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

Learn More →

Relation of influencing variables and weather conditions on rainfall partitioning by birch and pine trees

Relation of influencing variables and weather conditions on rainfall partitioning by birch and... J. Hydrol. Hydromech., 69, 2021, 4, 456–466 ©2021. This is an open access article distributed DOI: 10.2478/johh-2021-0023 under the Creative Commons Attribution ISSN 1338-4333 NonCommercial-NoDerivatives 4.0 License Relation of influencing variables and weather conditions on rainfall partitioning by birch and pine trees 1, 2 1* Katarina Zabret , Mojca Šraj University of Ljubljana, Faculty of Civil and Geodetic Engineering, Jamova 2, 1000 Ljubljana, Slovenia. Institute for Water of the Republic of Slovenia, Einspielerjeva 6, 1000 Ljubljana, Slovenia. Corresponding author. Tel.: +386 1 4768 684. E-mail: mojca.sraj@fgg.uni-lj.si Abstract: General weather conditions may have a strong influence on the individual elements of the hydrological cycle, an important part of which is rainfall interception. The influence of general weather conditions on this process was analysed, evaluating separately the influence of various variables on throughfall, stemflow, and rainfall interception for a wet (2014), a dry (2015), and an average (2016) year. The analysed data were measured for the case of birch and pine trees at a study site in the city of Ljubljana, Slovenia. The relationship between the components of rainfall partitioning and the influential variables for the selected years was estimated using two statistical models, namely boosted regression trees and random forest. The results of both implemented models complemented each other well, as both indicated the rainfall amount and the number of raindrops as the most influential variables. During the wet year 2014 rainfall duration seems to play an important role, correlating with the previously observed influence of the variables during the wetter leafless period. Similarly, during the dry year 2015, rainfall intensity had a significant influence on rainfall partitioning by the birch tree, again corresponding to the influences observed during the drier leafed period. Keywords: Throughfall; Stemflow; Rainfall interception; Rainfall microstructure; Boosted regression trees; Random forest. INTRODUCTION ecosystems, for which Hao et al. (2008) reported that both timing and frequency of rainfall events during the growing The hydrological cycle is altering due to climate change, as season significantly alter the capacity of steppe vegetation to differences in global redistribution of precipitation and varia- uptake CO . tions in seasonal precipitation patterns are observed (Inglezakis Forest ecosystems and trees in general also significantly in- et al., 2016). This results in a significant reduction of precipita- fluence the hydrological cycle through the process of rainfall tion in some parts of the world, while major variations in the interception (Dohnal et al., 2014; Klamerus-Iwan et al., 2020; timing and amount of precipitation per dry and wet season are Xu et al., 2013). Precipitation reaching the vegetation surface is expected elsewhere (Peng et al., 2021). The pronounced differ- distributed among the intercepted rainfall, which is captured by ences between the wet and dry periods significantly alter the the canopy and eventually evaporates back into to the atmos- water yield and the local water balance, the ecosystem services, phere, throughfall, which is described as the precipitation the water availability for vegetation, leading to changed occur- reaching the ground by dripping from the canopy or falling rences of floods and droughts (Bezak and Mikoš, 2014; Hun- directly to the ground through the gaps in the foliage, and gate and Hampton, 2012; Xu et al., 2020). stemflow, presenting the water flowing to the ground down the In the context of climate change, the relationship between branches and stems (Levia and Germer, 2015; Sadeghi et al., the water balance and vegetation in dry and wet periods is 2020; Staelens et al., 2008; Xiao et al., 2000; Yue et al., 2021; increasingly recognized. In this aspect, various influences of Zabret et al., 2018). Rainfall interception is influenced by vege- different vegetation systems were studied. Vegetation is an tation and meteorological characteristics. Vegetation character- important component, determining the ecosystem services, istics considered are mainly tree characteristics, such as the tree which were recognised to help mitigate the intensity of ex- height and surface area (e.g., projected tree canopy), smooth- tremely dry and wet conditions expected in the future (Peng et ness and absorbance of the bark, leaf area index, canopy cover- al., 2021). An important contribution to the ecosystem services age, and canopy storage capacity (Dohnal et al., 2014; Klamer- is also presented by the forest ecosystem affecting the global us-Iwan et al., 2020; Xu et al., 2013; Zabret, 2013). According carbon budget. The different response of a forest ecosystem in to the differences among the tree species, the different response wet and dry periods was analysed by Xiao et al. (2020), who of rainfall partitioning was analysed (Honda et al., 2014; concluded that in the dry season the precipitation generated Schooling and Carlyle-Moses, 2015). As characteristics of significantly positive effects to the cumulative CO emissions, some tree species (e.g., deciduous trees) are substantially influ- while the soil respiration rate was mainly influenced by the fine enced by the phenoseasons (presence and absence of leaves in root biomass regardless the season. An analysis of historical the tree canopy), the rainfall partitioning in leafed and leafless data from the tree rings was performed by Gao et al. (2020), period has also been frequently studied, mainly in relation to who observed that the growth of trees was improved by wet- the meteorological conditions (Brasil et al., 2020; Levia and ness, suggesting that tree growth is more sensitive to wetness Germer, 2015; Mużyło et al., 2012; Su et al., 2019; Zabret et than the forest coverage. Wetter conditions may, on the contra- al., 2018). Meteorological characteristics on the contrary ex- ry, reduce the carbon flux and evapotranspiration in steppe plain the characteristics of rainfall events, for example the 456 Relation of influencing variables and weather conditions on rainfall partitioning by birch and pine trees rainfall amount, duration and intensity, air temperature and in the east side there is a clearing. One group of trees in the humidity, vapour pressure deficit and wind conditions (Andre southern part consists of birch trees (Betula pendula Roth.), et al., 2008; Staelens et al., 2008; Zabret and Šraj, 2019a). which are on average 15.7 m high and have a total projected Although meteorological conditions are significantly associated crown area of 17.9 m and a diameter at breast height of 17.9 with dry and wet periods, which influence the hydrological cm. Their branches grow upwards, and its bark is smooth and cycle, the influence of these two water-related conditions has thin with a bark storage capacity estimated to be 0.7 mm (Za- been so far overlooked in the analysis of rainfall interception. bret and Šraj, 2021). Birch is a deciduous tree species with Rainfall interception is an important part of the hydrological distinct phenoseasons, which were determined according to the cycle and is, due to the inