wild re processes across multiple scales, while the complexity of some ML methods (e.g. DL) requires a dedicated and sophisticated knowledge of their application. Furthermore, a major obstacle for the adoption of ML methods to re modeling tasks is the perceived lack of interpretability of such methods, which are often considered to be black box models. The ML community, however, recognizes this problem and work is underway to develop methods that allow for greater interpretability of ML methods (see for example, [McGovern et al., 2019]). Data driven approaches are by de nition data dependent | if the re management community wants to more fully exploit powerful 35 ML methods, we need to consider data as a valuable resource and examine what further information on re events or operations are needed to apply ML approaches to management problems. Thus, wildland re science is a diverse multi-faceted discipline that requires a multi-pronged approach, a challenge made greater by the need to mitigate and adapt to current and future re regimes. Acknowledgments The motivation for this paper arose from the \Not the New Normal" BC AI Wild re Symposium held in Vancouver, BC, on 12 October 2018. The authors would also like to thank Intact Insurance, the Canadian Partnership for Wildland Fire Science, the NSERC Discovery Grants program and the Microsoft AI for Social Good program for their support. References Baker Abdalhaq, Ana Cort es, Tom as Margalef, and Emilio Luque. Enhancing wildland re prediction on cluster systems applying evolutionary optimization techniques. Future Generation Computer Systems, 21(1):61{67, 2005. ISSN 0167739X. doi: 10.1016/j.future.2004.09.013. Hamed Adab. Land re hazard assessment in the Caspian Hyrcanian forest ecoregion with the long-term MODIS active re data. Natural Hazards, 87(3):1807{1825, jul 2017. ISSN 0921-030X. doi: 10.1007/ s11069-017-2850-2. URL http://link.springer.com/10.1007/s11069-017-2850-2. Hamed Adab, Azadeh Atabati, Sandra Oliveira, and Ahmad Moghaddam Gheshlagh. Assessing re hazard potential and its main drivers in Mazandaran province, Iran: a data-driven approach. Environmental Monitoring and Assessment, 190(11):670, nov 2018. ISSN 0167-6369. doi: 10.1007/s10661-018-7052-1. URL http://link.springer.com/10.1007/s10661-018-7052-1. Moulay A. Akhlou , Roger Booto Tokime, and Hassan Elassady. Wildland res detection and segmentation using deep learning. In Mohammad S. Alam, editor, Pattern Recognition and Track- ing XXIX, page 11. SPIE, apr 2018. ISBN 9781510618091. doi: 10.1117/12.2304936. URL https://www.spiedigitallibrary.org/conference-proceedings-of-spie/10649/2304936/ Wildland-fires-detection-and-segmentation-using-deep-learning/10.1117/12.2304936. full. K. R. Al-Rawi, J. L. Casanova, and A. Calle. Burned area mapping system and re detection system, based on neural networks and NOAA-AVHRR imagery. International Journal of Remote Sensing, 22(10):2015{ 2032, jan 2001. ISSN 0143-1161. doi: 10.1080/01431160117531. URL https://www.tandfonline.com/ doi/full/10.1080/01431160117531. K R Al-Rawi, J L Casanova, A Romo, and E M Louakfaoui. Integrated re evolution monitoring system (IFEMS) for monitoring spatial-temporal behaviour of multiple re phenomena. International Journal of Remote Sensing, 23(10):1967{1983, 2002. ISSN 01431161. doi: 10.1080/01431160110069809. URL http://www.tandfonline.com/action/journalInformation?journalCode=tres20. Dima Alberg. An Interval Tree Approach to Predict Forest Fires using Meteorological Data. International Journal of Computer Applications, 132(4):17{22, 2015. doi: 10.5120/ijca2015907398. Andrew Aldersley, Steven J. Murray, and Sarah E. Cornell. Global and regional analysis of climate and human drivers of wild re. Science of the Total Environment, 409(18):3472{3481, 2011. ISSN 00489697. doi: 10.1016/j.scitotenv.2011.05.032. URL http://dx.doi.org/10.1016/j.scitotenv.2011.05.032. 36 Dmitriy Alexandrov, Elizaveta Pertseva, Ivan Berman, Igor Pantiukhin, and Aleksandr Kapitonov. Anal- ysis of machine learning methods for wild re security monitoring with an unmanned aerial vehicles. In Conference of Open Innovation Association, FRUCT, volume 2019-April, pages 3{9. IEEE Computer Society, may 2019. ISBN 9789526865386. doi: 10.23919/FRUCT.2019.8711917. Samaher Al Janabi, Ibrahim Al Shourbaji, and Mahdi A. Salman. Assessing the suitability of soft comput- ing approaches for forest res prediction. Applied Computing and Informatics, 14(2):214{224, jul 2018. ISSN 2210-8327. doi: 10.1016/J.ACI.2017.09.006. URL https://www.sciencedirect.com/science/ article/pii/S2210832717301539. A. Alonso-Benito, P. A. Hernandez-Leal, A. Gonzalez-Calvo, M. Arbelo, and A. Barreto. Analysis of Di erent Methods for Burnt Area Estimation using Remote Sensing and Ground Truth Data. In IGARSS 2008 - 2008 IEEE International Geoscience and Remote Sensing Symposium, pages III { 828{III { 831. IEEE, 2008. ISBN 978-1-4244-2807-6. doi: 10.1109/IGARSS.2008.4779477. URL http://ieeexplore. ieee.org/document/4779477/. Amparo Alonso-Betanzos, Oscar Fontenla-Romero, Bertha Guijarro-BerdinasBerdi, ~ Elena Hern andez- Pereira, Juan Canda, Eulogio Jimenez, Jos e Luis Legido, Susana MunizMu, ~ Cristina Paz-Andrade, and Mar a Inmaculada Paz-Andrade. A Neural Network Approach for Forestal Fire Risk Estimation. Technical report, 2002. Amparo Alonso-Betanzos, Oscar Fontenla-Romero, Bertha Guijarro-Berdin ~as, Elena Hern andez-Pereira, Mar a Inmaculada Paz Andrade, Eulogio Jim enez, Jose Luis Legido Soto, and Tarsy Carballas. An intelligent system for forest re risk prediction and re ghting management in Galicia. Expert Systems with Applications, 25(4):545{554, 2003. ISSN 09574174. doi: 10.1016/S0957-4174(03)00095-2. N. S. Altman. An introduction to kernel and nearest-neighbor nonparametric regression. American Statis- tician, 46(3):175{185, 1992. ISSN 15372731. doi: 10.1080/00031305.1992.10475879. Giuseppe Amatulli and Andrea Camia. Exploring the relationships of re occurrence variables by means of CART and MARS models. In Proceedings of the 4th International Wildland Fire Conference, pages 1{11, 2007. Giuseppe Amatulli, Maria Jo~ ao Rodrigues, Marco Trombetti, and Ra aella Lovreglio. Assessing long- term re risk at local scale by means of decision tree technique. Journal of Geophysical Research: Biogeosciences, 111(G4), dec 2006. ISSN 01480227. doi: 10.1029/2005JG000133. URL http://doi. wiley.com/10.1029/2005JG000133. Giuseppe Amatulli, Andrea Camia, and Jesus  San-Miguel-Ayanz. Estimating future burned areas un- der changing climate in the EU-Mediterranean countries. Science of The Total Environment, 450- 451:209{222, apr 2013. ISSN 0048-9697. doi: 10.1016/J.SCITOTENV.2013.02.014. URL https: //www.sciencedirect.com/science/article/pii/S0048969713001770. K. Angayarkkani and N. Radhakrishnan. An Intelligent System For E ective Forest Fire Detection Using Spatial Data. feb 2010. URL http://arxiv.org/abs/1002.2199. K. Angayarkkani and N. Radhakrishnan. An e ective technique to detect forest re region through AN- FIS with spatial data. In ICECT 2011 - 2011 3rd International Conference on Electronics Computer Technology, volume 3, pages 24{30, 2011. ISBN 9781424486779. doi: 10.1109/ICECTECH.2011.5941794. Bachisio Arca, Tiziano Ghisu, and Giuseppe A. Trun o. GPU-accelerated multi-objective optimization of fuel treatments for mitigating wild re hazard. Journal of Computational Science, 11:258{268, 2015. ISSN 18777503. doi: 10.1016/j.jocs.2015.08.009. 37 Sally Archibald, David P. Roy, Brian W. van Wilgen, and Robert J. Scholes. What limits re? An examination of drivers of burnt area in Southern Africa. Global Change Biology, 15(3):613{630, 2009. ISSN 13541013. doi: 10.1111/j.1365-2486.2008.01754.x. Juan P. Arganaraz, ~ Gregorio Gavier Pizarro, Marcelo Zak, Marcos A. Landi, and Laura M Bellis. Hu- man and biophysical drivers of res in Semiarid Chaco mountains of Central Argentina. Science of The Total Environment, 520:1{12, jul 2015. ISSN 0048-9697. doi: 10.1016/J.SCITOTENV.2015. 02.081. URL https://www-sciencedirect-com.login.ezproxy.library.ualberta.ca/science/ article/pii/S0048969715002338. Hilary Arksey and Lisa O'Malley. Scoping studies: towards a methodological framework. International Journal of Social Research Methodology, 8(1):19{32, feb 2005. ISSN 1364- 5579. doi: 10.1080/1364557032000119616. URL http://www.tandfonline.com/doi/abs/10.1080/ James D. Arnold, Simon C. Brewer, and Philip E. Dennison. Modeling Climate-Fire Connections within the Great Basin and Upper Colorado River Basin, Western United States. Fire Ecology, 10(2):64{75, aug 2014. ISSN 19339747. doi: 10.4996/ reecology.1002064. URL http://fireecologyjournal.org/ journal/abstract/?abstract=220. A. Arpaci, B. Malowerschnig, O. Sass, and H. Vacik. Using multi variate data mining techniques for estimating re susceptibility of Tyrolean forests. Applied Geography, 53:258{270, sep 2014. ISSN 0143-6228. doi: 10.1016/J.APGEOG.2014.05.015. URL https://www.sciencedirect.com/science/ article/abs/pii/S0143622814001106. B.C. Arrue, A. Ollero, and J.R. Matinez de Dios. An intelligent system for false alarm reduction in infrared forest- re detection. IEEE Intelligent Systems, 15(3):64{73, may 2000. ISSN 1094-7167. doi: 10.1109/5254.846287. URL http://ieeexplore.ieee.org/document/846287/. Tom as Art es, Andr es Cencerrado, Ana Cort es, and Tom as Margalef. Core Allocation Policies on Multicore Platforms to Accelerate Forest Fire Spread Predictions. PPAM 2013: Parallel Processing and Applied Mathematics, pages 151{160, 2014. ISSN 16113349. doi: 10.1007/978-3-642-55195-6. URL https: //link.springer.com/chapter/10.1007/978-3-642-55195-6{_}14{#}enumeration. Tom as Art es, Andr es Cencerrado, Ana Cort es, and Tom as Margalef. Real-time genetic spatial opti- mization to improve forest re spread forecasting in high-performance computing environments. In- ternational Journal of Geographical Information Science, 30(3):594{611, 2016. ISSN 13623087. doi: 10.1080/13658816.2015.1085052. URL http://dx.doi.org/10.1080/13658816.2015.1085052. Tom as Art es, Andr es Cencerrado, Ana Cort es, and Tom as Margalef. Time aware genetic algorithm for forest re propagation prediction: exploiting multi-core platforms. Concurrency Computation, 29(9): 1{18, 2017. ISSN 15320634. doi: 10.1002/cpe.3837. Davide Ascoli, Giorgio Vacchiano, Renzo Motta, and Giovanni Bovio. Building Rothermel re behaviour fuel models by genetic algorithm optimisation. International Journal of Wildland Fire, 24(3):317{328, 2015. ISSN 10498001. doi: 10.1071/WF14097. Rui Ba, Chen Chen, Jing Yuan, Weiguo Song, and Siuming Lo. SmokeNet: Satellite Smoke Scene Detection Using Convolutional Neural Network with Spatial and Channel-Wise Attention. Remote Sensing, 11(14): 1702, jul 2019. ISSN 2072-4292. doi: 10.3390/rs11141702. URL https://www.mdpi.com/2072-4292/ 11/14/1702. Shitai Bao, Ningchuan Xiao, Zehui Lai, Heyuan Zhang, and Changjoo Kim. Optimizing watchtower locations for forest re monitoring using location models. Fire Safety Journal, 71(December 2013):100{ 109, 2015. ISSN 03797112. doi: 10.1016/j. resaf.2014.11.016. URL http://dx.doi.org/10.1016/j. firesaf.2014.11.016. 38 Avi Bar Massada, Alexandra D. Syphard, Susan I. Stewart, and Volker C. Radelo . Wild re ignition- distribution modelling: a comparative study in the Huron?Manistee National Forest, Michigan, USA. International Journal of Wildland Fire, 22(2):174, apr 2013. ISSN 1049-8001. doi: 10.1071/WF11178. URL http://www.publish.csiro.au/?paper=WF11178. Panagiotis Barmpoutis, Kosmas Dimitropoulos, Kyriaki Kaza, and Nikos Grammalidis. Fire Detection from Images Using Faster R-CNN and Multidimensional Texture Analysis. In ICASSP, IEEE International Conference on Acoustics, Speech and Signal Processing - Proceedings, volume 2019-May, pages 8301{ 8305. Institute of Electrical and Electronics Engineers Inc., may 2019. ISBN 9781479981311. doi: 10.1109/ICASSP.2019.8682647. K. Barrett, A. D. McGuire, E. E. Hoy, and E. S. Kasischke. Potential shifts in dominant forest cover in interior Alaska driven by variations in re severity. Ecological Applications, 21(7):2380{2396, 2011. ISSN 10510761. doi: 10.1890/10-0896.1. Hossein Bashari, Ali Asghar Naghipour, Seyed Jamaleddin Khajeddin, Hamed Sangoony, and Pejman Tah- masebi. Risk of re occurrence in arid and semi-arid ecosystems of Iran: an investigation using Bayesian belief networks. Environmental Monitoring and Assessment, 188(9):531, sep 2016. ISSN 0167-6369. doi: 10.1007/s10661-016-5532-8. URL http://link.springer.com/10.1007/s10661-016-5532-8. Bryson C. Bates, Andrew J. Dowdy, Richard E. Chandler, Bryson C. Bates, Andrew J. Dowdy, and Richard E. Chandler. Classi cation of Australian Thunderstorms Using Multivariate Analyses of Large- Scale Atmospheric Variables. Journal of Applied Meteorology and Climatology, 56(7):1921{1937, jul 2017. ISSN 1558-8424. doi: 10.1175/JAMC-D-16-0271.1. URL http://journals.ametsoc.org/doi/ 10.1175/JAMC-D-16-0271.1. Enric Batllori, Marc-Andr e Parisien, Meg A. Krawchuk, and Max A. Moritz. Climate change-induced shifts in re for Mediterranean ecosystems. Global Ecology and Biogeography, 22(10):1118{1129, oct 2013. ISSN 1466822X. doi: 10.1111/geb.12065. URL http://doi.wiley.com/10.1111/geb.12065. Peter Bauer, Alan Thorpe, and Gilbert Brunet. The quiet revolution of numerical weather prediction. Nature, 525(7567):47{55, sep 2015. ISSN 0028-0836. doi: 10.1038/nature14956. URL http://www. nature.com/articles/nature14956. A. Belayneh, J. Adamowski, B. Khalil, and B. Ozga-Zielinski. Long-term SPI drought forecasting in the Awash River Basin in Ethiopia using wavelet neural networks and wavelet support vector regression models. Journal of Hydrology, 508:418{429, jan 2014. ISSN 00221694. doi: 10.1016/j.jhydrol.2013.10.052. Mar Bisquert, Eduardo Caselles, Juan Manuel S anchez, and Vicente Caselles. Application of arti cial neural networks and logistic regression to the prediction of forest re danger in Galicia using MODIS data. International Journal of Wildland Fire, 21(8):1025, dec 2012. ISSN 1049-8001. doi: 10.1071/WF11105. URL http://www.publish.csiro.au/?paper=WF11105. Karen D. Blouin, Mike D. Flannigan, Xianli Wang, and Bohdan Kochtubajda. Ensemble lightning predic- tion models for the province of Alberta, Canada. International Journal of Wildland Fire, 25(4):421{432, apr 2016. ISSN 1049-8001. doi: 10.1071/WF15111. URL http://www.publish.csiro.au/?paper= WF15111. William J. Bond and Jon E. Keeley. Fire as a global 'herbivore': The ecology and evolution of ammable ecosystems. Trends in Ecology and Evolution, 20(7):387{394, jul 2005. ISSN 01695347. doi: 10.1016/j. tree.2005.04.025. Yan Boulanger, Marc-Andr e Parisien, and Xianli Wang. Model-speci cation uncertainty in future area burned by wild res in Canada. International Journal of Wildland Fire, 27(3):164, apr 2018. ISSN 1049-8001. doi: 10.1071/WF17123. URL http://www.publish.csiro.au/?paper=WF17123. 39 Curtis M. Bradley, Chad T. Hanson, and Dominick A. DellaSala. Does increased forest protection corre- spond to higher re severity in frequent- re forests of the western United States? Ecosphere, 7(10):1{13, 2016. ISSN 21508925. doi: 10.1002/ecs2.1492. Jonathan Branham, Nicholas Hamilton, Dale Hamilton, and Barry Myers. Evaluation of Image Spatial Resolution for Machine Learning Mapping of Wildland Fire E ects, jul 2017. URL https://scholarworks.boisestate.edu/icur/2017/Poster{_}Session/26https:// link.springer.com/chapter/10.1007/978-3-030-01054-6{_}29. Leo Breiman. Statistical modeling: The two cultures. Statistical Science, 16(3):199{215, 2001. ISSN 08834237. doi: 10.1214/ss/1009213726. Leo Breiman. Classi cation and regression trees. Routledge, 2017. Leo Breiman, Jerome Friedman, Richard A Olshen, and Charles J Stone. Classi cation and regression trees Chapman & Hall. New York, 1984. N. D. Brenowitz and C. S. Bretherton. Prognostic Validation of a Neural Network Uni ed Physics Parameterization. Geophysical Research Letters, 45(12):6289{6298, jun 2018. ISSN 00948276. doi: 10.1029/2018GL078510. URL http://doi.wiley.com/10.1029/2018GL078510. Steven P. Brumby, Neal R. Harvey, Je rey J. Bloch, James P. Theiler, Simon J. Perkins, Aaron C. Young, and John J. Szymanski. Evolving forest re burn severity classi cation algorithms for multispectral imagery. Algorithms for Multispectral, Hyperspectral, and Ultraspectral Imagery VII, 4381(August 2001): 236{245, 2001. doi: 10.1117/12.437013. C. E. Buckland, R. M. Bailey, and D. S. G. Thomas. Using arti cial neural networks to predict future dryland responses to human and climate disturbances. Scienti c Reports, 9(1):3855, dec 2019. ISSN 2045-2322. doi: 10.1038/s41598-019-40429-5. URL http://www.nature.com/articles/ s41598-019-40429-5. Declan Butler. AI summit aims to help world's poorest, jun 2017. ISSN 14764687. Wenhua Cai, Jian Yang, Zhihua Liu, Yuanman Hu, and Peter J. Weisberg. Post- re tree recruitment of a boreal larch forest in Northeast China. Forest Ecology and Management, 307:20{29, 2013. ISSN 03781127. doi: 10.1016/j.foreco.2013.06.056. URL http://dx.doi.org/10.1016/j.foreco.2013.06.056. X. Cao, J. Chen, B. Matsushita, H. Imura, and L. Wang. An automatic method for burn scar mapping using support vector machines. International Journal of Remote Sensing, 30(3):577{594, feb 2009. ISSN 0143-1161. doi: 10.1080/01431160802220219. URL https://www.tandfonline.com/doi/full/ 10.1080/01431160802220219. Yichao Cao, Feng Yang, Qingfei Tang, and Xiaobo Lu. An Attention Enhanced Bidirectional LSTM for Early Forest Fire Smoke Recognition. IEEE Access, pages 1{1, oct 2019. doi: 10.1109/access.2019. Yinxue Cao, Ming Wang, and Kai Liu. Wild re Susceptibility Assessment in Southern China: A Compar- ison of Multiple Methods. International Journal of Disaster Risk Science, 8(2):164{181, jun 2017. ISSN 21926395. doi: 10.1007/s13753-017-0129-6. Adri an Cardil, Blas Mola-Yudego, Angela Bl azquez-Casado, and Jos e Ram on Gonz alez-Olabarria. Fire and burn severity assessment: Calibration of Relative Di erenced Normalized Burn Ratio (RdNBR) with eld data. Journal of Environmental Management, 235(January):342{349, 2019. ISSN 10958630. doi: 10.1016/j.jenvman.2019.01.077. URL https://doi.org/10.1016/j.jenvman.2019.01.077. 40 Carlos Carrillo, Tom as Art es, Ana Cort es, and Tom as Margalef. Error function impact in dynamic data- driven framework applied to forest re spread prediction. Procedia Computer Science, 80:418{427, 2016. ISSN 18770509. doi: 10.1016/j.procs.2016.05.342. URL http://dx.doi.org/10.1016/j.procs.2016. 05.342. Mauro Castelli, Leonardo Vanneschi, and Ale s Popovi c. PREDICTING BURNED AREAS OF FOR- EST FIRES: AN ARTIFICIAL INTELLIGENCE APPROACH. Fire Ecology, 11(1):106{118, apr 2015. ISSN 19339747. doi: 10.4996/ reecology.1101106. URL http://fireecologyjournal.org/journal/ abstract/?abstract=236. Turgay Celik. Change detection in satellite images using a genetic algorithm approach. IEEE Geoscience and Remote Sensing Letters, 7(2):386{390, 2010. ISSN 1545598X. doi: 10.1109/LGRS.2009.2037024. Andr es Cencerrado, Ana Cort es, and Tom as Margalef. Genetic algorithm characterization for the quality assessment of forest re spread prediction. Procedia Computer Science, 9:312{320, 2012. ISSN 18770509. doi: 10.1016/j.procs.2012.04.033. Andr es Cencerrado, Ana Cort es, and Tom as Margalef. Applying probability theory for the quality as- sessment of a wild re spread prediction framework based on genetic algorithms. The Scienti c World Journal, 2013, 2013. ISSN 1537744X. doi: 10.1155/2013/728414. Andr es Cencerrado, Ana Cort es, and Tom as Margalef. Response time assessment in forest re spread simulation: An integrated methodology for ecient exploitation of available prediction time. Environ- mental Modelling & Software, 54:153{164, apr 2014. ISSN 1364-8152. doi: 10.1016/J.ENVSOFT.2014. 01.008. URL https://www-sciencedirect-com.login.ezproxy.library.ualberta.ca/science/ article/pii/S1364815214000176. Supriyo Chakraborty, Richard Tomsett, Ramya Raghavendra, Daniel Harborne, Moustafa Alzantot, Federico Cerutti, Mani Srivastava, Alun Preece, Simon Julier, Raghuveer M Rao, and Others. In- terpretability of deep learning models: a survey of results. In 2017 IEEE SmartWorld, Ubiq- uitous Intelligence & Computing, Advanced & Trusted Computed, Scalable Computing & Commu- nications, Cloud & Big Data Computing, Internet of People and Smart City Innovation (Smart- World/SCALCOM/UIC/ATC/CBDCom/IOP/SCI), pages 1{6. IEEE, 2017. Fs Chapin, T.N. Hollingsworth, and Re Hewitt. Fire e ects on seedling establishment success across treeline: implications for future tree migration and ammability in a changing climate. 2014. URL http://digitalcommons.unl.edu/jfspresearch/82/. Feng Chen, Yongsheng Du, Shukui Niu, and Jinlong Zhao. Modeling Forest Lightning Fire Occurrence in the Daxinganling Mountains of Northeastern China with MAXENT. Forests, 6(12):1422{1438, apr 2015. ISSN 1999-4907. doi: 10.3390/f6051422. URL http://www.mdpi.com/1999-4907/6/5/1422. Tao Cheng and Jiaqiu Wang. Integrated Spatio-temporal Data Mining for Forest Fire Prediction. Trans- actions in GIS, 12(5):591{611, sep 2008. ISSN 13611682. doi: 10.1111/j.1467-9671.2008.01117.x. URL http://doi.wiley.com/10.1111/j.1467-9671.2008.01117.x. Khaled Chetehouna, Eddy El Tabach, Loubna Bouazaoui, and Nicolas Gascoin. Predicting the ame characteristics and rate of spread in res propagating in a bed of Pinus pinaster using Arti cial Neu- ral Networks. Process Safety and Environmental Protection, 98:50{56, nov 2015. ISSN 0957-5820. doi: 10.1016/J.PSEP.2015.06.010. URL https://www.sciencedirect.com/science/article/pii/ S0957582015001111. Tatenda T Chingono and C Mbohwa. Fire Hazard Modelling in Southern Africa. In Proceedings of the World Congress on Engineering and Computer Science, San Francisco, 2015. URL http://www.iaeng. org/publication/WCECS2015/WCECS2015{_}pp514-519.pdf. 41 G. Chirici, R. Scotti, A. Montaghi, A. Barbati, R. Cartisano, G. Lopez, M. Marchetti, R. E. Mcroberts, H. Olsson, and P. Corona. Stochastic gradient boosting classi cation trees for forest fuel types mapping through airborne laser scanning and IRS LISS-III imagery. International Journal of Applied Earth Observation and Geoinformation, 25(1):87{97, 2013. ISSN 15698432. doi: 10.1016/j.jag.2013.04.006. Emilio Chuvieco, F. Javier Salas, Luis Carvacho, and Francisco Rodr guez-Silva. Integrated re risk mapping. In Remote Sensing of Large Wild res, pages 61{100. Springer Berlin Heidelberg, Berlin, Heidelberg, 1999. doi: 10.1007/978-3-642-60164-4 5. URL http://link.springer.com/10.1007/ 978-3-642-60164-4{_}5. Hamish Clarke, Rebecca Gibson, Brett Cirulis, Ross A Bradstock, and Trent D Penman. Developing and testing models of the drivers of anthropogenic and lightning-caused wild re ignitions in south- eastern Australia. Journal of Environmental Management, 235:34{41, apr 2019. ISSN 0301-4797. doi: 10.1016/J.JENVMAN.2019.01.055. URL https://www-sciencedirect-com.login.ezproxy. library.ualberta.ca/science/article/pii/S0301479719300568. J. Coen. Some Requirements for Simulating Wildland Fire Behavior Using Insight from Coupled Weather|Wildland Fire Models. Fire, 1(1):6, feb 2018. ISSN 2571-6255. doi: 10.3390/ re1010006. URL http://www.mdpi.com/2571-6255/1/1/6. Shane R. Coeld, Casey A. Gra , Yang Chen, Padhraic Smyth, E Foufoula-Georgiou, and James T. Randerson. Machine learning to predict nal re size at the time of ignition. International Journal of Wildland Fire, 28(11):861, 2019. ISSN 1049-8001. doi: 10.1071/WF19023. URL http://www.publish. csiro.au/?paper=WF19023. Judah Cohen, Dim Coumou, Jessica Hwang, Lester Mackey, Paulo Orenstein, Sonja Totz, and Eli Tziper- man. S2S reboot: An argument for greater inclusion of machine learning in subseasonal to seasonal forecasts. Wiley Interdisciplinary Reviews: Climate Change, 10(2), mar 2019. ISSN 1757-7780. doi: 10.1002/wcc.567. URL https://onlinelibrary.wiley.com/doi/abs/10.1002/wcc.567. L. Collins, P. Grioen, G. Newell, and A. Mellor. The utility of Random Forests for wild re severity mapping. Remote Sensing of Environment, 216:374{384, oct 2018. ISSN 0034-4257. doi: 10.1016/J.RSE. 2018.07.005. URL https://www.sciencedirect.com/science/article/pii/S0034425718303328. Sean C.P. Coogan, Fran cois Nicolas Robinne, Piyush Jain, and Mike D. Flannigan. Scientists' warning on wild re | a canadian perspective. Canadian Journal of Forest Research, 49(9):1015{1023, 2019. ISSN 12086037. doi: 10.1139/cjfr-2019-0094. Michelle Coppoletta, Kyle E. Merriam, and Brandon M. Collins. Post- re vegetation and fuel development in uences re severity patterns in reburns. Ecological Applications, 26(3):686{699, apr 2016. ISSN 1051-0761. doi: 10.1890/15-0225. A. Cordoba, R. Vilar, A. Lavrov, A. B. Utkin, and A. Fernandes. Multi-objective optimisation of lidar parameters for forest- re detection on the basis of a genetic algorithm. Optics and Laser Technology, 36 (5):393{400, 2004. ISSN 00303992. doi: 10.1016/j.optlastec.2003.10.010. Paulo Cortez and An bal de Jesus Raimundo Morais. A data mining approach to predict forest res using meteorological data. 2007. URL https://repositorium.sdum.uminho.pt/handle/1822/8039. Sergi Costafreda-Aumedes, Adrian Cardil, Domingo M. Molina, Sarah N. Daniel, Robert Mavsar, and Cristina Vega-Garcia. Analysis of factors in uencing deployment of re suppression resources in Spain using arti cial neural networks. iForest, 9(Feb 2016):138{145, 2015. ISSN 19717458. doi: 10.3832/ ifor1329-008. 42 Michael A. Crimmins. Synoptic climatology of extreme re-weather conditions across the southwest United States. International Journal of Climatology, 26(8):1001{1016, jun 2006. ISSN 0899-8418. doi: 10.1002/ joc.1300. URL http://doi.wiley.com/10.1002/joc.1300. Morgan A. Crowley, Je rey A. Cardille, Joanne C. White, and Michael A. Wulder. Multi-sensor, multi- scale, Bayesian data synthesis for mapping within-year wild re progression. Remote Sensing Letters, 10(3):302{311, 2019. ISSN 2150-704X. doi: 10.1080/2150704X.2018.1536300. URL https://www. tandfonline.com/doi/full/10.1080/2150704X.2018.1536300. Thomas Curt, Laurent Borgniet, Thomas Ibanez, Vincent Moron, and Christelle H ely. Understanding re patterns and re drivers for setting a sustainable management policy of the New-Caledonian biodiversity hotspot. Forest Ecology and Management, 337:48{60, 2015. ISSN 03781127. doi: 10.1016/j.foreco.2014. 10.032. URL http://dx.doi.org/10.1016/j.foreco.2014.10.032. Thomas Curt, Thibaut Fr ejaville, and S ebastien Lahaye. Modelling the spatial patterns of ignition causes and re regime features in southern France: Implications for re prevention policy. International Journal of Wildland Fire, 25(7):785{796, 2016. ISSN 10498001. doi: 10.1071/WF15205. James Richard Davis, J R L Hoare, and P M Nanninga. Developing a re man- agement expert system for Kakadu National Park, Australia. Journal of Environmental Management, 1986. URL https://publications.csiro.au/rpr/pub?list=BRO{&}pid=procite: f221b911-2e97-4c9f-be9b-f1155bf48c24. J.R. Davis, P.M. Nanninga, J.R.L. Hoare, and A.J. Press. Transferring scienti c knowledge to natu- ral resource managers using arti cial intelligence concepts. Ecological Modelling, 46(1-2):73{89, jul 1989. ISSN 0304-3800. doi: 10.1016/0304-3800(89)90070-7. URL https://www.sciencedirect.com/ science/article/pii/0304380089900707. Raymond Davis, Zhiqiang Yang, Andrew Yost, Cole Belongie, and Warren Cohen. The normal re en- vironment|Modeling environmental suitability for large forest wild res using past, present, and fu- ture climate normals. Forest Ecology and Management, 390:173{186, apr 2017. ISSN 0378-1127. doi: 10.1016/J.FORECO.2017.01.027. URL https://www.sciencedirect.com/science/article/ pii/S0378112716309318. Antonella De Angelis, Carlo Ricotta, Marco Conedera, and Gianni Boris Pezzatti. Modelling the Meteoro- logical Forest Fire Niche in Heterogeneous Pyrologic Conditions. PLOS ONE, 10(2):e0116875, feb 2015. ISSN 1932-6203. doi: 10.1371/journal.pone.0116875. URL https://dx.plos.org/10.1371/journal. pone.0116875. P. P. de Bem, O. A. de Carvalho Ju nior, E. A. T. Matricardi, R. F. Guimar~ aes, and R. A. T. Gomes. Predicting wild re vulnerability using logistic regression and arti cial neural networks: a case study in Brazil's Federal District. International Journal of Wildland Fire, 28(1):35, feb 2018. ISSN 1049-8001. doi: 10.1071/wf18018. URL http://www.publish.csiro.au/?paper=WF18018. Marla Jose Perestrello De Vasconcelos, Sara Sllva, Margarlda Tome, Margarlda Alvim, Jose Mlguel, and Cardoso Perelra. Spatial Prediction of Fire Ignition Probabilities: Comparing Logistic Regression and Neural Networks. Photogrammetric Engineering & Remote Sensing, 67(1):73{81, 2001. Haifa Debouk, Ramon Riera-Tatch e, and Cristina Vega-Garc a. Assessing Post-Fire Regeneration in a Mediterranean Mixed Forest Using Lidar Data and Arti cial Neural Networks. Photogrammetric Engi- neering & Remote Sensing, 2013. ISSN 00991112. doi: 10.14358/PERS.79.12.1121. Rosario Delgado, Jos e-Luis Gonz alez, Andr es Sotoca, and Xavier-Andoni Tibau. Archetypes of Wild re Arsonists: An Approach by Using Bayesian Networks. In Forest Fire. InTech, may 43 2018. doi: 10.5772/intechopen.72615. URL http://www.intechopen.com/books/forest-fire/ archetypes-of-wildfire-arsonists-an-approach-by-using-bayesian-networks. M onica Denham and Karina Laneri. Using ecient parallelization in Graphic Processing Units to param- eterize stochastic re propagation models. Journal of Computational Science, 25:76{88, 2018. ISSN 18777503. doi: 10.1016/j.jocs.2018.02.007. URL https://doi.org/10.1016/j.jocs.2018.02.007. M onica Denham, Kerstin Wendt, Germ an Bianchini, Ana Cort es, and Tom as Margalef. Dynamic Data- Driven Genetic Algorithm for forest re spread prediction. Journal of Computational Science, 3(5): 398{404, 2012. ISSN 18777503. doi: 10.1016/j.jocs.2012.06.002. URL http://dx.doi.org/10.1016/j. jocs.2012.06.002. Peter J. Diggle, Raquel Menezes, and Ting-li Su. Geostatistical inference under preferential sam- pling. Journal of the Royal Statistical Society: Series C (Applied Statistics), 59(2):191{232, mar 2010. ISSN 00359254. doi: 10.1111/j.1467-9876.2009.00701.x. URL http://doi.wiley.com/10.1111/j. 1467-9876.2009.00701.x. Luca Antonio Dimuccio, Rui Ferreira, Lucio Cunha, and Antonio Campar de Almeida. Regional forest- re susceptibility analysis in central Portugal using a probabilistic ratings procedure and arti cial neural network weights assignment. International Journal of Wildland Fire, 20(6):776, oct 2011. ISSN 1049- 8001. doi: 10.1071/WF09083. URL http://www.publish.csiro.au/?paper=WF09083. T. L. Divya and M. N. Vijayalakshmi. Inference of Replanting in Forest Fire A ected Land Using Data Mining Technique. pages 121{129. Springer, New Delhi, 2016. doi: 10.1007/978-81-322-2734-2 13. URL http://link.springer.com/10.1007/978-81-322-2734-2{_}13. Wisdom M. Dlamini. A Bayesian belief network analysis of factors in uencing wild re occurrence in Swaziland. Environmental Modelling and Software, 25(2):199{208, 2010. ISSN 13648152. doi: 10.1016/ j.envsoft.2009.08.002. URL http://dx.doi.org/10.1016/j.envsoft.2009.08.002. Wisdom Mdumiseni Dlamini. Application of Bayesian networks for re risk mapping using GIS and remote sensing data. GeoJournal, 76(3):283{296, jun 2011. ISSN 0343-2521. doi: 10.1007/s10708-010-9362-x. URL http://link.springer.com/10.1007/s10708-010-9362-x. E. Dragozi, I. Gitas, D. Stavrakoudis, and J. Theocharis. Burned Area Mapping Using Support Vector Machines and the FuzCoC Feature Selection Method on VHR IKONOS Imagery. Remote Sensing, 6 (12):12005{12036, dec 2014. ISSN 2072-4292. doi: 10.3390/rs61212005. URL http://www.mdpi.com/ 2072-4292/6/12/12005. Andrea Duane, M riam Piqu e, Marc Castellnou, and Llu s Brotons. Predictive modelling of re occurrences from di erent re spread patterns in Mediterranean landscapes. International Journal of Wildland Fire, 24(3):407, jun 2015. ISSN 1049-8001. doi: 10.1071/WF14040. URL http://www.publish.csiro.au/ ?paper=WF14040. Ritaban Dutta, Jagannath Aryal, Aruneema Das, and Jamie B. Kirkpatrick. Deep cognitive imaging systems enable estimation of continental-scale re incidence from climate data. Scienti c Reports, 3, 2013. ISSN 20452322. doi: 10.1038/srep03188. Ritaban Dutta, Aruneema Das, and Jagannath Aryal. Big data integration shows Australian bush- re frequency is increasing signi cantly. Royal Society Open Science, 3(2):150241, feb 2016. ISSN 2054-5703. doi: 10.1098/rsos.150241. URL https://royalsocietypublishing.org/doi/10.1098/rsos.150241. Francis K. Dwomoh and Michael C. Wimberly. Fire regimes and their drivers in the Upper Guinean Region of West Africa. Remote Sensing, 9(11), 2017. ISSN 20724292. doi: 10.3390/rs9111117. 44 J.B. Theocharis E. Dragozi, I. Z. Gitas, D.G. Stavrakoudis. A Performance Evaluation Of Sup- port Vector Machines And The Nearest Neighbor Classi er In Classifying Image Objects For Burned Area Mapping. In Proceedings of the 8th International EARSeL FF-SIG Workshop, Stresa, Italy, 2011. URL https://publications.europa.eu/en/publication-detail/-/publication/ 5e90be79-efe8-43dd-8670-2518a43155f4/language-en. Hamid Ebrahimy, Aliakbar Rasuly, *, and Davoud Mokhtari. Development of a Web GIS Sys- tem Based on the MaxEnt Approach for Wild re Management: A Case Study of East Azer- baijan. ECOPERSIA, 5(3):1859{1873, 2017. URL https://pdfs.semanticscholar.org/46b2/ fd74419232dbe2dedccaaca40bab6dbf50b8.pdf. J. Elith, J. R. Leathwick, and T. Hastie. A working guide to boosted regression trees. Journal of Animal Ecology, 77(4):802{813, jul 2008. ISSN 0021-8790. doi: 10.1111/j.1365-2656.2008.01390.x. URL http: //doi.wiley.com/10.1111/j.1365-2656.2008.01390.x. Jane Elith, Steven J. Phillips, Trevor Hastie, Miroslav Dud k, Yung En Chee, and Colin J. Yates. A statistical explanation of MaxEnt for ecologists. Diversity and Distributions, 17(1):43{57, jan 2011. ISSN 13669516. doi: 10.1111/j.1472-4642.2010.00725.x. URL http://doi.wiley.com/10.1111/j. 1472-4642.2010.00725.x. Thomas A. Fairman, Lauren T. Bennett, Shauna Tupper, and Craig R. Nitschke. Frequent wild res erode tree persistence and alter stand structure and initial composition of a re-tolerant sub-alpine forest. Journal of Vegetation Science, 28(6):1151{1165, 2017. ISSN 16541103. doi: 10.1111/jvs.12575. Lei Fang, Jian Yang, Megan White, and Zhihua Liu. Predicting potential re severity using vegetation, topography and surface moisture availability in a Eurasian boreal forest landscape. Forests, 9(3):1{26, 2018. ISSN 19994907. doi: 10.3390/f9030130. Armando M. Fernandes, Andrei B. Utkin, Alexander V. Lavrov, and Rui M. Vilar. Neural Network Based Recognition of Smoke Signatures from Lidar Signals. Neural Processing Letters, 19(3):175{189, jun 2004a. ISSN 1370-4621. doi: 10.1023/B:NEPL.0000035598.19042.42. URL http://link.springer. com/10.1023/B:NEPL.0000035598.19042.42. Armando M. Fernandes, Andrei B. Utkin, Alexander V. Lavrov, and Rui M. Vilar. Development of neural network committee machines for automatic forest re detection using lidar. Pattern Recognition, 37(10): 2039{2047, oct 2004b. ISSN 00313203. doi: 10.1016/j.patcog.2004.04.002. URL http://linkinghub. elsevier.com/retrieve/pii/S0031320304001360. Paulo M. Fernandes, Tiago Monteiro-Henriques, Nuno Guiomar, Carlos Loureiro, and Ana M.G. Barros. Bottom-Up Variables Govern Large-Fire Size in Portugal. Ecosystems, 19(8):1362{1375, 2016. ISSN 14350629. doi: 10.1007/s10021-016-0010-2. Alfonso Fernandez-Manso, Carmen Quintano, and Dar A. Roberts. Burn severity analysis in Mediterranean forests using maximum entropy model trained with EO-1 Hyperion and LiDAR data. ISPRS Journal of Photogrammetry and Remote Sensing, 155(July):102{118, 2019. ISSN 09242716. doi: 10.1016/j.isprsjprs. 2019.07.003. URL https://doi.org/10.1016/j.isprsjprs.2019.07.003. Mark A Finney. FARSITE: Fire Area Simulator-Model Development and Evaluation. Technical report, Rocky Mountain Research Station, 2004. URL http://www.farsite.org. Mark A. Finney. The challenge of quantitative risk analysis for wildland re. In Forest Ecology and Management, volume 211, pages 97{108, jun 2005. doi: 10.1016/j.foreco.2005.02.010. 45 Marisa G. Fonseca, Luiz Eduardo O. C. Arag~ ao, Andr e Lima, Yosio E. Shimabukuro, Egidio Arai, and Liana O. Anderson. Modelling re probability in the Brazilian Amazon using the maximum entropy method. International Journal of Wildland Fire, 25(9):955, sep 2016. doi: 10.1071/WF15216. URL http://www.publish.csiro.au/?paper=WF15216. Matthias Forkel, Wouter Dorigo, Gitta Lasslop, Irene Teubner, Emilio Chuvieco, and Kirsten Thonicke. A data-driven approach to identify controls on global re activity from satellite and climate observations (SOFIA V1). Geoscienti c Model Development, 10(12):4443{4476, dec 2017. ISSN 1991-9603. doi: 10.5194/gmd-10-4443-2017. URL https://www.geosci-model-dev.net/10/4443/2017/. Matthias Forkel, Niels Andela, Sandy P Harrison, Gitta Lasslop, Margreet Van Marle, Emilio Chuvieco, Wouter Dorigo, Matthew Forrest, Stijn Hantson, Angelika Heil, Fang Li, Joe Melton, Stephen Sitch, Chao Yue, and Almut Arneth. Emergent relationships with respect to burned area in global satellite observations and re-enabled vegetation models. Biogeosciences, 16(1):57{76, 2019. ISSN 17264189. doi: 10.5194/bg-16-57-2019. Yoav Freund and Robert E Shapire. A decision-theoretic generalization of on-line learning and an ap- plication to boosting. In Computational Learning Theory: Eurocolt '95, pages 23{37. Springer-Verlag, Jerome H. Friedman. Greedy function approximation: A gradient boosting machine. Ann. Statist., 29(5): 1189{1232, 10 2001. doi: 10.1214/aos/1013203451. URL https://doi.org/10.1214/aos/1013203451. Sigfredo Fuentes, Eden Jane Tongson, Roberta De Bei, Claudia Gonzalez Viejo, Renata Ristic, Stephen Tyerman, and Kerry Wilkinson. Non-Invasive Tools to Detect Smoke Contamination in Grapevine Canopies, Berries and Wine: A Remote Sensing and Machine Learning Modeling Approach. Sensors, 19 (15):3335, jul 2019. ISSN 1424-8220. doi: 10.3390/s19153335. URL https://www.mdpi.com/1424-8220/ 19/15/3335. Mariano Garc a, David Riano, ~ Emilio Chuvieco, Javier Salas, and F. Mark Danson. Multispectral and LiDAR data fusion for fuel type mapping using Support Vector Machine and decision rules. Remote Sensing of Environment, 115(6):1369{1379, 2011. ISSN 00344257. doi: 10.1016/j.rse.2011.01.017. URL http://dx.doi.org/10.1016/j.rse.2011.01.017. Paula Garc a-Llamas, Susana Su arez-Seoane, Angela Taboada, Alfonso Fern andez-Manso, Carmen Quin- tano, V ctor Fern andez-Garc a, Jos e Manuel Fern andez-Guisuraga, Elena Marcos, and Leonor Calvo. Environmental drivers of re severity in extreme re events that a ect Mediterranean pine forest ecosystems. Forest Ecology and Management, 433(October 2018):24{32, 2019. ISSN 03781127. doi: 10.1016/j.foreco.2018.10.051. URL https://doi.org/10.1016/j.foreco.2018.10.051. Stuart Geman, Elie Bienenstock, and Ren e Doursat. Neural Networks and the Bias/Variance Dilemma. Neural Computation, 4(1):1{58, jan 1992. ISSN 0899-7667. doi: 10.1162/neco.1992.4.1.1. Andre Gensler, Janosch Henze, Bernhard Sick, and Nils Raabe. Deep Learning for solar power forecasting - An approach using AutoEncoder and LSTM Neural Networks. In 2016 IEEE International Conference on Systems, Man, and Cybernetics, SMC 2016 - Conference Proceedings, pages 2858{2865. Institute of Electrical and Electronics Engineers Inc., feb 2017. ISBN 9781509018970. doi: 10.1109/SMC.2016. Omid Ghorbanzadeh, Thomas Blaschke, Khalil Gholamnia, and Jagannath Aryal. Forest Fire Susceptibility and Risk Mapping Using Social/Infrastructural Vulnerability and Environmental Variables. Fire, 2(3): 50, sep 2019a. ISSN 2571-6255. doi: 10.3390/ re2030050. URL https://www.mdpi.com/2571-6255/2/ 3/50. 46 Omid Ghorbanzadeh, Khalil Valizadeh Kamran, Thomas Blaschke, Jagannath Aryal, Amin Naboureh, Jamshid Einali, and Jinhu Bian. Spatial Prediction of Wild re Susceptibility Using Field Survey GPS Data and Machine Learning Approaches. Fire, 2(3):43, jul 2019b. ISSN 2571-6255. doi: 10.3390/ re2030043. URL https://www.mdpi.com/2571-6255/2/3/43. Louis Giglio, Luigi Boschetti, David P. Roy, Michael L. Humber, and Christopher O. Justice. The Collection 6 MODIS burned area mapping algorithm and product. Remote Sensing of Environment, 217:72{85, nov 2018. ISSN 00344257. doi: 10.1016/j.rse.2018.08.005. Ljubomir Gigovi c, Hamid Reza Pourghasemi, Sini sa Drobnjak, and Shibiao Bai. Testing a New Ensemble Model Based on SVM and Random Forest in Forest Fire Susceptibility Assessment and Its Mapping in Serbia's Tara National Park. Forests, 10(5):408, may 2019. ISSN 1999-4907. doi: 10.3390/f10050408. URL https://www.mdpi.com/1999-4907/10/5/408. Y.J. Goldarag, Ali Mohammadzadeh, and A. S. Ardakani. Fire Risk Assessment Using Neural Network and Logistic Regression. Journal of the Indian Society of Remote Sensing, 44(6):885{894, 2016. ISSN 09743006. doi: 10.1007/s12524-016-0557-6. URL http://dx.doi.org/10.1007/s12524-016-0557-6. Carla Gomes. Computational Sustainability: Computational Methods for a Sustainable Environment, Economy, and Society. The Bridge, National Academy of Engineering, 39(4), 2009. Israel G omez and M. Pilar Mart n. Prototyping an arti cial neural network for burned area mapping on a regional scale in Mediterranean areas using MODIS images. International Journal of Applied Earth Observation and Geoinformation, 13(5):741{752, 2011. ISSN 15698432. doi: 10.1016/j.jag.2011.05.002. Noel Gorelick, Matt Hancher, Mike Dixon, Simon Ilyushchenko, David Thau, and Rebecca Moore. Google Earth Engine: Planetary-scale geospatial analysis for everyone. Remote Sensing of Environment, 2017. doi: 10.1016/j.rse.2017.06.031. URL https://doi.org/10.1016/j.rse.2017.06.031. Futao Guo, Guangyu Wang, Zhangwen Su, Huiling Liang, Wenhui Wang, Fangfang Lin, and Aiqin Liu. What drives forest re in Fujian, China? Evidence from logistic regression and Random Forests. In- ternational Journal of Wildland Fire, 25(5):505, may 2016a. ISSN 1049-8001. doi: 10.1071/WF15121. URL http://www.publish.csiro.au/?paper=WF15121. Futao Guo, Lianjun Zhang, Sen Jin, Mulualem Tigabu, Zhangwen Su, Wenhui Wang, Futao Guo, Lianjun Zhang, Sen Jin, Mulualem Tigabu, Zhangwen Su, and Wenhui Wang. Modeling Anthropogenic Fire Occurrence in the Boreal Forest of China Using Logistic Regression and Random Forests. Forests, 7(12): 250, oct 2016b. ISSN 1999-4907. doi: 10.3390/f7110250. URL http://www.mdpi.com/1999-4907/7/ 11/250. Dale Hamilton, Barry Myers, and Jonathan Branham. Evaluation Of Texture As An Input Of Spatial Context For Machine Learning Mapping Of Wildland Fire E ects. An International Journal (SIPIJ), 8 (5), 2017. doi: 10.5121/sipij.2017.8501. Jie Han, Zehao Shen, Lingxiao Ying, Guixiang Li, and Anping Chen. Early post- re regeneration of a re-prone subtropical mixed Yunnan pine forest in Southwest China: E ects of pre- re vegetation, re severity and topographic factors. Forest Ecology and Management, 356(2015):31{40, 2015. ISSN 03781127. doi: 10.1016/j.foreco.2015.06.016. URL http://dx.doi.org/10.1016/j.foreco.2015.06. Lucas Harris and Alan H. Taylor. Previous burns and topography limit and reinforce re severity in a large wild re. Ecosphere, 8(11), 2017. ISSN 21508925. doi: 10.1002/ecs2.2019. Trevor Hastie, Jerome Friedman, and Robert Tibshirani. The Elements of Statistical Learning: Data Mining, Inference, and Prediction. Springer, New York, NY, 2009. URL https://doi.org/10.1007/ 978-0-387-21606-5. 47 Todd J. Hawbaker, Melanie K. Vanderhoof, Yen Ju Beal, Joshua D. Takacs, Gail L. Schmidt, Je T. Falgout, Brad Williams, Nicole M. Fairaux, Megan K. Caldwell, Joshua J. Picotte, Stephen M. Howard, Susan Stitt, and John L. Dwyer. Mapping burned areas using dense time-series of Landsat data. Remote Sensing of Environment, 198:504{522, sep 2017. ISSN 00344257. doi: 10.1016/j.rse.2017.06.027. Marti A. Hearst, Susan T Dumais, Edgar Osuna, John Platt, and Bernhard Scholkopf. Support vector machines. IEEE Intelligent Systems and their applications, 13(4):18{28, 1998. Robert Hecht-Nielsen. Theory of the backpropagation neural network. In Neural networks for perception, pages 65{93. Elsevier, 1992. Risto K. Heikkinen, Mathieu Marmion, and Miska Luoto. Does the interpolation accuracy of species distribution models come at the expense of transferability? Ecography, 35(3):276{288, mar 2012. ISSN 09067590. doi: 10.1111/j.1600-0587.2011.06999.x. URL http://doi.wiley.com/10.1111/j. 1600-0587.2011.06999.x. Txomin Hermosilla, Michael A. Wulder, Joanne C. White, Nicholas C. Coops, and Geordie W. Hobart. Regional detection, characterization, and attribution of annual forest change from 1984 to 2012 using Landsat-derived time-series metrics. Remote Sensing of Environment, 170:121{132, 2015. ISSN 00344257. doi: 10.1016/j.rse.2015.09.004. URL http://dx.doi.org/10.1016/j.rse.2015.09.004. Sepp Hochreiter and Jurgen  Schmidhuber. Long short-term memory. Neural computation, 9(8):1735{1780, Jonathan L. Hodges and Brian Y. Lattimer. Wildland Fire Spread Modeling Using Convolutional Neural Networks. Fire Technology, nov 2019. ISSN 15728099. doi: 10.1007/s10694-019-00846-4. C. M. Ho man, J. Can eld, R. R. Linn, W. Mell, C. H. Sieg, F. Pimont, and J. Ziegler. Evaluating Crown Fire Rate of Spread Predictions from Physics-Based Models. Fire Technology, 52(1):221{237, jan 2016. ISSN 15728099. doi: 10.1007/s10694-015-0500-3. Zachary A. Holden, Penelope Morgan, and Je rey S. Evans. A predictive model of burn severity based on 20-year satellite-inferred burn severity data in a large southwestern US wilderness area. Forest Ecology and Management, 258(11):2399{2406, 2009. ISSN 03781127. doi: 10.1016/j.foreco.2009.08.017. Baisravan Homchaudhuri, Sheng Zhao, Kelly Cohen, and Manish Kumar. Generation of optimal re-line for ghting wildland res using genetic algorithms. Proceedings of the ASME Dynamic Systems and Control Conference 2009, DSCC2009, (PART A):111{118, 2010. doi: 10.1115/DSCC2009-2707. Haoyuan Hong, Paraskevas Tsangaratos, Ioanna Ilia, Junzhi Liu, A-Xing Zhu, and Chong Xu. Applying genetic algorithms to set the optimal combination of forest re related variables and model forest re susceptibility based on data mining models. The case of Dayu County, China. Science of The Total Environment, 630:1044{1056, jul 2018. ISSN 0048-9697. doi: 10.1016/J.SCITOTENV.2018.02.278. URL https://www.sciencedirect.com/science/article/pii/S004896971830679X. F M Anim Hossain, Youmin Zhang, Chi Yuan, and Chun-Yi Su. Wild re Flame and Smoke Detection Using Static Image Features and Arti cial Neural Network. In 2019 1st International Conference on Industrial Arti cial Intelligence (IAI), pages 1{6. Institute of Electrical and Electronics Engineers (IEEE), oct 2019. doi: 10.1109/iciai.2019.8850811. Bronwyn A. Hradsky, Trent D. Penman, Dan Ababei, Anca Hanea, Euan G. Ritchie, Alan York, and Julian Di Stefano. Bayesian networks elucidate interactions between re and other drivers of terrestrial fauna distributions. Ecosphere, 8(8):e01926, aug 2017. ISSN 21508925. doi: 10.1002/ecs2.1926. URL http://doi.wiley.com/10.1002/ecs2.1926. 48 Carolynne Hultquist, Gang Chen, and Kaiguang Zhao. A comparison of Gaussian process regression, random forests and support vector regression for burn severity assessment in diseased forests. Re- mote Sensing Letters, 5(8):723{732, aug 2014. ISSN 2150-704X. doi: 10.1080/2150704X.2014.963733. URL http://dx.doi.org/10.1080/2150704X.2014.963733http://www.tandfonline.com/doi/abs/ 10.1080/2150704X.2014.963733. L.S. Iliadis. A decision support system applying an integrated fuzzy model for long-term forest re risk estimation. Environmental Modelling & Software, 20(5):613{621, may 2005. ISSN 1364-8152. doi: 10.1016/J.ENVSOFT.2004.03.006. URL https://www.sciencedirect.com/science/article/ pii/S1364815204000933. Abolfazl Jaafari. Factors In uencing Regional-Scale Wild re Probability in Iran: An Applica- tion of Random Forest and Support Vector Machine. Spatial Modeling in GIS and R for Earth and Environmental Sciences, pages 607{619, jan 2019. doi: 10.1016/B978-0-12-815226-3. 00028-4. URL https://www-sciencedirect-com.login.ezproxy.library.ualberta.ca/science/ article/pii/B9780128152263000284. Abolfazl Jaafari, Eric K. Zenner, and Binh Thai Pham. Wild re spatial pattern analysis in the Za- gros Mountains, Iran: A comparative study of decision tree based classi ers. Ecological Informat- ics, 43:200{211, jan 2018. ISSN 1574-9541. doi: 10.1016/J.ECOINF.2017.12.006. URL https: //www.sciencedirect.com/science/article/pii/S157495411730167X. Abolfazl Jaafari, Eric K. Zenner, Mahdi Panahi, and Himan Shahabi. Hybrid arti cial intelligence mod- els based on a neuro-fuzzy system and metaheuristic optimization algorithms for spatial prediction of wild re probability. Agricultural and Forest Meteorology, 266-267:198{207, mar 2019. ISSN 01681923. doi: 10.1016/j.agrformet.2018.12.015. URL https://www.sciencedirect.com/science/article/pii/ S0168192318304088. Jacek Jakubowski, Maciej Solarczyk, and Micha l Wi snios. Smoke detection in a digital image with the use of convolutional network. page 14. SPIE-Intl Soc Optical Eng, mar 2019. ISBN 9781510627857. doi: 10.1117/12.2524560. J. . R. Jang. An s: adaptive-network-based fuzzy inference system. IEEE Transactions on Systems, Man, and Cybernetics, 23(3):665{685, May 1993. doi: 10.1109/21.256541. Maria Jo~ ao Sousa, Alexandra Moutinho, and Miguel Almeida. Wild re detection using transfer learning on augmented datasets. Expert Systems with Applications, page 112975, sep 2019. ISSN 09574174. doi: 10.1016/j.eswa.2019.112975. Torres Jo~ ao, Gon calves Jo~ ao, Marcos Bruno, and Honrado Jo~ ao. Indicator-based assessment of post- re re- covery dynamics using satellite NDVI time-series. Ecological Indicators, 89(January):199{212, 2018. ISSN 1470160X. doi: 10.1016/j.ecolind.2018.02.008. URL https://doi.org/10.1016/j.ecolind.2018.02. Jill F. Johnstone, Teresa N. Hollingsworth, F. Stuart Chapin, and Michelle C. Mack. Changes in re regime break the legacy lock on successional trajectories in Alaskan boreal forest. Global Change Biology, 16 (4):1281{1295, 2010. ISSN 13541013. doi: 10.1111/j.1365-2486.2009.02051.x. Kyle D. Julian and Mykel J. Kochenderfer. Autonomous distributed wild re surveillance using deep reinforcement learning. In AIAA Guidance, Navigation, and Control Conference, 2018, number 210039. American Institute of Aeronautics and Astronautics Inc, AIAA, jan 2018a. ISBN 9781624105265. doi: 10.2514/6.2018-1589. 49 Kyle D. Julian and Mykel J. Kochenderfer. Distributed Wild re Surveillance with Autonomous Aircraft using Deep Reinforcement Learning. Journal of Guidance, Control, and Dynamics, 42(8):1768{1778, oct 2018b. URL http://arxiv.org/abs/1810.04244. Martin Jung, Susanne Tautenhahn, Christian Wirth, and Jens Kattge. Estimating Basal Area of Spruce and Fir in Post- re Residual Stands in Central Siberia Using Quickbird, Feature Selection, and Random Forests. Procedia Computer Science, 18:2386{2395, jan 2013. ISSN 1877-0509. doi: 10.1016/J.PROCS. 2013.05.410. URL https://www.sciencedirect.com/science/article/pii/S187705091300553X. M. Njoki Kahiu and N. P. Hanan. Fire in sub-Saharan Africa: The fuel, cure and connectivity hypothesis. Global Ecology and Biogeography, 27(8):946{957, aug 2018. ISSN 1466822X. doi: 10.1111/geb.12753. URL http://doi.wiley.com/10.1111/geb.12753. Van R. Kane, C. Alina Cansler, Nicholas A. Povak, Jonathan T. Kane, Robert J. McGaughey, James A. Lutz, Derek J. Churchill, and Malcolm P. North. Mixed severity re e ects within the Rim re: Relative importance of local climate, re weather, topography, and forest structure. Forest Ecology and Management, 358:62{79, dec 2015. ISSN 0378-1127. doi: 10.1016/J.FORECO.2015.09.001. URL https://www.sciencedirect.com/science/article/pii/S0378112715004697. Anuj Karpatne, Imme Ebert-Upho , Sai Ravela, Hassan Ali Babaie, and Vipin Kumar. Machine Learning for the Geosciences: Challenges and Opportunities. nov 2017. URL http://arxiv.org/abs/1711. Robert E. Keane, Geo rey J. Cary, Ian D. Davies, Michael D. Flannigan, Robert H. Gardner, Sandra Lavorel, James M. Lenihan, Chao Li, and T. Scott Rupp. A classi cation of landscape re succession models: Spatial simulations of re and vegetation dynamics. Ecological Modelling, 179(1-2):3{27, nov 2004. ISSN 03043800. doi: 10.1016/j.ecolmodel.2004.03.015. Nima Khakzad. Modeling wild re spread in wildland-industrial interfaces using dynamic Bayesian network. Reliability Engineering & System Safety, 189:165{176, sep 2019. ISSN 0951-8320. doi: 10.1016/J.RESS. 2019.04.006. URL https://www.sciencedirect.com/science/article/pii/S0951832018313887. Sea Jin Kim, Chul-Hee Lim, Gang Sun Kim, Jongyeol Lee, Tobias Geiger, Omid Rahmati, Yowhan Son, and Woo-Kyun Lee. Multi-Temporal Analysis of Forest Fire Probability Using Socio-Economic and Environmental Variables. Remote Sensing, 11(1):86, jan 2019. ISSN 2072-4292. doi: 10.3390/rs11010086. URL https://www.mdpi.com/2072-4292/11/1/86. Seongchan Kim, Seungkyun Hong, Minsu Joh, and Sa-kwang Song. DeepRain: ConvLSTM Network for Precipitation Prediction using Multichannel Radar Data. nov 2017. URL http://arxiv.org/abs/ 1711.02316. ByoungChul Ko, Kwang-Ho Cheong, and Jae-Yeal Nam. Early re detection algorithm based on irreg- ular patterns of ames and hierarchical Bayesian Networks. Fire Safety Journal, 45(4):262{270, jun 2010. ISSN 0379-7112. doi: 10.1016/J.FIRESAF.2010.04.001. URL https://www.sciencedirect.com/ science/article/pii/S0379711210000378. Alexandru Korotcov, Valery Tkachenko, Daniel P. Russo, and Sean Ekins. Comparison of Deep Learning with Multiple Machine Learning Methods and Metrics Using Diverse Drug Discovery Data Sets. Molec- ular Pharmaceutics, 14(12):4462{4475, dec 2017. ISSN 15438392. doi: 10.1021/acs.molpharmaceut. 7b00578. Peter Kourtz. Arti cial intelligence: a new tool for forest management. Canadian Journal of For- est Research, 20(4):428{437, apr 1990. ISSN 0045-5067. doi: 10.1139/x90-060. URL http://www. nrcresearchpress.com/doi/10.1139/x90-060. 50 P.H. Kourtz. Arti cial intelligence applications in the next generation Canadian forest re control system, 1993. URL https://cfs.nrcan.gc.ca/publications?id=10775. V. I. Kozik, E. S. Nezhevenko, and A. S. Feoktistov. Adaptive prediction of forest re behavior on the basis of recurrent neural networks. Optoelectronics, Instrumentation and Data Processing, 49(3):250{259, may 2013. ISSN 8756-6990. doi: 10.3103/S8756699013030060. URL http://link.springer.com/10.3103/ S8756699013030060. V. I. Kozik, E. S. Nezhevenko, and A. S. Feoktistov. Studying the method of adaptive prediction of forest re evolution on the basis of recurrent neural networks. Optoelectronics, Instrumentation and Data Processing, 50(4):395{401, jul 2014. ISSN 8756-6990. doi: 10.3103/S8756699014040116. URL http://link.springer.com/10.3103/S8756699014040116. Max Kuhn and Kjell Johnson. Applied predictive modeling. Springer New York, jan 2013. ISBN 9781461468493. doi: 10.1007/978-1-4614-6849-3. Jan Kuka cka, Vladimir Golkov, and Daniel Cremers. Regularization for Deep Learning: A Taxonomy. oct 2017. URL http://arxiv.org/abs/1710.10686. Nataliia Kussul, Mykola Lavreniuk, Sergii Skakun, and Andrii Shelestov. Deep Learning Classi cation of Land Cover and Crop Types Using Remote Sensing Data. IEEE Geoscience and Remote Sensing Letters, 14(5):778{782, may 2017. ISSN 1545-598X. doi: 10.1109/LGRS.2017.2681128. URL http: //ieeexplore.ieee.org/document/7891032/. Ryan Lagerquist, Mike D. Flannigan, Xianli Wang, and Ginny A. Marshall. Automated prediction of extreme re weather from synoptic patterns in northern Alberta, Canada. Canadian Journal of Forest Research, 47(9):1175{1183, sep 2017. ISSN 0045-5067. doi: 10.1139/cjfr-2017-0063. URL http://www. nrcresearchpress.com/doi/10.1139/cjfr-2017-0063. Zachary Langford, Jitendra Kumar, and Forrest Ho man. Wild re mapping in interior alaska using deep neural networks on imbalanced datasets. In IEEE International Conference on Data Mining Workshops, ICDMW, volume 2018-Novem, pages 770{778. IEEE Computer Society, feb 2019. ISBN 9781538692882. doi: 10.1109/ICDMW.2018.00116. David J. Lary, Amir H. Alavi, Amir H. Gandomi, and Annette L. Walker. Machine learning in geosciences and remote sensing. Geoscience Frontiers, 7(1):3{10, jan 2016. ISSN 16749871. doi: 10.1016/j.gsf.2015. 07.003. URL https://www.sciencedirect.com/science/article/pii/S1674987115000821. Don J Latham. Arti cial Intelligence Applications To Fire Management. In Proceedings Of The Symposium On Wildland Fire, South Lake Tahoe, 1987. Christopher J. Lauer, Claire A. Montgomery, and Thomas G. Dietterich. Spatial interactions and opti- mal forest management on a re-threatened landscape. Forest Policy and Economics, 83:107{120, oct 2017. ISSN 1389-9341. doi: 10.1016/J.FORPOL.2017.07.006. URL https://www.sciencedirect.com/ science/article/pii/S1389934116304749. Yann LeCun, Yoshua Bengio, and Geo rey Hinton. Deep learning. nature, 521(7553):436, 2015. Michael Leuenberger, Joana Parente, Marj Tonini, M ario Gonzalez Pereira, and Mikhail Kanevski. Wild re susceptibility mapping: Deterministic vs. stochastic approaches. Environmental Modelling & Software, 101:194{203, mar 2018. ISSN 1364-8152. doi: 10.1016/J.ENVSOFT.2017.12.019. URL https://www. sciencedirect.com/science/article/pii/S1364815217303316. 51 Danielle Levac, Heather Colquhoun, and Kelly K O'Brien. Scoping studies: advanc- ing the methodology. Implementation science : IS, 5:69, sep 2010. ISSN 1748-5908. doi: 10.1186/1748-5908-5-69. URL http://www.ncbi.nlm.nih.gov/pubmed/20854677http://www. pubmedcentral.nih.gov/articlerender.fcgi?artid=PMC2954944. Berangere A. Leys, Julie L. Commerford, and Kendra K. McLauchlan. Reconstructing grassland re history using sedimentary charcoal: Considering count, size and shape. PLOS ONE, 12(4):e0176445, apr 2017. ISSN 1932-6203. doi: 10.1371/journal.pone.0176445. URL https://dx.plos.org/10.1371/journal. pone.0176445. Hanchao Li, Xiang Fei, and Chaobo He. Study on Most Important Factor and Most Vulnerable Location for a Forest Fire Case Using Various Machine Learning Techniques. 2018 Sixth International Conference on Advanced Cloud and Big Data (CBD), pages 298{303, 2018a. doi: 10.1109/CBD.2018.00060. URL https://ieeexplore.ieee.org/document/8530856/. Jin Li, Andrew D. Heap, Anna Potter, and James J. Daniell. Application of machine learning methods to spatial interpolation of environmental variables. Environmental Modelling and Software, 26(12):1647{ 1659, dec 2011. ISSN 13648152. doi: 10.1016/j.envsoft.2011.07.004. Li-Ming Li, Wei-Guo Song, Jian Ma, and Kohyu Satoh. Arti cial neural network approach for modeling the impact of population density and weather parameters on forest re risk. International Journal of Wildland Fire, 18(6):640, oct 2009. ISSN 1049-8001. doi: 10.1071/WF07136. URL http://www. publish.csiro.au/?paper=WF07136. Shufeng Li, Alice C. Hughes, Tao Su, Julie Lebreton Anberr ee, Alexei A. Oskolski, Mei Sun, David K. Ferguson, and Zhekun Zhou. Fire dynamics under monsoonal climate in Yunnan, SW China: past, present and future. Palaeogeography, Palaeoclimatology, Palaeoecology, 465:168{176, jan 2017. ISSN 0031-0182. doi: 10.1016/J.PALAEO.2016.10.028. URL https://www.sciencedirect.com/science/ article/pii/S0031018216306411. T. Li, E. Zhao, J. Zhang, and C. Hu. Detection of Wild re Smoke Images Based on a Densely Dilated Con- volutional Network. Electronics, 8(10):1131, oct 2019. ISSN 2079-9292. doi: 10.3390/electronics8101131. URL https://www.mdpi.com/2079-9292/8/10/1131. Xiaolian Li, Weiguo Song, Liping Lian, and Xiaoge Wei. Forest Fire Smoke Detection Using Back- Propagation Neural Network Based on MODIS Data. Remote Sensing, 7(4):4473{4498, apr 2015. ISSN 2072-4292. doi: 10.3390/rs70404473. URL http://www.mdpi.com/2072-4292/7/4/4473. Xiuqing Li, Zhenxue Chen, Q. M.J. Wu, and Chengyun Liu. 3D Parallel Fully Convolutional Networks for Real-time Video Wild re Smoke Detection, 2018b. ISSN 10518215. Hao Liang, Meng Zhang, and Hailan Wang. A Neural Network Model for Wild re Scale Prediction Using Meteorological Factors. IEEE Access, 7:176746{176755, 2019. ISSN 21693536. doi: 10.1109/ACCESS. 2019.2957837. Chul-Hee Lim, You Seung Kim, Myungsoo Won, Sea Jin Kim, and Woo-Kyun Lee. Can satellite-based data substitute for surveyed data to predict the spatial probability of forest re? A geostatistical approach to forest re in the Republic of Korea. Geomatics, Natural Hazards and Risk, 10(1):719{739, jan 2019. ISSN 1947-5705. doi: 10.1080/19475705.2018.1543210. URL https://www.tandfonline.com/doi/full/10. 1080/19475705.2018.1543210. Y Liu, Y Yang, C Liu, and Yu Gu. Forest Fire Detection Using Arti cial Neural Network Algorithm Imple- mented in Wireless Sensor Networks. ZTE Communications, jun 2015. URL http://wwwen.zte.com. cn/endata/magazine/ztecommunications/2015/2/articles/201507/t20150724{_}443252.html. 52 Zelin Liu, Changhui Peng, Timothy Work, Jean-Noel Candau, Annie DesRochers, and Daniel Kneeshaw. Application of machine-learning methods in forest ecology: recent progress and future challenges. En- vironmental Reviews, 26(4):339{350, dec 2018. ISSN 1181-8700. doi: 10.1139/er-2018-0034. URL http://www.nrcresearchpress.com/doi/10.1139/er-2018-0034. Zhihua Liu and Michael C. Wimberly. Climatic and Landscape In uences on Fire Regimes from 1984 to 2010 in the Western United States. PLOS ONE, 10(10):e0140839, oct 2015. ISSN 1932-6203. doi: 10.1371/journal.pone.0140839. URL http://dx.plos.org/10.1371/journal.pone.0140839. Zhihua Liu and Michael C. Wimberly. Direct and indirect e ects of climate change on projected future re regimes in the western United States. Science of The Total Environment, 542:65{75, jan 2016. ISSN 0048-9697. doi: 10.1016/J.SCITOTENV.2015.10.093. URL https://www.sciencedirect.com/ science/article/pii/S0048969715309098. Zhihua Liu and Jian Yang. Quantifying ecological drivers of ecosystem productivity of the early-successional boreal Larix gmelinii forest. Ecosphere, 5(7):art84, jul 2014. ISSN 2150-8925. doi: 10.1890/ES13-00372.1. URL http://doi.wiley.com/10.1890/ES13-00372.1. Zhihua Liu, Jian Yang, and Hong S. He. Identifying the Threshold of Dominant Controls on Fire Spread in a Boreal Forest Landscape of Northeast China. PLoS ONE, 8(1):e55618, jan 2013. ISSN 1932-6203. doi: 10.1371/journal.pone.0055618. URL https://dx.plos.org/10.1371/journal.pone.0055618. Pablito M. L opez-Serrano, Carlos A. L opez-S anchez, Juan G. Alvarez-Gonz alez, and Jorge Garc a- Guti errez. A Comparison of Machine Learning Techniques Applied to Landsat-5 TM Spectral Data for Biomass Estimation. Canadian Journal of Remote Sensing, 42(6):690{705, nov 2016. ISSN 0703- 8992. doi: 10.1080/07038992.2016.1217485. URL https://www.tandfonline.com/doi/full/10.1080/ 07038992.2016.1217485. F. Javier Lozano, S. Su arez-Seoane, M. Kelly, and E. Luis. A multi-scale approach for modeling re occurrence probability using satellite data and classi cation trees: A case study in a mountain- ous Mediterranean region. Remote Sensing of Environment, 112(3):708{719, mar 2008. ISSN 0034- 4257. doi: 10.1016/J.RSE.2007.06.006. URL https://www.sciencedirect.com/science/article/ pii/S003442570700243X. V Lozhkin, D Tarkhov, V Timofeev, O Lozhkina, and A Vasilyev. Di erential neural network ap- proach in information process for prediction of roadside air pollution by peat re. IOP Conference Series: Materials Science and Engineering, 158(1):012063, nov 2016. ISSN 1757-8981. doi: 10. 1088/1757-899X/158/1/012063. URL http://stacks.iop.org/1757-899X/158/i=1/a=012063?key= crossref.7abf8c3fd66f7ce48986b4554f7aecd5. Guilan Luo, Mei Zhang, Zizhong Yang, and Mingmei Song. Data mining of correlation between re disturbance habitat factors and spider communities. In 2017 4th International Conference on Systems and Informatics (ICSAI), pages 1471{1476. IEEE, nov 2017. ISBN 978-1-5386-1107-4. doi: 10.1109/ ICSAI.2017.8248518. URL http://ieeexplore.ieee.org/document/8248518/. Ruisen Luo, Yingying Dong, Muye Gan, Dejun Li, Shuli Niu, Amy Oliver, Ke Wang, and Yiqi Luo. Global Analysis of In uencing Forces of Fire Activity: the Threshold Relationships between Vegetation and Fire. Life Science Journal, 10(2):15{24, 2013. ISSN 0300-9165. Duncan C Lutes, Robert E Keane, John F Caratti, Carl H Key, Nathan C Benson, Steve Sutherland, and Larry J Gangi. FIREMON: Fire e ects monitoring and inventory system. Gen. Tech. Rep. RMRS- GTR-164. Fort Collins, CO: US Department of Agriculture, Forest Service, Rocky Mountain Research Station. 1 CD., 164, 2006. 53 Jamie M. Lydersen, Malcolm P. North, and Brandon M. Collins. Severity of an uncharacteristically large wild re, the Rim Fire, in forests with relatively restored frequent re regimes. Forest Ecology and Management, 328:326{334, sep 2014. ISSN 0378-1127. doi: 10.1016/J.FORECO.2014.06.005. URL https://www.sciencedirect.com/science/article/pii/S0378112714003661. Jamie M. Lydersen, Brandon M. Collins, Matthew L. Brooks, John R. Matchett, Kristen L. Shive, Nicholas A. Povak, Van R. Kane, and Douglas F. Smith. Evidence of fuels management and re weather in uencing re severity in an extreme re event. Ecological Applications, 27(7):2013{2030, oct 2017. ISSN 10510761. doi: 10.1002/eap.1586. URL http://doi.wiley.com/10.1002/eap.1586. James MacQueen et al. Some methods for classi cation and analysis of multivariate observations. In Proceedings of the fth Berkeley symposium on mathematical statistics and probability, volume 1, pages 281{297. Oakland, CA, USA, 1967. Eduardo Eiji Maeda, Antonio Roberto Formaggio, Yosio Edemir Shimabukuro, Gustavo Felipe Balu e Arcoverde, and Matthew C. Hansen. Predicting forest re in the Brazilian Amazon using MODIS imagery and arti cial neural networks. International Journal of Applied Earth Observation and Geoinformation, 11(4):265{272, aug 2009. ISSN 0303-2434. doi: 10.1016/J.JAG.2009.03.003. URL https://www.sciencedirect.com/science/article/pii/S0303243409000233. Nyasha Magadzire, Helen M. Klerk, Karen J. Esler, and Jasper A. Slingsby. Fire and life history a ect the distribution of plant species in a biodiversity hotspot. Diversity and Distributions, 25(7):1012{1023, jul 2019. ISSN 1366-9516. doi: 10.1111/ddi.12921. URL https://onlinelibrary.wiley.com/doi/abs/ 10.1111/ddi.12921. G. Mallinis, F. Maris, I. Kalinderis, and N. Koutsias. Assessment of Post- re Soil Erosion Risk in Fire- A ected Watersheds Using Remote Sensing and GIS. GIScience & Remote Sensing, 46(4):388{410, oct 2009. ISSN 1548-1603. doi: 10.2747/1548-1603.46.4.388. URL https://www.tandfonline.com/doi/ full/10.2747/1548-1603.46.4.388. Nicolas Mansuy, Carol Miller, Marc-Andr e Parisien, Sean A Parks, Enric Batllori, and Max A Moritz. Con- trasting human in uences and macro-environmental factors on re activity inside and outside protected areas of North America. Environmental Research Letters, 14(6):064007, may 2019. ISSN 1748-9326. doi: 10.1088/1748-9326/ab1bc5. URL https://iopscience.iop.org/article/10.1088/1748-9326/ ab1bc5. Natasha Markuzon and Stephan Kolitz. Data driven approach to estimating re danger from satellite images and weather information. In 2009 IEEE Applied Imagery Pattern Recognition Workshop (AIPR 2009), pages 1{7. IEEE, oct 2009. ISBN 978-1-4244-5146-3. doi: 10.1109/AIPR.2009.5466309. URL http://ieeexplore.ieee.org/document/5466309/. David L. Martell. A review of recent forest and wildland re management decision support systems research, jun 2015. ISSN 21986436. Yago Mart n, Mar a Zu niga-An ~ t on, and Marcos Rodrigues Mimbrero. Modelling temporal variation of re-occurrence towards the dynamic prediction of human wild re ignition danger in northeast Spain. Geomatics, Natural Hazards and Risk, 10(1):385{411, jan 2019. ISSN 1947-5705. doi: 10.1080/19475705. 2018.1526219. URL https://www.tandfonline.com/doi/full/10.1080/19475705.2018.1526219. S. Mart n-Alc on and L Coll. Unraveling the relative importance of factors driving post- re regenera- tion trajectories in non-serotinous Pinus nigra forests. Forest Ecology and Management, 361:13{22, feb 2016. ISSN 0378-1127. doi: 10.1016/J.FORECO.2015.11.006. URL https://www-sciencedirect-com. login.ezproxy.library.ualberta.ca/science/article/pii/S037811271500612X. 54 Arif Masrur, Andrey N. Petrov, and John DeGroote. Circumpolar spatio-temporal patterns and contribut- ing climatic factors of wild re activity in the Arctic tundra from 2001-2015. Environmental Research Letters, 13(1), jan 2018. ISSN 17489326. doi: 10.1088/1748-9326/aa9a76. Robert Mavsar, Armando Gonz alez Cab an, and Elsa Varela. The state of development of re management decision support systems in America and Europe. Forest Policy and Economics, 29:45{55, apr 2013. ISSN 13899341. doi: 10.1016/j.forpol.2012.11.009. R. Stockton Maxwell, Alan H. Taylor, Carl N. Skinner, Hugh D. Sa ord, Rachel E. Isaacs, Catherine Airey, and Amanda B. Young. Landscape-scale modeling of reference period forest conditions and re behavior on heavily logged lands. Ecosphere, 5(3):art32, mar 2014. ISSN 2150-8925. doi: 10.1890/ES13-00294.1. URL http://doi.wiley.com/10.1890/ES13-00294.1. M.J. Mayr, K.A. Vanselow, and C. Samimi. Fire regimes at the arid fringe: A 16-year remote sensing perspective (2000{2016) on the controls of re activity in Namibia from spatial predictive models. Eco- logical Indicators, 91:324{337, aug 2018. ISSN 1470-160X. doi: 10.1016/J.ECOLIND.2018.04.022. URL https://www.sciencedirect.com/science/article/pii/S1470160X18302759. Ronald J Mccormick, Thomas A Brandner, and Timothy F H Allen. TOWARD A THEORY OF MESO- SCALE WILDFIRE MODELING-A COMPLEX SYSTEMS APPROACH USING ARTIFICIAL NEU- RAL NETWORKS. In The Joint Fire Science Conference And Workshop, Boise, Idaho, 1999. Amy McGovern, Kimberly L. Elmore, David John Gagne, Sue Ellen Haupt, Christopher D. Karstens, Ryan Lagerquist, Travis Smith, John K. Williams, Amy McGovern, Kimberly L. Elmore, David John Gagne II, Sue Ellen Haupt, Christopher D. Karstens, Ryan Lagerquist, Travis Smith, and John K. Williams. Using Arti cial Intelligence to Improve Real-Time Decision-Making for High-Impact Weather. Bulletin of the American Meteorological Society, 98(10):2073{2090, oct 2017. ISSN 0003-0007. doi: 10.1175/ BAMS-D-16-0123.1. URL http://journals.ametsoc.org/doi/10.1175/BAMS-D-16-0123.1. Amy McGovern, Ryan Lagerquist, David John Gagne, G. Eli Jergensen, Kimberly L. Elmore, Cameron R. Homeyer, and Travis Smith. Making the Black Box More Transparent: Understanding the Physical Implications of Machine Learning. Bulletin of the American Meteorological Society, 100(11):2175{2199, nov 2019. ISSN 0003-0007. doi: 10.1175/BAMS-D-18-0195.1. URL http://journals.ametsoc.org/ doi/10.1175/BAMS-D-18-0195.1. Sean McGregor, Rachel Houtman, Hailey Buckingham, Claire Montgomery, Ronald Metoyer, and Thomas G Dietterich. Fast simulation for computational sustainability sequential decision making prob- lems. In Proceedings of the 4th International Conference on Computational Sustainability, pages 5{7, Sean McGregor, Rachel Houtman, Claire Montgomery, Ronald Metoyer, and Thomas G. Dietterich. Fast Optimization of Wild re Suppression Policies with SMAC. arXiv preprint, mar 2017. URL http: //arxiv.org/abs/1703.09391. James P. Minas, John W. Hearne, and John W. Handmer. A review of operations research methods applicable to wild re management. International Journal of Wildland Fire, 21(3):189, may 2012. ISSN 1049-8001. doi: 10.1071/WF10129. URL http://www.publish.csiro.au/?paper=WF10129. Yosune Miquelajauregui, Steven G. Cumming, and Sylvie Gauthier. Modelling Variable Fire Severity in Boreal Forests: E ects of Fire Intensity and Stand Structure. PLOS ONE, 11(2):e0150073, feb 2016. ISSN 1932-6203. doi: 10.1371/journal.pone.0150073. URL https://dx.plos.org/10.1371/journal. pone.0150073. Melanie Mitchell. An introduction to genetic algorithms. MIT Press, 1996. ISBN 9780262133166. 55 T.M. Mitchell. Machine Learning. McGraw-HIll, 1997. ISBN 0071154671. Varun Mithal, Guruprasad Nayak, Ankush Khandelwal, Vipin Kumar, Ramakrishna Nemani, and Nikunj Oza. Mapping burned areas in tropical forests using a novel machine learning framework. Remote Sensing, 10(1):69, 2018. Nikolaos E. Mitrakis, Giorgos Mallinis, Nikos Koutsias, and John B. Theocharis. Burned area mapping in Mediterranean environment using medium-resolution multi-spectral data and a neuro-fuzzy classi er. International Journal of Image and Data Fusion, 3(4):299{318, dec 2012. ISSN 1947-9832. doi: 10. 1080/19479832.2011.635604. URL http://www.tandfonline.com/doi/abs/10.1080/19479832.2011. Ioannis Mitsopoulos and Giorgos Mallinis. A data-driven approach to assess large re size generation in Greece. Natural Hazards, 88(3):1591{1607, sep 2017. ISSN 0921-030X. doi: 10.1007/s11069-017-2934-z. URL http://link.springer.com/10.1007/s11069-017-2934-z. J. R. Molina, A. Lora, C. Prades, and F. Rodr guez y Silva. Roadside vegetation planning and con- servation: New approach to prevent and mitigate wild res based on re ignition potential. Forest Ecology and Management, 444:163{173, jul 2019. ISSN 0378-1127. doi: 10.1016/J.FORECO.2019. 04.034. URL https://www-sciencedirect-com.login.ezproxy.library.ualberta.ca/science/ article/pii/S0378112719301501. Max A. Moritz, Marc-Andr e Parisien, Enric Batllori, Meg A. Krawchuk, Je Van Dorn, David J. Ganz, and Katharine Hayhoe. Climate change and disruptions to global re activity. Ecosphere, 3(6):art49, jun 2012. ISSN 2150-8925. doi: 10.1890/ES11-00345.1. URL http://doi.wiley.com/10.1890/ES11-00345.1. Amir Mosavi, Pinar Ozturk, Kwok-wing Chau, Amir Mosavi, Pinar Ozturk, and Kwok-wing Chau. Flood Prediction Using Machine Learning Models: Literature Review. Water, 10(11):1536, oct 2018. ISSN 2073-4441. doi: 10.3390/w10111536. URL http://www.mdpi.com/2073-4441/10/11/1536. Mohsen MOSTAFA, Shaban SHATAEE JOUIBARY, Majid LOTFALIAN, and Amir SADODDIN. WA- TERSHED ROAD NETWORK ANALYSIS WITH AN EMPHASIS ON FIRE FIGHTING MANAGE- MENT. Journal of Environmental Engineering and Landscape Management, 25(4):342{353, dec 2017. ISSN 1648-6897. doi: 10.3846/16486897.2017.1281816. URL http://journals.vgtu.lt/index.php/ JEELM/article/view/1712. Kudzai S. Mpakairi, Paradzayi Tagwireyi, Henry Ndaimani, and Hilary T. Madiri. Distribution of wild- land res and possible hotspots for the Zimbabwean component of Kavango-Zambezi Transfrontier Con- servation Area. South African Geographical Journal, 101(1):110{120, jan 2019. ISSN 0373-6245. doi: 10.1080/03736245.2018.1541023. URL https://www.tandfonline.com/doi/full/10.1080/03736245. 2018.1541023. Khan Muhammad, Jamil Ahmad, and Sung Wook Baik. Early re detection using convolutional neural networks during surveillance for e ective disaster management. Neurocomputing, 288:30{42, may 2018. ISSN 18728286. doi: 10.1016/j.neucom.2017.04.083. Kevin Murphy. Machine Learning: A Probabilistic Perspective. MIT Press, 2012. URL http://www. amazon.com/Machine-Learning-Probabilistic-Perspective-Computation/dp/0262018020. Khurram Nadeem, S. W. Taylor, Douglas G. Woolford, and C. B. Dean. Mesoscale spatiotemporal pre- dictive models of daily human- and lightning-caused wildland re occurrence in British Columbia. In- ternational Journal of Wildland Fire, 29(1):11, 2020. ISSN 1049-8001. doi: 10.1071/WF19058. URL http://www.publish.csiro.au/?paper=WF19058. 56 Hariharan Naganathan, Sudarshan P Seshasayee, Jonghoon Kim, Wai K Chong, and Jui-Sheng Chou. Wild re Predictions: Determining Reliable Models using Fused Dataset. Global Journal of Computer Science and Technology, 16(4C), 2016. URL https://computerresearch.org/index.php/computer/ article/view/1437. Nicholas J. Nauslar, Benjamin J. Hatchett, Timothy J. Brown, Michael L. Kaplan, and John F. Mejia. Impact of the North American monsoon on wild re activity in the southwest United States. International Journal of Climatology, 39(3):1539{1554, mar 2019. ISSN 08998418. doi: 10.1002/joc.5899. URL http://doi.wiley.com/10.1002/joc.5899. Trisalyn A. Nelson, Wiebe Nijland, Mathieu L. Bourbonnais, and Michael A. Wulder. Regression Tree Modeling of Spatial Pattern and Process Interactions. In Mapping Forest Landscape Patterns, pages 187{212. Springer New York, New York, NY, 2017. doi: 10.1007/978-1-4939-7331-6 5. URL http: //link.springer.com/10.1007/978-1-4939-7331-6{_}5. Nguyen Ngoc Thach, Dang Bao-Toan Ngo, Pham Xuan-Canh, Nguyen Hong-Thi, Bui Hang Thi, Hoang Nhat-Duc, and Tien Bui Dieu. Spatial pattern assessment of tropical forest re danger at Thuan Chau area (Vietnam) using GIS-based advanced machine learning algorithms: A comparative study. Ecological Informatics, 46:74{85, jul 2018. ISSN 1574-9541. doi: 10.1016/J.ECOINF.2018.05.009. URL https: //www.sciencedirect.com/science/article/pii/S1574954118300852. I. Nitze, G. Grosse, B. M. Jones, V. E. Romanovsky, and J. Boike. Remote sensing quanti es widespread abundance of permafrost region disturbances across the Arctic and Subarctic. Nature Communications, 9(1), dec 2018. ISSN 20411723. doi: 10.1038/s41467-018-07663-3. Christopher D. O' O'Connor, David E. Calkin, Matthew P. Thompson, Christopher D.O. O'Connor, David E. Calkin, and Matthew P. Thompson. An empirical machine learning method for predict- ing potential re control locations for pre- re planning and operational re management. Interna- tional Journal of Wildland Fire, 26(7):587, jul 2017. ISSN 1049-8001. doi: 10.1071/WF16135. URL http://www.publish.csiro.au/?paper=WF16135. Julian D. Olden, Joshua J. Lawler, and N. LeRoy Po . Machine Learning Methods Without Tears: A Primer for Ecologists. The Quarterly Review of Biology, 83(2):171{193, jun 2008. ISSN 0033-5770. doi: 10.1086/587826. URL https://www.journals.uchicago.edu/doi/10.1086/587826. Sandra Oliveira, Friderike Oehler, Jesus  San-Miguel-Ayanz, Andrea Camia, and Jos e M.C. Pereira. Modeling spatial patterns of re occurrence in Mediterranean Europe using Multiple Regression and Random Forest. Forest Ecology and Management, 275:117{129, jul 2012. ISSN 0378-1127. doi: 10.1016/J.FORECO.2012.03.003. URL https://www.sciencedirect.com/science/article/ pii/S0378112712001272. A. Murat Ozbayo glu and Recep Bozer. Estimation of the Burned Area in Forest Fires Using Compu- tational Intelligence Techniques. Procedia Computer Science, 12:282{287, jan 2012. ISSN 1877-0509. doi: 10.1016/J.PROCS.2012.09.070. URL https://www.sciencedirect.com/science/article/pii/ S1877050912006618. P. Papakosta, G. Xanthopoulos, and D. Straub. Probabilistic prediction of wild re economic losses to housing in Cyprus using Bayesian network analysis. International Journal of Wildland Fire, 26(1):10, feb 2017. ISSN 1049-8001. doi: 10.1071/WF15113. URL http://www.publish.csiro.au/?paper=WF15113. Marc-Andr e Parisien and Max A. Moritz. Environmental controls on the distribution of wild re at multiple spatial scales. Ecological Monographs, 79(1):127{154, feb 2009. ISSN 0012-9615. doi: 10.1890/07-1289.1. URL http://doi.wiley.com/10.1890/07-1289.1. 57 Marc-Andr e Parisien, Sean A. Parks, Meg A. Krawchuk, John M. Little, Mike D. Flannigan, Lynn M. Gowman, and Max A. Moritz. An analysis of controls on re activity in boreal Canada: comparing models built with di erent temporal resolutions. Ecological Applications, 24(6):1341{1356, sep 2014. ISSN 1051-0761. doi: 10.1890/13-1477.1. URL http://doi.wiley.com/10.1890/13-1477.1. Marc And r Parisien, Susan Snetsinger, Jonathan A. Greenberg, Cara R. Nelson, Tania Schoennagel, Solomon Z. Dobrowski, and Max A. Moritz. Spatial variability in wild re probability across the west- ern United States. International Journal of Wildland Fire, 21(4):313{327, 2012. ISSN 10498001. doi: 10.1071/WF11044. Sean A. Parks, Carol Miller, Marc-Andr e Parisien, Lisa M. Holsinger, Solomon Z. Dobrowski, and John Abatzoglou. Wildland re de cit and surplus in the western United States, 1984{2012. Ecosphere, 6 (12):art275, dec 2015. ISSN 2150-8925. doi: 10.1890/ES15-00294.1. URL http://doi.wiley.com/10. 1890/ES15-00294.1. Sean A Parks, Carol Miller, John T Abatzoglou, Lisa M Holsinger, Marc-Andr e Parisien, and Solomon Z Dobrowski. How will climate change a ect wildland re severity in the west- ern US? Environmental Research Letters, 11(3):035002, mar 2016. ISSN 1748-9326. doi: 10.1088/1748-9326/11/3/035002. URL http://stacks.iop.org/1748-9326/11/i=3/a=035002?key= crossref.4d33abcb068f5458baf3b94828ca073e. Sean A Parks, Lisa M Holsinger, Matthew H Panunto, W Matt Jolly, Solomon Z Dobrowski, and Gregory K Dillon. High-severity re: evaluating its key drivers and mapping its probability across western US forests. Environmental Research Letters, 13(4):044037, apr 2018. ISSN 1748-9326. doi: 10.1088/1748-9326/aab791. URL http://stacks.iop.org/1748-9326/13/i=4/a=044037?key= crossref.5c2b6b1d5870d4a9269af3badf873e81. Judea Pearl. Probabilistic Reasoning in Intelligent Systems:Networks of Plausible Inference. Morgan Kauf- mann, San Mateo, California, 1988. T. D. Penman, O. Price, and R. A. Bradstock. Bayes Nets as a method for analysing the in uence of management actions in re planning. International Journal of Wildland Fire, 20(8):909{920, 2011. ISSN 10498001. doi: 10.1071/WF10076. T. D. Penman, R. A. Bradstock, and O. F. Price. Reducing wild re risk to urban developments: Simulation of cost-e ective fuel treatment solutions in south eastern Australia. Environmental Mod- elling and Software, 52:166{175, 2014. ISSN 13648152. doi: 10.1016/j.envsoft.2013.09.030. URL http://dx.doi.org/10.1016/j.envsoft.2013.09.030. T. D. Penman, A. E. Nicholson, R. A. Bradstock, L. Collins, S. H. Penman, and O. F. Price. Reducing the risk of house loss due to wild res. Environmental Modelling and Software, 67:12{25, 2015. ISSN 13648152. doi: 10.1016/j.envsoft.2014.12.020. Allan Pereira, Jos e Pereira, Renata Libonati, Duarte Oom, Alberto Setzer, Fabiano Morelli, Fausto Machado-Silva, Luis de Carvalho, Allan A. Pereira, Jos e M. C. Pereira, Renata Libonati, Duarte Oom, Alberto W. Setzer, Fabiano Morelli, Fausto Machado-Silva, and Luis Marcelo Tavares de Carvalho. Burned Area Mapping in the Brazilian Savanna Using a One-Class Support Vector Machine Trained by Active Fires. Remote Sensing, 9(11):1161, nov 2017. ISSN 2072-4292. doi: 10.3390/rs9111161. URL http://www.mdpi.com/2072-4292/9/11/1161. George L. W. Perry, Janet M. Wilmshurst, Matt S. McGlone, and Aaron Napier. Reconstructing spatial vulnerability to forest loss by re in pre-historic New Zealand. Global Ecology and Biogeography, 21 (10):1029{1041, oct 2012. ISSN 1466822X. doi: 10.1111/j.1466-8238.2011.00745.x. URL http://doi. wiley.com/10.1111/j.1466-8238.2011.00745.x. 58 Matthew P. Peters and Louis R. Iverson. Incorporating ne-scale drought information into an eastern US wild re hazard model. International Journal of Wildland Fire, 26(5):393, may 2017. ISSN 1049-8001. doi: 10.1071/WF16130. URL http://www.publish.csiro.au/?paper=WF16130. Matthew P. Peters, Louis R. Iverson, Stephen N. Matthews, and Anantha M. Prasad. Wild re hazard mapping: exploring site conditions in eastern US wildland{urban interfaces. International Journal of Wildland Fire, 22(5):567, aug 2013. ISSN 1049-8001. doi: 10.1071/WF12177. URL http://www. publish.csiro.au/?paper=WF12177. G. P. Petropoulos, W. Knorr, M. Scholze, L. Boschetti, and G. Karantounias. Combining ASTER multispectral imagery analysis and support vector machines for rapid and cost-e ective post- re as- sessment: a case study from the Greek wildland res of 2007. Natural Hazards and Earth Sys- tem Science, 10(2):305{317, feb 2010. ISSN 1684-9981. doi: 10.5194/nhess-10-305-2010. URL http://www.nat-hazards-earth-syst-sci.net/10/305/2010/. George P. Petropoulos, Charalambos Kontoes, and Iphigenia Keramitsoglou. Burnt area delineation from a uni-temporal perspective based on Landsat TM imagery classi cation using Support Vector Ma- chines. International Journal of Applied Earth Observation and Geoinformation, 13(1):70{80, feb 2011. ISSN 0303-2434. doi: 10.1016/J.JAG.2010.06.008. URL https://www.sciencedirect.com/science/ article/pii/S0303243410000784. M.T. Pham, A. Raji c, J.D. Greig, J.M. Sargeant, A. Papadopoulos, and S.A. Mcewen. A scoping review of scoping reviews: Advancing the approach and enhancing the consistency. Research Synthesis Methods, 5(4):371{385, dec 2014. ISSN 17592887. doi: 10.1002/jrsm.1123. Thanh Cong Phan and Thanh Tam Nguyen. Remote Sensing meets Deep Learning: Exploiting Spatio- Temporal-Spectral Satellite Images for Early Wild re Detection. Technical report, 2019. URL https: //infoscience.epfl.ch/record/270339. Sharon B. Phillips, Viney P. Aneja, Daiwen Kang, and S. Pal Arya. Modelling and analysis of the atmo- spheric nitrogen deposition in North Carolina. In International Journal of Global Environmental Issues, volume 6, pages 231{252, 2006. doi: 10.1016/j.ecolmodel.2005.03.026. Andrew D. Pierce, Calvin A. Farris, and Alan H. Taylor. Use of random forests for modeling and mapping forest canopy fuels for re behavior analysis in Lassen Volcanic National Park, California, USA. Forest Ecology and Management, 279:77{89, sep 2012. ISSN 0378-1127. doi: 10.1016/J.FORECO.2012.05.010. URL https://www.sciencedirect.com/science/article/pii/S0378112712002654. David L Poole and Alan K Mackworth. Arti cial Intelligence: foundations of computational agents. Cam- bridge University Press, 2010. Patrick Poon, Alicia Kinoshita, Patrick K. Poon, and Alicia M. Kinoshita. Estimating Evapotranspiration in a Post-Fire Environment Using Remote Sensing and Machine Learning. Remote Sensing, 10(11):1728, nov 2018. ISSN 2072-4292. doi: 10.3390/rs10111728. URL http://www.mdpi.com/2072-4292/10/11/ Zohre Sadat Pourtaghi, Hamid Reza Pourghasemi, Roberta Aretano, and Teodoro Semeraro. Investigation of general indicators in uencing on forest re and its susceptibility modeling using di erent data mining techniques. Ecological Indicators, 64:72{84, may 2016. ISSN 1470-160X. doi: 10.1016/J.ECOLIND. 2015.12.030. URL https://www.sciencedirect.com/science/article/pii/S1470160X15007359. Ruiliang Pu and Peng Gong. Determination of Burnt Scars Using Logistic Regression and Neural Network Techniques from a Single Post-Fire Landsat 7 ETM + Image. Photogram- metric Engineering & Remote Sensing, 70(7):841{850, jul 2004. ISSN 00991112. doi: 10. 59 14358/PERS.70.7.841. URL http://openurl.ingenta.com/content/xref?genre=article{&}issn= 0099-1112{&}volume=70{&}issue=7{&}spage=841. J Ross Quinlan. C 4.5: Programs for machine learning. The Morgan Kaufmann Series in Machine Learning, San Mateo, CA: Morgan Kaufmann,| c1993, 1993. Carmen Quintano, Alfonso Fern andez-Manso, Leonor Calvo, and Dar A. Roberts. Vegetation and Soil Fire Damage Analysis Based on Species Distribution Modeling Trained with Multispectral Satellite Data. Remote Sensing, 11(15):1832, aug 2019. ISSN 2072-4292. doi: 10.3390/rs11151832. URL https: //www.mdpi.com/2072-4292/11/15/1832. Natalia Quintero, Olga Viedma, Itziar R. Urbieta, and Jos e M. Moreno. Assessing Landscape Fire Hazard by Multitemporal Automatic Classi cation of Landsat Time Series Using the Google Earth Engine in West-Central Spain. Forests, 10(6):518, jun 2019. ISSN 1999-4907. doi: 10.3390/f10060518. URL https://www.mdpi.com/1999-4907/10/6/518. David Radke, Anna Hessler, and Dan Ellsworth. FireCast: Leveraging Deep Learning to Predict Wild re Spread. In Proceedings of the Twenty-Eighth International Joint Conference on Arti cial Intelligence (IJCAI-19), pages 4575{4581. International Joint Conferences on Arti cial Intelligence, jul 2019. doi: 10.24963/ijcai.2019/636. M. Raissi, P. Perdikaris, and G. E. Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial di erential equations. Journal of Computational Physics, 378:686{707, feb 2019. ISSN 10902716. doi: 10.1016/j.jcp.2018.10. Maziar Raissi and George Em Karniadakis. Hidden physics models: Machine learning of nonlinear partial di erential equations. Journal of Computational Physics, 357:125{141, mar 2018. ISSN 10902716. doi: 10.1016/j.jcp.2017.11.039. Carl Edward Rasmussen and Christopher KI Williams. Gaussian processes for machine learning, volume 1. MIT press Cambridge, 2006. Stephan Rasp and Sebastian Lerch. Neural Networks for Postprocessing Ensemble Weather Forecasts. Monthly Weather Review, 146(11):3885{3900, nov 2018. ISSN 0027-0644. doi: 10.1175/MWR-D-18-0187. 1. URL http://journals.ametsoc.org/doi/10.1175/MWR-D-18-0187.1. Markus Reichstein, Gustau Camps-Valls, Bjorn Stevens, Martin Jung, Joachim Denzler, Nuno Carvalhais, and Prabhat. Deep learning and process understanding for data-driven Earth system science. Nature, 566(7743):195{204, feb 2019. ISSN 0028-0836. doi: 10.1038/s41586-019-0912-1. URL http://www. nature.com/articles/s41586-019-0912-1. Colleen E. Reid, Michael Jerrett, Maya L. Petersen, Gabriele G. P ster, Philip E. More eld, Ira B. Tager, Sean M. Ra use, and John R. Balmes. Spatiotemporal Prediction of Fine Particulate Matter During the 2008 Northern California Wild res Using Machine Learning. Environmental Science & Technology, 49 (6):3887{3896, mar 2015. ISSN 0013-936X. doi: 10.1021/es505846r. URL http://pubs.acs.org/doi/ 10.1021/es505846r. Quentin Renard, Rapha el P elissier, B. R. Ramesh, and Narendran Kodandapani. Environmental sus- ceptibility model for predicting forest re occurrence in the Western Ghats of India. International Journal of Wildland Fire, 21(4):368, jul 2012. ISSN 1049-8001. doi: 10.1071/WF10109. URL http://www.publish.csiro.au/?paper=WF10109. 60 April E. Reside, Jeremy VanDerWal, Alex Kutt, Ian Watson, and Stephen Williams. Fire regime shifts af- fect bird species distributions. Diversity and Distributions, 18(3):213{225, mar 2012. ISSN 13669516. doi: 10.1111/j.1472-4642.2011.00818.x. URL http://doi.wiley.com/10.1111/j.1472-4642.2011.00818. x. D. Riano, ~ S. L. Ustin, L. Usero, and M. A. Patricio. Estimation of Fuel Moisture Content Using Neural Networks. pages 489{498. Springer, Berlin, Heidelberg, 2005. doi: 10.1007/11499305 50. URL http: //link.springer.com/10.1007/11499305{_}50. Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. Model-Agnostic Interpretability of Machine Learning. jun 2016. URL http://arxiv.org/abs/1606.05386. Rihan, Zhao, Zhang, Guo, Ying, Deng, and Li. Wild res on the Mongolian Plateau: Identifying Drivers and Spatial Distributions to Predict Wild re Probability. Remote Sensing, 11(20):2361, oct 2019. ISSN 2072-4292. doi: 10.3390/rs11202361. URL https://www.mdpi.com/2072-4292/11/20/2361. David R. Roberts, Volker Bahn, Simone Ciuti, Mark S. Boyce, Jane Elith, Gurutzeta Guillera-Arroita, Sev- erin Hauenstein, Jos e J. Lahoz-Monfort, Boris Schr oder, Wilfried Thuiller, David I. Warton, Brendan A. Wintle, Florian Hartig, and Carsten F. Dormann. Cross-validation strategies for data with temporal, spatial, hierarchical, or phylogenetic structure. Ecography, 40(8):913{929, aug 2017. ISSN 09067590. doi: 10.1111/ecog.02881. URL http://doi.wiley.com/10.1111/ecog.02881. Marcos Rodrigues and Juan De la Riva. An insight into machine-learning algorithms to model human- caused wild re occurrence. Environmental Modelling and Software, 57:192{201, 2014. ISSN 13648152. doi: 10.1016/j.envsoft.2014.03.003. URL http://dx.doi.org/10.1016/j.envsoft.2014.03.003. Marcos Rodrigues, Ferm n Alcasena, and Cristina Vega-Garc a. Modeling initial attack success of wild re suppression in Catalonia, Spain. Science of The Total Environment, 666:915{927, may 2019. ISSN 00489697. doi: 10.1016/j.scitotenv.2019.02.323. URL https://linkinghub.elsevier.com/retrieve/ pii/S0048969719308319. Roque Rodriguez, Ana Cort es, Tom as Margalef, and Emilio Luque. An Adaptive System for Forest Fire Behavior Prediction. In 2008 11th IEEE International Conference on Computational Science and Engineering, pages 275{282. IEEE, jul 2008. ISBN 978-0-7695-3193-9. doi: 10.1109/CSE.2008.15. URL http://ieeexplore.ieee.org/document/4578243/. Roque Rodr guez, Ana Cort es, and Tom as Margalef. Injecting Dynamic Real-Time Data into a DDDAS for Forest Fire Behavior Prediction. pages 489{499. Springer, Berlin, Heidelberg, 2009. doi: 10.1007/ 978-3-642-01973-9 55. URL http://link.springer.com/10.1007/978-3-642-01973-9{_}55. Yuji Roh, Geon Heo, and Steven Euijong Whang. A Survey on Data Collection for Machine Learning: a Big Data { AI Integration Perspective. IEEE Transactions on Knowledge and Data Engineering, pages 1{1, nov 2018. URL http://arxiv.org/abs/1811.03402. David Rolnick, Priya L. Donti, Lynn H. Kaack, Kelly Kochanski, Alexandre Lacoste, Kris Sankaran, An- drew Slavin Ross, Nikola Milojevic-Dupont, Natasha Jaques, Anna Waldman-Brown, Alexandra Luc- cioni, Tegan Maharaj, Evan D. Sherwin, S. Karthik Mukkavilli, Konrad P. Kording, Carla Gomes, Andrew Y. Ng, Demis Hassabis, John C. Platt, Felix Creutzig, Jennifer Chayes, and Yoshua Bengio. Tackling Climate Change with Machine Learning. jun 2019. URL http://arxiv.org/abs/1906.05433. J. Ru ault and F. Mouillot. How a new re-suppression policy can abruptly reshape the re-weather relationship. Ecosphere, 6(10):art199, oct 2015. ISSN 2150-8925. doi: 10.1890/ES15-00182.1. URL http://doi.wiley.com/10.1890/ES15-00182.1. 61 Jakob Runge, Sebastian Bathiany, Erik Bollt, Gustau Camps-Valls, Dim Coumou, Ethan Deyle, Clark Glymour, Marlene Kretschmer, Miguel D. Mahecha, Jordi Munoz-Mar ~  , Egbert H. van Nes, Jonas Peters, Rick Quax, Markus Reichstein, Marten Sche er, Bernhard Sch olkopf, Peter Spirtes, George Sugihara, Jie Sun, Kun Zhang, and Jakob Zscheischler. Inferring causation from time series in Earth system sciences. Nature Communications, 10(1):1{13, dec 2019. ISSN 20411723. doi: 10.1038/s41467-019-10105-3. A. C. L. S a, J. M. C. Pereira, M. J. P. Vasconcelos, J. M. N. Silva, N. Ribeiro, and A. Awasse. Assessing the feasibility of sub-pixel burned area mapping in miombo woodlands of northern Mozambique using MODIS imagery. International Journal of Remote Sensing, 24(8):1783{1796, jan 2003. ISSN 0143- 1161. doi: 10.1080/01431160210144750. URL https://www.tandfonline.com/doi/full/10.1080/ Shruti Sachdeva, Tarunpreet Bhatia, and A. K. Verma. GIS-based evolutionary optimized Gradient Boosted Decision Trees for forest re susceptibility mapping. Natural Hazards, 92(3):1399{1418, jul 2018. ISSN 0921-030X. doi: 10.1007/s11069-018-3256-5. URL http://link.springer.com/10.1007/ s11069-018-3256-5. Y. Sa and A. Bouroumi. Prediction of forest res using Arti cial neural networks. Applied Mathematical Sciences, 7(6):271{286, 2013. doi: 10.12988/ams.2013.13025. George E. Sakr, Imad H. Elhajj, George Mitri, and Uchechukwu C. Wejinya. Arti cial intelligence for forest re prediction. In IEEE/ASME International Conference on Advanced Intelligent Mechatronics, AIM, 2010. ISBN 9781424480319. doi: 10.1109/AIM.2010.5695809. George E. Sakr, Imad H. Elhajj, and George Mitri. Ecient forest re occurrence prediction for developing countries using two weather parameters. Engineering Applications of Arti cial Intelli- gence, 24(5):888{894, aug 2011. ISSN 0952-1976. doi: 10.1016/J.ENGAPPAI.2011.02.017. URL https://www.sciencedirect.com/science/article/abs/pii/S0952197611000418. Jess San-Miguel-Ayanz, Ernst Schulte, Guido Schmuck, Andrea Camia, Peter Strobl, Giorgio Liberta, Cris- tiano Giovando, Roberto Boca, Fernando Sedano, Pieter Kempeneers, Daniel McInerney, Ceri Withmore, Sandra Santos de Oliveira, Marcos Rodrigues, Tracy Durrant, Paolo Corti, Friderike Oehler, Lara Vilar, and Giuseppe Amatulli. Comprehensive Monitoring of Wild res in Europe: The European Forest Fire Information System (EFFIS). In Approaches to Managing Disaster - Assessing Hazards, Emergencies and Disaster Impacts. InTech, mar 2012. doi: 10.5772/28441. L.A. Sanabria, X. Qin, J. Li, R.P. Cechet, and C. Lucas. Spatial interpolation of McArthur's Forest Fire Danger Index across Australia: Observational study. Environmental Modelling & Software, 50:37{50, dec 2013. ISSN 1364-8152. doi: 10.1016/J.ENVSOFT.2013.08.012. URL https://www.sciencedirect. com/science/article/pii/S1364815213001916?via{%}3Dihub. Onur Satir, Suha Berberoglu, and Cenk Donmez. Mapping regional forest re probability using arti cial neural network model in a Mediterranean forest ecosystem. Geomatics, Natural Hazards and Risk, 7 (5):1645{1658, sep 2016. ISSN 1947-5705. doi: 10.1080/19475705.2015.1084541. URL http://www. tandfonline.com/doi/full/10.1080/19475705.2015.1084541. Younes Oulad Sayad, Hajar Mousannif, and Hassan Al Moatassime. Predictive modeling of wild res: A new dataset and machine learning approach. Fire Safety Journal, 104:130{146, mar 2019. ISSN 03797112. doi: 10.1016/j. resaf.2019.01.006. Daniel L. Schmoldt. Application of Arti cial Intelligence to Risk Analysis for Forested Ecosystems. pages 49{74. Springer, Dordrecht, 2001. doi: 10.1007/978-94-017-2905-5 3. URL http://link.springer. com/10.1007/978-94-017-2905-5{_}3. 62 Frederic Paik Schoenberg. A NOTE ON THE CONSISTENT ESTIMATION OF SPATIAL-TEMPORAL POINT PROCESS PARAMETERS, 2016. URL https://www.jstor.org/stable/24721302. Gavin Shaddick and James V. Zidek. A case study in preferential sampling: Long term monitoring of air pollution in the UK. Spatial Statistics, 9:51{65, aug 2014. ISSN 2211-6753. doi: 10.1016/J.SPASTA.2014. 03.008. URL https://www.sciencedirect.com/science/article/abs/pii/S2211675314000219. Chaopeng Shen. A Transdisciplinary Review of Deep Learning Research and Its Relevance for Water Resources Scientists. Water Resources Research, 54(11):8558{8593, nov 2018. ISSN 0043-1397. doi: 10.1029/2018WR022643. URL https://onlinelibrary.wiley.com/doi/abs/10.1029/2018WR022643. Kirk R. Sherrill and William H. Romme. Spatial Variation in Post re Cheatgrass: Dinosaur National Monument, USA. Fire Ecology, 8(2):38{56, aug 2012. ISSN 19339747. doi: 10.4996/ reecology.0802038. URL http://fireecologyjournal.org/journal/abstract/?abstract=162. Mengyun Shi, Fengying Xie, Yue Zi, and Jihao Yin. Cloud detection of remote sensing images by deep learning. In International Geoscience and Remote Sensing Symposium (IGARSS), volume 2016-Novem, pages 701{704. Institute of Electrical and Electronics Engineers Inc., nov 2016. ISBN 9781509033324. doi: 10.1109/IGARSS.2016.7729176. Guruh Fajar Shidik and Khabib Mustofa. Predicting Size of Forest Fire Using Hybrid Model. pages 316{327. Springer, Berlin, Heidelberg, 2014. doi: 10.1007/978-3-642-55032-4 31. URL http://link. springer.com/10.1007/978-3-642-55032-4{_}31. Albert J. Simard. Fire severity, changing scales, and how things hang together. International Journal of Wildland Fire, 1(1):23{34, 1991. ISSN 10498001. doi: 10.1071/WF9910023. Imas Sukaesih Sitanggang and Mohd Hasmadi Ismail. Classi cation model for hotspot occurrences using a decision tree method. Geomatics, Natural Hazards and Risk, 2(2):111{121, jun 2011. ISSN 19475705. doi: 10.1080/19475705.2011.565807. I.S. Sitanggang, R. Yaakob, N. Mustapha, and A.N. Ainuddin. Predictive Models for Hotspots Occurrence using Decision Tree Algorithms and Logistic Regression. Journal of Applied Sciences, 13(2):252{261, feb 2013. ISSN 18125654. doi: 10.3923/jas.2013.252.261. URL http://www.scialert.net/abstract/ ?doi=jas.2013.252.261. Natasa Ski c and Jennifer Francis. Self-organizing maps: a powerful tool for the atmospheric sciences. Applications of Self-Organizing Maps, pages 251{268, 2012. W. R. Skinner, M. D. Flannigan, B. J. Stocks, D. L. Martell, B. M. Wotton, J. B. Todd, J. A. Mason, K. A. Logan, and E. M. Bosch. A 500 hPa synoptic wildland re climatology for large Canadian forest res, 1959-1996. Theoretical and Applied Climatology, 71(3-4):157{169, 2002. ISSN 0177798X. doi: 10.1007/s007040200002. Hamdy Soliman, Komal Sudan, and Ashish Mishra. A smart forest- re early detection sensory system: Another approach of utilizing wireless sensor and neural networks. In 2010 IEEE Sensors, pages 1900{ 1904. IEEE, nov 2010. ISBN 978-1-4244-8170-5. doi: 10.1109/ICSENS.2010.5690033. URL http: //ieeexplore.ieee.org/document/5690033/. Chao Song, Mei-Po Kwan, Weiguo Song, and Jiping Zhu. A Comparison between Spatial Econometric Models and Random Forest for Modeling Fire Occurrence. Sustainability, 9(5):819, may 2017. ISSN 2071-1050. doi: 10.3390/su9050819. URL http://www.mdpi.com/2071-1050/9/5/819. Kalli Srinivasa, Nageswara Prasad, and S Ramakrishna. An Autonomous Forest Fire Detection System Based On Spatial Data Mining and Fuzzy Logic. Technical Report 12, 2008. 63 B. J. Stocks and David L. Martell. Forest re management expenditures in Canada: 1970-2013. Forestry Chronicle, 92(3):298{306, jun 2016. ISSN 00157546. doi: 10.5558/tfc2016-056. Daniela Stojanova, Andrej Kobler, Sa so D zeroski, and Katerina Ta skova. LEARNING TO PREDICT FOREST FIRES WITH DIFFERENT DATA MINING TECHNIQUES. In Conference on data mining and data warehouses (SiKDD 2006), pages 255{258, 2006. URL http://www.academia.edu/download/ 30570649/10.1.1.116.2555.pdf. Daniela Stojanova, Andrej Kobler, Peter Ogrinc, Bernard Zenko, and Sa so D zeroski. Estimating the risk of re outbreaks in the natural environment. Data Mining and Knowledge Discovery, 24(2):411{442, mar 2012. ISSN 1384-5810. doi: 10.1007/s10618-011-0213-2. URL http://link.springer.com/10.1007/ s10618-011-0213-2. Jeremy Storer and Robert Green. PSO trained Neural Networks for predicting forest re size: A comparison of implementation and performance. In Proceedings of the International Joint Conference on Neural Networks, volume 2016-Octob, pages 676{683. Institute of Electrical and Electronics Engineers Inc., oct 2016. ISBN 9781509006199. doi: 10.1109/IJCNN.2016.7727265. Diana Stralberg, Xianli Wang, Marc-Andr e Parisien, Fran cois-Nicolas Robinne, P eter S olymos, C. Lisa Mahon, Scott E. Nielsen, and Erin M. Bayne. Wild re-mediated vegetation change in boreal forests of Alberta, Canada. Ecosphere, 9(3):e02156, 2018. ISSN 2150-8925. doi: 10.1002/ecs2.2156. Carolin Strobl, Anne-Laure Boulesteix, Achim Zeileis, and Torsten Hothorn. Bias in random forest variable importance measures: Illustrations, sources and a solution. BMC Bioinformatics, 8(1):25, dec 2007. ISSN 1471-2105. doi: 10.1186/1471-2105-8-25. URL https://bmcbioinformatics.biomedcentral. com/articles/10.1186/1471-2105-8-25. Esther D. Stroh, Matthew A. Struckho , Michael C. Stambaugh, and Richard P. Guyette. Fire and Climate Suitability for Woody Vegetation Communities in the South Central United States. Fire Ecology 2018 14:1, 14(1):106{124, feb 2018. ISSN 1933-9747. doi: 10.4996/FIREECOLOGY.140110612. URL https://fireecology.springeropen.com/articles/10.4996/fireecology.140110612. Sriram Ganapathi Subramanian and Mark Crowley. Learning Forest Wild re Dynamics from Satellite Images Using Reinforcement Learning. In Conference on Reinforcement Learning and Decision Making, page 5, 2017. URL http://www.ospo.noaa.gov/Products/atmosphere/wind.html. Sriram Ganapathi Subramanian and Mark Crowley. Using Spatial Reinforcement Learning to Build Forest Wild re Dynamics Models From Satellite Images. Frontiers in ICT, 5:6, apr 2018. ISSN 2297-198X. doi: 10.3389/ ct.2018.00006. URL http://journal.frontiersin.org/article/10.3389/fict.2018. 00006/full. AL Sullivan. A review of wildland re spread modelling, 1990-present 3: Mathematical analogues and simulation models. arXiv preprint arXiv:0706.4130, 2007. Andrew L. Sullivan. Wildland surface re spread modelling, 1990 - 2007. 3: Simulation and mathematical analogue models. International Journal of Wildland Fire, 18(4):387, 2009a. ISSN 1049-8001. doi: 10.1071/wf06144. Andrew L. Sullivan. Wildland surface re spread modelling, 1990 - 2007. 2: Empirical and quasi-empirical models. International Journal of Wildland Fire, 18(4):369, 2009b. ISSN 1049-8001. doi: 10.1071/ wf06142. Andrew L. Sullivan. Wildland surface re spread modelling, 1990 - 2007. 1: Physical and quasi-physical models. International Journal of Wildland Fire, 18(4):349, 2009c. ISSN 1049-8001. doi: 10.1071/wf06143. 64 Brian L. Sullivan, Jocelyn L. Aycrigg, Jessie H. Barry, Rick E. Bonney, Nicholas Bruns, Caren B. Cooper, Theo Damoulas, Andr e A. Dhondt, Tom Dietterich, Andrew Farnsworth, Daniel Fink, John W. Fitz- patrick, Thomas Fredericks, Je Gerbracht, Carla Gomes, Wesley M. Hochachka, Marshall J. Ili , Carl Lagoze, Frank A. La Sorte, Matthew Merri eld, Will Morris, Tina B. Phillips, Mark Reynolds, Amanda D. Rodewald, Kenneth V. Rosenberg, Nancy M. Trautmann, Andrea Wiggins, David W. Win- kler, Weng Keen Wong, Christopher L. Wood, Jun Yu, and Steve Kelling. The eBird enterprise: An integrated approach to development and application of citizen science, jan 2014. ISSN 00063207. Alexander Y. Sun and Bridget R Scanlon. How can big data and machine learning bene t environment and water management: A survey of methods, applications, and future directions. Environmental Research Letters, apr 2019. ISSN 1748-9326. doi: 10.1088/1748-9326/ab1b7d. URL http://iopscience.iop. org/article/10.1088/1748-9326/ab1b7d. F. Sunar and C. Ozkan. Forest re analysis with remote sensing data. International Journal of Remote Sensing, 22(12):2265{2277, jan 2001. ISSN 0143-1161. doi: 10.1080/01431160118510. URL https: //www.tandfonline.com/doi/full/10.1080/01431160118510. Richard S Sutton and Andrew G Barto. Introduction to reinforcement learning, volume 135. MIT press Cambridge, 1998. Richard S Sutton and Andrew G Barto. Reinforcement learning: An introduction. MIT press, 2018. Alexandra D. Syphard, Jon E. Keeley, Avi Bar Massada, Teresa J. Brennan, and Volker C. Radelo . Housing Arrangement and Location Determine the Likelihood of Housing Loss Due to Wild re. PLoS ONE, 7(3):e33954, mar 2012. ISSN 1932-6203. doi: 10.1371/journal.pone.0033954. URL http://dx. plos.org/10.1371/journal.pone.0033954. Alexandra D. Syphard, Avi Bar Massada, Van Butsic, and Jon E. Keeley. Land Use Planning and Wild re: Development Policies In uence Future Probability of Housing Loss. PLoS ONE, 8(8):e71708, aug 2013. ISSN 1932-6203. doi: 10.1371/journal.pone.0071708. URL https://dx.plos.org/10.1371/journal. pone.0071708. Alexandra D. Syphard, Van Butsic, Avi Bar-Massada, Jon E. Keeley, Je A. Tracey, and Robert N. Fisher. Setting priorities for private land conservation in re-prone landscapes: Are re risk reduction and biodiversity conservation competing or compatible objectives? Ecology and Society, 21(3):art2, jul 2016. ISSN 1708-3087. doi: 10.5751/ES-08410-210302. URL http://www.ecologyandsociety.org/ vol21/iss3/art2/. Alexandra D. Syphard, Heather Rustigian-Romsos, Michael Mann, Erin Conlisk, Max A. Moritz, and David Ackerly. The relative in uence of climate and housing development on current and projected future re patterns and structure loss across three California landscapes. Global Environmental Change, 56:41{55, may 2019. ISSN 0959-3780. doi: 10.1016/J.GLOENVCHA.2019.03.007. URL https://www. sciencedirect.com/science/article/pii/S0959378018313293. S. W. Taylor, Douglas G. Woolford, C. B. Dean, and David L. Martell. Wild re Prediction to Inform Fire Management: Statistical Science Challenges. Statistical Science, 28(4):586{615, 2013. ISSN 0883-4237. doi: 10.1214/13-STS451. S.W. Taylor. Atmospheric cascades shape wild re re management decision spaces - a theory unifying re weather and re management. Submitted, 2020. Mahyat Shafapour Tehrany, Simon Jones, Farzin Shabani, Francisco Mart nez-Alvarez, and Dieu Tien Bui. A novel ensemble modeling approach for the spatial prediction of tropical forest re susceptibility using LogitBoost machine learning classi er and multi-source geospatial data. Theoretical and Applied 65 Climatology, pages 1{17, sep 2018. ISSN 0177-798X. doi: 10.1007/s00704-018-2628-9. URL http: //link.springer.com/10.1007/s00704-018-2628-9. Jonathan R. Thompson and Thomas A. Spies. Factors associated with crown damage following recurring mixed-severity wild res and post- re management in southwestern Oregon. Landscape Ecology, 25(5): 775{789, may 2010. ISSN 0921-2973. doi: 10.1007/s10980-010-9456-3. URL http://link.springer. com/10.1007/s10980-010-9456-3. Matthew P. Thompson and Dave E. Calkin. Uncertainty and risk in wildland re management: A re- view. Journal of Environmental Management, 92(8):1895{1909, aug 2011. ISSN 0301-4797. doi: 10.1016/J.JENVMAN.2011.03.015. URL https://www.sciencedirect.com/science/article/pii/ S0301479711000818. Dieu Tien Bui, Quang-Thanh Bui, Quoc-Phi Nguyen, Biswajeet Pradhan, Haleh Nampak, and Phan Trong Trinh. A hybrid arti cial intelligence approach using GIS-based neural-fuzzy inference system and particle swarm optimization for forest re susceptibility modeling at a tropical area. Agricultural and Forest Meteorology, 233:32{44, feb 2017. ISSN 0168-1923. doi: 10.1016/J.AGRFORMET.2016.11.002. URL https://www.sciencedirect.com/science/article/pii/S0168192316304269. Ahmed Toujani, Hammadi Achour, and Sami Fa z. Estimating Forest Fire Losses Using Stochas- tic Approach: Case Study of the Kroumiria Mountains (Northwestern Tunisia). Applied Arti cial Intelligence, pages 1{25, sep 2018. ISSN 0883-9514. doi: 10.1080/08839514.2018.1514808. URL https://www.tandfonline.com/doi/full/10.1080/08839514.2018.1514808. James L. Tracy, Antonio Trabucco, A. Michelle Lawing, J. Tomasz Giermakowski, Maria Tchakerian, Gail M. Drus, and Robert N. Coulson. Random subset feature selection for ecological niche models of wild re activity in Western North America. Ecological Modelling, 383:52{68, sep 2018. ISSN 0304- 3800. doi: 10.1016/J.ECOLMODEL.2018.05.019. URL https://www.sciencedirect.com/science/ article/pii/S0304380018301868. Cordy Tymstra, Brian J. Stocks, Xinli Cai, and Mike D. Flannigan. Wild re management in Canada: Review, challenges and opportunities. Progress in Disaster Science, page 100045, 2019. ISSN 25900617. doi: 10.1016/j.pdisas.2019.100045. URL https://linkinghub.elsevier.com/retrieve/ pii/S2590061719300456. A.B. Utkin, Armando M Fernandes, Fernando Sim~ oes, and R. Vilar. Forest- re detection by means of lidar. Proceedings of IV International Conference on Forest Fire Research, (1993):1{14, 2002. Giorgio Vacchiano, Cristiano Foderi, Roberta Berretti, Enrico Marchi, and Renzo Motta. Modeling anthropogenic and natural re ignitions in an inner-alpine valley. Natural Hazards and Earth Sys- tem Sciences, 18(3):935{948, mar 2018. ISSN 1684-9981. doi: 10.5194/nhess-18-935-2018. URL https://www.nat-hazards-earth-syst-sci.net/18/935/2018/. D. Vakalis, H. Sarimveis, C. Kiranoudis, A. Alexandridis, and G. Bafas. A GIS based operational system for wildland re crisis management I. Mathematical modelling and simulation. Applied Mathematical Modelling, 28(4):389{410, 2004. ISSN 0307904X. doi: 10.1016/j.apm.2003.10.005. Miguel Conrado Valdez, Kang-Tsung Chang, Chi-Farn Chen, Shou-Hao Chiang, and Jorge Luis Santos. Modelling the spatial variability of wild re susceptibility in Honduras using remote sensing and geo- graphical information systems. Geomatics, Natural Hazards and Risk, 8(2):876{892, dec 2017. ISSN 1947-5705. doi: 10.1080/19475705.2016.1278404. URL https://www.tandfonline.com/doi/full/10. 1080/19475705.2016.1278404. 66 Ashley E. Van Beusekom, William A. Gould, A. Carolina Monmany, Azad Henareh Khalyani, Maya Quinones, ~ Stephen J. Fain, Maria Jos e Andrade-Nu nez, ~ and Grizelle Gonz alez. Fire weather and likelihood: characterizing climate space for re occurrence and extent in Puerto Rico. Climatic Change, 146(1-2):117{131, jan 2018. ISSN 0165-0009. doi: 10.1007/s10584-017-2045-6. URL http: //link.springer.com/10.1007/s10584-017-2045-6. P. van Breugel, I. Friis, Sebsebe Demissew, Jens-Peter Barnekow Lilles, and Roeland Kindt. Current and Future Fire Regimes and Their In uence on Natural Vegetation in Ethiopia. Ecosystems, 19(2):369{386, mar 2016. ISSN 1432-9840. doi: 10.1007/s10021-015-9938-x. URL http://link.springer.com/10. 1007/s10021-015-9938-x. Ce Van Wagner. Development and Structure of the Canadian Forest Fire Weather Index System. 1987. ISBN 0662151984. doi: 19927. Thomas Vandal, Evan Kodra, and Auroop R. Ganguly. Intercomparison of machine learning methods for statistical downscaling: the case of daily and extreme precipitation. Theoretical and Applied Climatology, pages 1{14, sep 2018. ISSN 0177-798X. doi: 10.1007/s00704-018-2613-3. URL http://link.springer. com/10.1007/s00704-018-2613-3. Christos Vasilakos, Kostas Kalabokidis, John Hatzopoulos, George Kallos, and Yiannis Matsinos. Integrat- ing new methods and tools in re danger rating. International Journal of Wildland Fire, 16(3):306{316, 2007. ISSN 10498001. doi: 10.1071/WF05091. Christos Vasilakos, Kostas Kalabokidis, John Hatzopoulos, and Ioannis Matsinos. Identifying wildland re ignition factors through sensitivity analysis of a neural network. Natural Hazards, 50(1):125{143, jul 2009. ISSN 0921-030X. doi: 10.1007/s11069-008-9326-3. URL http://link.springer.com/10.1007/ s11069-008-9326-3. Daniel Vec n-Arias, Fernando Castedo-Dorado, Celestino Ord onez, ~ and Jos e Ram on Rodr guez-P erez. Bio- physical and lightning characteristics drive lightning-induced re occurrence in the central plateau of the Iberian Peninsula. Agricultural and Forest Meteorology, 225:36{47, sep 2016. ISSN 0168- 1923. doi: 10.1016/J.AGRFORMET.2016.05.003. URL https://www.sciencedirect.com/science/ article/pii/S0168192316302593. C. Vega-Garcia, B.S. Lee, P.M. Woodard, and S.J. Titus. Applying neural network technology to human- caused wild re occurrence prediction. AI Applications, pages 9{18, 1996. URL https://cfs.nrcan. gc.ca/publications?id=18949. O Viedma, L A Arroyo, R Mateo, A De Santis, and J M Moreno. EFFECTS OF ENVIRONMEN- TAL PROPERTIES, BURNING CONDITIONS AND HUMAN-RELATED VARIABLES ON FIRE SEVERITY DERIVED FROM LANDSAT TM IMAGES FOR A LARGE FIRE IN CENTRAL SPAIN. In Advances in Remote Sensing and GIS applications in Forest Fire Management From local to global assessments, page 157. 2011. Olga Viedma, Juan Quesada, Iv an Torres, Angela De Santis, and Jos e M. Moreno. Fire Severity in a Large Fire in a Pinus pinaster Forest is Highly Predictable from Burning Conditions, Stand Structure, and Topography. Ecosystems, 18(2):237{250, mar 2015. ISSN 1432-9840. doi: 10.1007/s10021-014-9824-y. URL http://link.springer.com/10.1007/s10021-014-9824-y. Domingos Xavier Viegas, Karin L. Riley, Isaac C. Grenfell, Mark A. Finney, and Nicholas L. Crookston. Utilizing random forests imputation of forest plot data for landscape-level wild- re analyses. Imprensa da Universidade de Coimbra, Coimbra, 2014. ISBN 978-989-26-0884- 6 (PDF). doi: 10.14195/978-989-26-0884-6 67. URL https://digitalis.uc.pt/en/livro/ utilizing{_}random{_}forests{_}imputation{_}forest{_}plot{_}data{_}landscape{_}level{_}wildfire{_}analyses. 67 Dinesh Babu Irulappa Pillai Vijayakumar, Fr ed eric Raulier, Pierre Y. Bernier, Sylvie Gauthier, Yves Bergeron, and David Pothier. Lengthening the historical records of re history over large areas of boreal forest in eastern Canada using empirical relationships. Forest Ecology and Management, 347:30{39, jul 2015. ISSN 0378-1127. doi: 10.1016/J.FORECO.2015.03.011. URL https://www.sciencedirect.com/ science/article/pii/S0378112715001310{#}b0140. Dinesh Babu Irulappa Pillai Vijayakumar, Fr ed eric Raulier, Pierre Bernier, David Par e, Sylvie Gauthier, Yves Bergeron, and David Pothier. Cover density recovery after re disturbance controls landscape aboveground biomass carbon in the boreal forest of eastern Canada. Forest Ecology and Management, 360:170{180, jan 2016. ISSN 0378-1127. doi: 10.1016/J.FORECO.2015.10.035. URL https://www. sciencedirect.com/science/article/pii/S0378112715005927?via{%}3Dihub. Lara Vilar, Israel G omez, Javier Mart nez-Vega, Pilar Echavarr a, David Rian ~o, and M. Pilar Mart n. Multitemporal Modelling of Socio-Economic Wild re Drivers in Central Spain between the 1980s and the 2000s: Comparing Generalized Linear Models to Machine Learning Algorithms. PLOS ONE, 11(8): e0161344, aug 2016. ISSN 1932-6203. doi: 10.1371/journal.pone.0161344. URL http://dx.plos.org/ 10.1371/journal.pone.0161344. Yan Wang, Chunyu Yu, Ran Tu, and Yongming Zhang. Fire detection model in Tibet based on grey-fuzzy neural network algorithm. Expert Systems with Applications, 38(8):9580{9586, aug 2011. ISSN 0957- 4174. doi: 10.1016/J.ESWA.2011.01.163. URL https://www.sciencedirect.com/science/article/ pii/S0957417411001965. Yuanbin Wang, Langfei Dang, and Jieying Ren. Forest re image recognition based on convolutional neural network. Journal of Algorithms & Computational Technology, 13:174830261988768, jan 2019. ISSN 1748-3026. doi: 10.1177/1748302619887689. Gregory L. Watson, Donatello Telesca, Colleen E. Reid, Gabriele G. P ster, and Michael Jerrett. Machine learning models accurately predict ozone exposure during wild re events. Environmental Pollution, 254: 112792, nov 2019. ISSN 18736424. doi: 10.1016/j.envpol.2019.06.088. David H Wolpert. The lack of a priori distinctions between learning algorithms. Neural computation, 8(7): 1341{1390, 1996. Zhiwei Wu, Hong S. He, Jian Yang, Zhihua Liu, and Yu Liang. Relative e ects of climatic and local factors on re occurrence in boreal forest landscapes of northeastern China. Science of The Total Environment, 493:472{480, sep 2014. ISSN 0048-9697. doi: 10.1016/J.SCITOTENV.2014.06.011. URL https://www.sciencedirect.com/science/article/pii/S0048969714008547. Zhiwei Wu, Hong S. He, Jian Yang, and Yu Liang. De ning re environment zones in the boreal forests of northeastern China. Science of The Total Environment, 518-519:106{116, jun 2015. ISSN 0048-9697. doi: 10.1016/J.SCITOTENV.2015.02.063. URL https://www.sciencedirect.com/science/article/ pii/S0048969715002065. Dexen D Z Xi, Stephen W Taylor, Douglas G Woolford, and C B Dean. Statistical Models of Key Components of Wild re Risk. 2019. doi: 10.1146/annurev-statistics-031017-100450. URL www.annualreviews.org. Dao Wen Xie and Shi Liang Shi. Prediction for Burned Area of Forest Fires Based on SVM Model. Applied Mechanics and Materials, 513-517:4084{4089, feb 2014. ISSN 1662-7482. doi: 10.4028/www.scienti c. net/AMM.513-517.4084. URL https://www.scientific.net/AMM.513-517.4084. Ying Xie and Minggang Peng. Forest re forecasting using ensemble learning approaches. Neural Computing and Applications, 31(9):4541{4550, sep 2019. ISSN 0941-0643. doi: 10.1007/s00521-018-3515-0. URL http://link.springer.com/10.1007/s00521-018-3515-0. 68 Jiayun Yao, Michael Brauer, Sean Ra use, and Sarah B. Henderson. Machine Learning Approach To Estimate Hourly Exposure to Fine Particulate Matter for Urban, Rural, and Remote Populations during Wild re Seasons. Environmental Science & Technology, 52(22):13239{13249, nov 2018a. ISSN 0013- 936X. doi: 10.1021/acs.est.8b01921. URL http://pubs.acs.org/doi/10.1021/acs.est.8b01921. Jiayun Yao, Sean M. Ra use, Michael Brauer, Grant J. Williamson, David M.J.S. Bowman, Fay H. Johnston, and Sarah B. Henderson. Predicting the minimum height of forest re smoke within the atmosphere using machine learning and data from the CALIPSO satellite. Remote Sensing of Environment, 206:98{106, mar 2018b. ISSN 0034-4257. doi: 10.1016/J.RSE.2017.12.027. URL https://www.sciencedirect.com/science/article/abs/pii/S003442571730603X. Lingxiao Ying, Jie Han, Yongsheng Du, and Zehao Shen. Forest re characteristics in China: Spatial pat- terns and determinants with thresholds. Forest Ecology and Management, 424:345{354, sep 2018. ISSN 0378-1127. doi: 10.1016/J.FORECO.2018.05.020. URL https://www.sciencedirect.com/science/ article/pii/S0378112717317668. Adam M. Young, Philip E. Higuera, Paul A. Du y, and Feng Sheng Hu. Climatic thresholds shape northern high-latitude re regimes and imply vulnerability to future climate change. Ecography, 40(5):606{617, may 2017. ISSN 16000587. doi: 10.1111/ecog.02205. Adam M. Young, Philip E. Higuera, John T. Abatzoglou, Paul A. Du y, and Feng Sheng Hu. Consequences of climatic thresholds for projecting re activity and ecological change. Global Ecology and Biogeography, 28(4):521{532, apr 2019. ISSN 14668238. doi: 10.1111/geb.12872. Bo Yu, Fang Chen, Bin Li, Li Wang, and Mingquan Wu. Fire Risk Prediction Using Remote Sensed Products: A Case of Cambodia. Photogrammetric Engineering & Remote Sensing, 83(1):19{25, jan 2017. ISSN 0099-1112. doi: 10.14358/PERS.83.1.19. URL http://www.ingentaconnect.com/content/10. 14358/PERS.83.1.19. Yong Poh Yu, Rosli Omar, Rhett D Harrison, Mohan Kumar Sammathuria, and Abdul Rahim Nik. Pattern clustering of forest res based on meteorological variables and its classi cation using hybrid data mining methods. Journal of Computational Biology and Bioinformatics Research, 3(4):47{52, 2011. URL http: //www.academicjournals.org/jcbbr. Jie Yuan, Lidong Wang, Peng Wu, Chao Gao, and Lingqing Sun. Detection of Wild res along Transmission Lines Using Deep Time and Space Features. Pattern Recognition and Image Analysis, 28(4):805{812, oct 2018. ISSN 15556212. doi: 10.1134/S1054661818040168. Bianca Zadrozny. Learning and evaluating classi ers under sample selection bias. In Twenty- rst interna- tional conference on Machine learning - ICML '04, page 114, New York, New York, USA, 2004. ACM Press. ISBN 1581138285. doi: 10.1145/1015330.1015425. URL http://portal.acm.org/citation. cfm?doid=1015330.1015425. Harold S.J. Zald and Christopher J. Dunn. Severe re weather and intensive forest management increase re severity in a multi-ownership landscape. Ecological Applications, 28(4):1068{1080, 2018. ISSN 19395582. doi: 10.1002/eap.1710. Olivier Zammit, Xavier Descombes, and Josiane Zerubia. Burnt area mapping using Support Vector Machines. Forest Ecology and Management, 234:S240, 2006. ISSN 03781127. doi: 10.1016/j.foreco.2006. 08.269. URL http://linkinghub.elsevier.com/retrieve/pii/S0378112706008097. Bin Zhang, Wei Wei, Bingqian He, and Chuanlei Guo. Early wild re smoke detection based on improved codebook model and convolutional neural networks. In Xudong Jiang and Jenq- Neng Hwang, editors, Tenth International Conference on Digital Image Processing (ICDIP 69 2018), page 120. SPIE, aug 2018a. ISBN 9781510621992. doi: 10.1117/12.2502974. URL https://www.spiedigitallibrary.org/conference-proceedings-of-spie/10806/2502974/ Early-wildfire-smoke-detection-based-on-improved-codebook-model-and/10.1117/12. 2502974.full. Guoli Zhang, Ming Wang, and Kai Liu. Forest Fire Susceptibility Modeling Using a Convolutional Neural Network for Yunnan Province of China. International Journal of Disaster Risk Science, 10(3):1{18, sep 2019. ISSN 2095-0055. doi: 10.1007/s13753-019-00233-1. URL http://link.springer.com/10.1007/ s13753-019-00233-1. Qi Xing Zhang, Gao Hua Lin, Yong Ming Zhang, Gao Xu, and Jin Jun Wang. Wildland Forest Fire Smoke Detection Based on Faster R-CNN using Synthetic Smoke Images. In Procedia Engineering, volume 211, pages 441{446. Elsevier Ltd, 2018b. doi: 10.1016/j.proeng.2017.12.034. Qingjie Zhang, Jiaolong Xu, Liang Xu, and Haifeng Guo. Deep Convolutional Neural Networks for Forest Fire Detection. Atlantis Press, 2016. doi: 10.2991/ifmeita-16.2016.105. Zhanqing Li, A. Khananian, R.H. Fraser, and J. Cihlar. Automatic detection of re smoke using arti cial neural networks and threshold approaches applied to AVHRR imagery. IEEE Transactions on Geoscience and Remote Sensing, 39(9):1859{1870, 2001. ISSN 01962892. doi: 10.1109/36.951076. URL http: //ieeexplore.ieee.org/document/951076/. Feng Zhao, Chengquan Huang, and Zhiliang Zhu. Use of Vegetation Change Tracker and Support Vector Machine to Map Disturbance Types in Greater Yellowstone Ecosystems in a 1984{2010 Landsat Time Series. IEEE Geoscience and Remote Sensing Letters, 12(8):1650{1654, aug 2015. ISSN 1545-598X. doi: 10.1109/LGRS.2015.2418159. URL http://ieeexplore.ieee.org/document/7088596/. Jianhui Zhao, Zhong Zhang, Shizhong Han, Chengzhang Qu, Zhiyong Yuan, and Dengyi Zhang. SVM based forest re detection using static and dynamic features. Computer Science and Information Systems, 8 (3):821{841, 2011. ISSN 1820-0214. doi: 10.2298/csis101012030z. Yi Zhao, Jiale Ma, Xiaohui Li, and Jie Zhang. Saliency Detection and Deep Learning-Based Wild re Identi cation in UAV Imagery. Sensors, 18(3):712, feb 2018. ISSN 1424-8220. doi: 10.3390/s18030712. URL http://www.mdpi.com/1424-8220/18/3/712. Zhong Zheng, Wei Huang, Songnian Li, and Yongnian Zeng. Forest re spread simulating model us- ing cellular automaton with extreme learning machine. Ecological Modelling, 348(May 2018):33{43, 2017. ISSN 03043800. doi: 10.1016/j.ecolmodel.2016.12.022. URL http://dx.doi.org/10.1016/j. ecolmodel.2016.12.022. Yufei Zou, Susan M. O'Neill, Narasimhan K. Larkin, Ernesto C. Alvarado, Robert Solomon, Cli ord Mass, Yang Liu, M. Talat Odman, and Huizhong Shen. Machine Learning-Based Integration of High-Resolution Wild re Smoke Simulations and Observations for Regional Health Impact Assessment. International Journal of Environmental Research and Public Health, 16(12):2137, jun 2019. ISSN 1660-4601. doi: 10.3390/ijerph16122137. URL https://www.mdpi.com/1660-4601/16/12/2137. K. Zwirglmaier, P. Papakosta, and D. Straub. Learning a Bayesian network model for predicting wild re behavior. In ICOSSAR 2013, 2013. 70 Supplementary Material This supplemental contains all papers identi ed in this review with ML applications for wild re science and management, organized by problem domains. Note that some papers are repeated in multiple problem domains. S.1. Fuels Characterization, Fire Detection And Mapping S.1.1 Fuels characterization Citation ML methods used Study Region Riano ~ et al. [2005] ANN Not speci ed Garc a et al. [2011] SVM Alto Tajo Natural Park, central Spain Pierce et al. [2012] RF Lassen Volcanic National Park, Cal- ifornia, USA Chirici et al. [2013] DT, RF, BRT Sicily, Italy Viegas et al. [2014] RF Eastern Oregon, USA L opez-Serrano et al. SVM, KNN, RF Sierra Madre Occidental, Mexico [2016] S.1.2 Fire detection Citation ML methods used Study Region Arrue et al. [2000] ANN Experiments at University of Seville Al-Rawi et al. [2001] ANN Eastern Spain Zhanqing Li et al. [2001] ANN Canada Utkin et al. [2002] ANN Not speci ed Cordoba et al. [2004] GA Experiments Fernandes et al. [2004a] ANN Not speci ed Fernandes et al. [2004b] ANN Not speci ed Srinivasa et al. [2008] KM Not speci ed Angayarkkani and Rad- ANN Not speci ed hakrishnan [2010] Ko et al. [2010] BN test images Soliman et al. [2010] ANN Laboratory experiments Angayarkkani and Rad- ANFIS not speci ed hakrishnan [2011] Wang et al. [2011] ANFIS Tibet Zhao et al. [2011] SVM, GMM Test images Li et al. [2015] ANN China, North East Asia, Russia, Canada, Australia Liu et al. [2015] ANN Laboratory experiments Zhang et al. [2016] CNN, SVM Test images Akhlou et al. [2018] CNN Corsica Li et al. [2018b] CNN Test images Continued on next page 71 Citation ML methods used Study Region Muhammad et al. [2018] CNN Test images Yuan et al. [2018] CNN Test images Zhang et al. [2018a] CNN Test images Zhang et al. [2018b] CNN Synthetic data Zhao et al. [2018] CNN Test images Alexandrov et al. [2019] CNN, Test images HAAR CASCADES, YOLO Ba et al. [2019] CNN Satellite test images Barmpoutis et al. [2019] CNN Test images Cao et al. [2019] CNN, LSTM Test images and video Hossain et al. [2019] ANN Test images Jakubowski et al. [2019] CNN Test images Jo~ ao Sousa et al. [2019] CNN Corsica Li et al. [2019] CNN Test images Phan and Nguyen CNN American Continent [2019] Sayad et al. [2019] ANN, SVM Canada Wang et al. [2019] CNN Test images S.1.3 Fire perimeter and severity mapping Citation ML methods used Study Region Al-Rawi et al. [2001] ANN Eastern Spain Brumby et al. [2001] GA Cerro Grande Fire, New Mexico, USA Sunar and Ozkan [2001] ANN, ISODATA south coast of Turkey Al-Rawi et al. [2002] ANN Valencia, Spain S a et al. [2003] DT, BAG Northern Mozambique Pu and Gong [2004] ANN Northern California, USA Zammit et al. [2006] SVM, KM, KNN Southern France Alonso-Benito et al. SVM Tenerife and Gran Canaria [2008] Cao et al. [2009] SVM Mongolia and China Celik [2010] GA Reno Lake Tahoe area, Nevada, USA Petropoulos et al. [2010] SVM Greece E. Dragozi, I. Z. Gi- SVM, KNN Stresa, Italy tas, D.G. Stavrakoudis [2011] G omez and Pilar ANN Iberian Peninsula Mart n [2011] Petropoulos et al. [2011] SVM Greece Mitrakis et al. [2012] NFM, ANN, SVM, AD- Greece ABOOST Dragozi et al. [2014] SVM Parnitha and Rhodes, Greece Continued on next page 72 Citation ML methods used Study Region Hultquist et al. [2014] GP, RF, SVM Big Sur, California, USA Zhao et al. [2015] SVM Greater Yellowstone Ecosystem, USA Hamilton et al. [2017] SVM Idaho, USA Hawbaker et al. [2017] BRT USA Pereira et al. [2017] SVM Cerrado savanna, Brazil Collins et al. [2018] RF Victoria, Australia Nitze et al. [2018] RF Alaska, Eastern Canada, Western Siberia and Ea... Crowley et al. [2019] BULC Elephant Hill Fire, British Columbia, Canada Langford et al. [2019] DNN Interior Alaska, USA Quintano et al. [2019] MAXENT La Cabrera, Spain Quintero et al. [2019] RF West Central Spain S.2. Fire Weather And Climate Change S.2.1 Fire weather prediction Citation ML methods used Study Region Skinner et al. [2002] KM Canada Crimmins [2006] SOM Southwest USA Sanabria et al. [2013] RF Australia Lagerquist et al. [2017] SOM Alberta, Canada Nauslar et al. [2019] SOM Southwest USA S.2.2 Lightning prediction Citation ML methods used Study Region Blouin et al. [2016] RF Alberta, Canada Bates et al. [2017] DT, RF Australia S.2.3 Climate change Citation ML methods used Study Region Moritz et al. [2012] MAXENT Global Amatulli et al. [2013] RF Mediterranean Europe Batllori et al. [2013] MAXENT Mediterranen ecosystems, Global Liu and Wimberly BRT, RF Western USA [2016] Parks et al. [2016] BRT Western USA van Breugel et al. [2016] RF, SVM, BRT, MAX- Ethiopia ENT, ANN, DT Continued on next page 73 Citation ML methods used Study Region Davis et al. [2017] MAXENT Paci c Northwest, USA Li et al. [2017] MAXENT Yunnan Province, Southwest China Young et al. [2017] BRT Alaska, USA Boulanger et al. [2018] RF, BRT, DT Canada Stralberg et al. [2018] RF Alberta, Canada Stroh et al. [2018] MAXENT South central USA Tracy et al. [2018] MAXENT Western North America Buckland et al. [2019] ANN Nebraska, USA Young et al. [2019] BRT Alaska, USA S.3. Fire Occurrence, Susceptibility and Risk S.3.1 Fire occurrence prediction Citation ML methods used Study Region Vega-Garcia et al. ANN Alberta, Canada [1996] Alonso-Betanzos et al. ANN Galicia, Northwest Spain [2002] Alonso-Betanzos et al. ANN Galicia, Northwest Spain [2003] Vasilakos et al. [2007] ANN Lesvos Island, Greece Sakr et al. [2010] SVM Lebanon Sakr et al. [2011] SVM, ANN Lebanon Stojanova et al. [2012] KNN, NB, DT, SVM, BN, Slovenia ADABOOST, BAG, RF Dutta et al. [2013] ANN, DNN Australia Chen et al. [2015] MAXENT Daxinganling Mountains, North- eastern China De Angelis et al. [2015] MAXENT Canton Ticino, Switzerland Dutta et al. [2016] DNN Australia Vec n-Arias et al. [2016] RF Central Iberian Peninsula, Spain Cao et al. [2017] ANN, RF Yunnan Province, China Yu et al. [2017] RF Cambodia Van Beusekom et al. RF Puerto Rico [2018] S.3.2 Landscape-scale Burned area prediction Citation ML methods used Study Region Cheng and Wang [2008] RNN Canada Archibald et al. [2009] RF Southern Africa Arnold et al. [2014] HCL, MAXENT Interior Western USA Mayr et al. [2018] DT, RF, SVM, KM Namibia Continued on next page 74 Citation ML methods used Study Region de Bem et al. [2018] ANN, GA Federal District, Brazil S.3.3 Fire Susceptibility Mapping Citation ML methods used Study Region Chuvieco et al. [1999] ANN Mediterranean Europe De Vasconcelos et al. ANN central Portugal [2001] Amatulli et al. [2006] DT Gargano Peninsula, Italy Amatulli and Camia DT Tuscany, Italy [2007] Lozano et al. [2008] DT Northwestern Spain Holden et al. [2009] RF Gila National Forest, New Mexico, USA Maeda et al. [2009] ANN Brazil Mallinis et al. [2009] DT, KM Greece Parisien and Moritz MAXENT, BRT USA [2009] Barrett et al. [2011] RF Alaska Dimuccio et al. [2011] ANN Central Portugal Dlamini [2011] BN Swaziland Bisquert et al. [2012] ANN Galicia, Northwest Spain Moritz et al. [2012] MAXENT Global Oliveira et al. [2012] RF Mediterranean Europe Parisien et al. [2012] MAXENT Western USA Renard et al. [2012] MAXENT Western Ghats, India Syphard et al. [2012] MAXENT Southern California, USA Bar Massada et al. RF, MAXENT Michigan, USA [2013] Luo et al. [2013] RF global Peters et al. [2013] MAXENT Northeast USA Syphard et al. [2013] MAXENT South coast ecoregion, San Diego County, USA Arpaci et al. [2014] RF, MAXENT Tyrol, European Alps Parisien et al. [2014] DT Canada Rodrigues and De la RF, BRT, SVM Peninsular Spain Riva [2014] Duane et al. [2015] MAXENT Catalonia, Spain Bashari et al. [2016] BN Isfahan province, Iran Curt et al. [2016] BRT Southeastern France Fonseca et al. [2016] MAXENT Brazil Goldarag et al. [2016] ANN Golestan province, Northern Iran Guo et al. [2016b] RF Daxing'an Mountains, Northeast China Guo et al. [2016a] RF Fujian province, China Pourtaghi et al. [2016] BRT, RF Golestan province, Northern Iran Continued on next page 75 Citation ML methods used Study Region Satir et al. [2016] ANN Upper Seyhan Basin, Turkey van Breugel et al. [2016] RF, SVM, BRT, MAX- Ethiopia ENT, ANN, DT Vilar et al. [2016] MAXENT Madrid region, Spain Adab [2017] ANN Northeast Iran Cao et al. [2017] ANN, RF Yunnan Province, China Davis et al. [2017] MAXENT Paci c Northwest, USA Ebrahimy et al. [2017] MAXENT Eastern Azerbaijan Li et al. [2017] MAXENT Yunnan Province, Southwest China MOSTAFA et al. [2017] SVM Golestan province, Northern Iran Peters and Iverson MAXENT Northeast USA [2017] Song et al. [2017] RF Hefei City, China Tien Bui et al. [2017] NFM, PSO, RF, SVM Lam Dong province, Vietnam Valdez et al. [2017] RF Honduras Adab et al. [2018] MAXENT Mazandaran province, Iran Hong et al. [2018] SVM, RF, GA Southwest Jiangxi Province Jaafari et al. [2018] DT, DT, NB Zagros Mountains, Iran Kahiu and Hanan [2018] BRT sub Saharan Africa Leuenberger et al. RF, ANN Dao, Lafoes, Portugal [2018] Ngoc Thach et al. [2018] SVM, RF, ANN Thuan Chau district, Vietnam Parks et al. [2018] BRT Western USA Sachdeva et al. [2018] BRT, ANN, RF, SVM, Northern India DT, NB, GA, PSO Tehrany et al. [2018] LB, SVM, RF Lao Cai Province, Vietnam Tracy et al. [2018] MAXENT Western North America Vacchiano et al. [2018] MAXENT Aosta Valley, Northern Italy Fernandez-Manso et al. MAXENT Valencia, Spain [2019] Ghorbanzadeh et al. ANN Mazandaran Province, Northern [2019a] Iran Ghorbanzadeh et al. ANN, SVM, RF Mazandaran Province, Northern [2019b] Iran Gigovi c et al. [2019] SVM, RF Tara National Park, Serbia Jaafari [2019] RF, SVM Zagros Mountains, Iran Jaafari et al. [2019] NFM, GA Hyrcanian ecoregion, Iran Kim et al. [2019] MAXENT, RF South Korea Lim et al. [2019] MAXENT South Korea Mart n et al. [2019] MAXENT Northeast Spain Mpakairi et al. [2019] MAXENT northwestern Zimbabwe Quintano et al. [2019] MAXENT La Cabrera, Spain Rihan et al. [2019] RF Mongolian Plateau Syphard et al. [2019] MAXENT California, USA Zhang et al. [2019] CNN, RF, SVM, ANN Yunnan Province, China S.3.4 Landscape controls on re 76 Citation ML methods used Study Region Amatulli et al. [2006] DT Gargano Peninsula, Italy Amatulli and Camia DT Tuscany, Italy [2007] Archibald et al. [2009] RF Southern Africa Holden et al. [2009] RF Gila National Forest, New Mexico, USA Li et al. [2009] ANN Japan Parisien and Moritz MAXENT, BRT USA [2009] Vasilakos et al. [2009] ANN Lesvos Island, Greece Dlamini [2010] BN Swaziland Aldersley et al. [2011] DT, RF global Dimuccio et al. [2011] ANN Central Portugal Sitanggang and Ismail DT Rokan Hilir District, Sumatra, In- [2011] donesia Viedma et al. [2011] BRT, DT Guadalajara province, Central Spain Bisquert et al. [2012] ANN Galicia, Northwest Spain Moritz et al. [2012] MAXENT Global Oliveira et al. [2012] RF Mediterranean Europe Parisien et al. [2012] MAXENT Western USA Renard et al. [2012] MAXENT Western Ghats, India Syphard et al. [2012] MAXENT Southern California, USA Bar Massada et al. RF, MAXENT Michigan, USA [2013] Batllori et al. [2013] MAXENT Mediterranen ecosystems, Global Liu et al. [2013] BRT Great Xing'an Mountains, North- eastern China Luo et al. [2013] RF global Peters et al. [2013] MAXENT Northeast USA Sitanggang et al. [2013] DT Rokan Hilir District, Sumatra, In- donesia Syphard et al. [2013] MAXENT South coast ecoregion, San Diego County, USA Arpaci et al. [2014] RF, MAXENT Tyrol, European Alps Lydersen et al. [2014] RF central Sierra Nevada, California, USA Maxwell et al. [2014] RF Parisien et al. [2014] DT Canada Rodrigues and De la RF, BRT, SVM Peninsular Spain Riva [2014] Wu et al. [2014] RF, DT Great Xing'an Mountains, China Arganaraz ~ et al. [2015] BRT Central Argentina Chen et al. [2015] MAXENT Daxinganling Mountains, North- eastern China Continued on next page 77 Citation ML methods used Study Region Chingono and Mbohwa MAXENT Southern Africa [2015] Curt et al. [2015] BRT New Caledonia Duane et al. [2015] MAXENT Catalonia, Spain Kane et al. [2015] RF Sierra Nevada, California, USA Liu and Wimberly BRT Western USA [2015] Parks et al. [2015] BRT Western USA Viedma et al. [2015] BRT Guadalajara province, Central Spain Vijayakumar et al. RF central Quebec, Canada [2015] Wu et al. [2015] KNN, DT Great Xing'an Mountains, China Bashari et al. [2016] BN Isfahan province, Iran Coppoletta et al. [2016] RF, DT Northern Sierra Nevada, California, USA Curt et al. [2016] BRT Southeastern France Fernandes et al. [2016] BRT Portugal Fonseca et al. [2016] MAXENT Brazil Goldarag et al. [2016] ANN Golestan province, Northern Iran Guo et al. [2016b] RF Daxing'an Mountains, Northeast China Guo et al. [2016a] RF Fujian province, China Miquelajauregui et al. RF, DT central Quebec, Canada [2016] Pourtaghi et al. [2016] BRT, RF Golestan province, Northern Iran Satir et al. [2016] ANN Upper Seyhan Basin, Turkey Syphard et al. [2016] MAXENT South coast ecoregion, San Diego county, USA Vilar et al. [2016] MAXENT Madrid region, Spain Adab [2017] ANN Northeast Iran Cao et al. [2017] ANN, RF Yunnan Province, China Davis et al. [2017] MAXENT Paci c Northwest, USA Dwomoh and Wimberly BRT Upper Guinean Region, West Africa [2017] Ebrahimy et al. [2017] MAXENT Eastern Azerbaijan Forkel et al. [2017] RF, GA global Harris and Taylor [2017] RF Sierra Nevada, California, USA Leys et al. [2017] RF Central Great Plains, USA Li et al. [2017] MAXENT Yunnan Province, Southwest China Lydersen et al. [2017] RF central Sierra Nevada, California, USA MOSTAFA et al. [2017] SVM Golestan province, Northern Iran Nelson et al. [2017] DT, BRT, RF British Columbia, Canada Peters and Iverson MAXENT Northeast USA [2017] Song et al. [2017] RF Hefei City, China Continued on next page 78 Citation ML methods used Study Region Tien Bui et al. [2017] NFM, PSO, RF, SVM Lam Dong province, Vietnam Valdez et al. [2017] RF Honduras Young et al. [2017] BRT Alaska, USA Adab et al. [2018] MAXENT Mazandaran province, Iran Fang et al. [2018] BRT Great Xing'an Mountains, China Hong et al. [2018] SVM, RF, GA Southwest Jiangxi Province Jaafari et al. [2018] DT, DT, NB Zagros Mountains, Iran Kahiu and Hanan [2018] BRT sub Saharan Africa Masrur et al. [2018] RF circumpolar arctic Mayr et al. [2018] DT, RF, SVM, KM Namibia Parks et al. [2018] BRT Western USA Tehrany et al. [2018] LB, SVM, RF Lao Cai Province, Vietnam Tracy et al. [2018] MAXENT Western North America Vacchiano et al. [2018] MAXENT Aosta Valley, Northern Italy Ying et al. [2018] RF China Clarke et al. [2019] MAXENT Southeast Australia Fernandez-Manso et al. MAXENT Valencia, Spain [2019] Forkel et al. [2019] RF global Garc a-Llamas et al. RF Northwest Spain [2019] Ghorbanzadeh et al. ANN Mazandaran Province, Northern [2019a] Iran Ghorbanzadeh et al. ANN, SVM, RF Mazandaran Province, Northern [2019b] Iran Gigovi c et al. [2019] SVM, RF Tara National Park, Serbia Jaafari [2019] RF, SVM Zagros Mountains, Iran Jaafari et al. [2019] NFM, GA Hyrcanian ecoregion, Iran Kim et al. [2019] MAXENT, RF South Korea Mansuy et al. [2019] MAXENT North America Molina et al. [2019] MAXENT Andalusia, southern Spain Mpakairi et al. [2019] MAXENT northwestern Zimbabwe Rihan et al. [2019] RF Mongolian Plateau Syphard et al. [2019] MAXENT California, USA Young et al. [2019] BRT Alaska, USA Zhang et al. [2019] CNN, RF, SVM, ANN Yunnan Province, China S.4. Fire Behaviour Prediction S.4.1 Fire Spread and Growth Citation ML methods used Study Region Vakalis et al. [2004] ANN Attica region, Greece Abdalhaq et al. [2005] GA not speci ed Rodriguez et al. [2008] GA Catalonia, Spain Continued on next page 79 Citation ML methods used Study Region Markuzon and Kolitz RF, BN, KNN Southwest USA [2009] Rodr guez et al. [2009] GA Catalonia, Spain Cencerrado et al. [2012] GA Ashley National Forest, Utah, USA Denham et al. [2012] GA Gestosa, Portugal Cencerrado et al. [2013] GA Catalonia, Spain Kozik et al. [2013] ANN not speci ed Art es et al. [2014] GA Northeast Spain Cencerrado et al. [2014] GA Catalonia, Spain Kozik et al. [2014] RNN not speci ed Ascoli et al. [2015] GA Southern Europe Chetehouna et al. [2015] ANN Experimental data Art es et al. [2016] GA Northeast Catalonia, Spain Carrillo et al. [2016] GA Arkadia region, Greece Art es et al. [2017] GA Northeast Catalonia, Spain Subramanian and RL Northern Alberta, Canada Crowley [2017] Zheng et al. [2017] ANN western USA Denham and Laneri GA Northern Patagonia Andean region [2018] Subramanian and RL Northern Alberta, Canada Crowley [2018] Hodges and Lattimer CNN California, USA [2019] Khakzad [2019] BN Canada Radke et al. [2019] CNN USA S.4.2 Burned area and re severity prediction Citation ML methods used Study Region Cortez and Morais DT, RF, ANN, SVM Montesinho natural park, Portugal [2007] Yu et al. [2011] SOM, ANN Montesinho natural park, Portugal Ozbayo glu and Bozer ANN, SVM Turkey [2012] Sa and Bouroumi ANN Montesinho natural park, Portugal [2013] Zwirglmaier et al. [2013] BN Cyprus Shidik and Mustofa ANN, NB, DT, RF, KNN, Montesinho natural park, Portugal [2014] SVM Xie and Shi [2014] SVM Guangzhou City area, China Alberg [2015] DT Montesinho natural park, Portugal Castelli et al. [2015] GA Montesinho natural park, Portugal Naganathan et al. SVM, DT, KNN USA [2016] Storer and Green [2016] ANN, PSO Montesinho natural park, Portugal Continued on next page 80 Citation ML methods used Study Region Mitsopoulos and Malli- BRT, RF Greece nis [2017] Al Janabi et al. [2018] ANN, SVM Montesinho natural park, Portugal Li et al. [2018a] DT Montesinho natural park, Portugal Toujani et al. [2018] HMM, SOM northwest Tunisia Zald and Dunn [2018] RF southwest Oregon, USA Coeld et al. [2019] DT, RF, ANN, KNN, BRT Alaska, USA Liang et al. [2019] ANN, RNN, LSTM Alberta, Canada Xie and Peng [2019] BRT, DNN, DT, SVM, Montesinho Natural Park, Portugal ANN, RF S.5. Fire E ects S.5.1 Soil erosion and deposits Citation ML methods used Study Region Mallinis et al. [2009] DT, KM Greece Buckland et al. [2019] ANN Nebraska, USA Quintano et al. [2019] MAXENT La Cabrera, Spain S.5.2 Smoke and particulate levels Citation ML methods used Study Region Reid et al. [2015] RF, BAG, BRT, SVM, GP, California, USA KNN Lozhkin et al. [2016] ANN Irkutsk Region, Russia Yao et al. [2018b] RF British Columbia, Canada Yao et al. [2018a] RF British Columbia, Canada Fuentes et al. [2019] ANN South Australia Watson et al. [2019] BRT, RF, SVM, KNN, Northern California, USA ANN Zou et al. [2019] BRT, RF Paci c Northwest, USA S.5.3 Post- re regeneration and ecology Citation ML methods used Study Region Johnstone et al. [2010] BRT Alaska, USA Thompson and Spies RF, DT northwest California, southwest [2010] Oregon, USA Barrett et al. [2011] RF Alaska Perry et al. [2012] BRT New Zealand Reside et al. [2012] MAXENT Australia Continued on next page 81 Citation ML methods used Study Region Sherrill and Romme BRT Dinosaur National Monument, USA [2012] Cai et al. [2013] BRT Huzhong National Reserve, China Debouk et al. [2013] ANN Catalonia, Spain Jung et al. [2013] GA, RF Central Siberian Plateau, Russia Chapin et al. [2014] RF Alaska, USA Liu and Yang [2014] BRT Huzhong Natural Reserve, China Han et al. [2015] RF Yunnan province, China Hermosilla et al. [2015] RF Saskatchewan, Canada Divya and Vijayalak- NB India shmi [2016] Mart n-Alc on and Coll RF Catalonia, Spain [2016] Vijayakumar et al. RF Quebec, Canada [2016] Fairman et al. [2017] RF Victoria, Australia Luo et al. [2017] DT Cangshan Mountain, China Papakosta et al. [2017] BN Cyprus Jo~ ao et al. [2018] RF northern Portugal Poon et al. [2018] SVM San Bernardino, California, USA Cardil et al. [2019] BRT southwestern Europe Magadzire et al. [2019] MAXENT Cape Floristic Region, South Africa S.5.4 Socioeconomic e ects Citation ML methods used Study Region Hradsky et al. [2017] BN Otway Ranges, Australia S.6. Fire Management S.6.1 Planning and policy Citation ML methods used Study Region Penman et al. [2011] BN Wollemi National Park, Australia Bao et al. [2015] GA Longdong Forest Park, Guangzhou, China Ru ault and Mouillot [2015] Bradley et al. [2016] RF western USA McGregor et al. [2016] MDP not speci ed McGregor et al. [2017] RF Deschutes National Forest, Oregon, USA S.6.2 Fuel treatment 82 Citation ML methods used Study Region Penman et al. [2014] BN Southeastern Australia Arca et al. [2015] GA Southern Sardinia, Italy Lauer et al. [2017] RL southwest Oregon S.6.3 Wild re preparedness and response Citation ML methods used Study Region Homchaudhuri et al. GA not speci ed [2010] Costafreda-Aumedes ANN Spain et al. [2015] Penman et al. [2015] BN Sydney Basin Bioregion, Australia O'Connor et al. [2017] BRT, MAXENT Southern Idaho, Northern Nevada, USA Julian and Kochender- RL, DL simulation fer [2018b] Rodrigues et al. [2019] BN Catalonia, Spain S.6.4 Social factors Citation ML methods used Study Region Delgado et al. [2018] BN Spain

Journal

StatisticsarXiv (Cornell University)

Published: Mar 2, 2020

There are no references for this article.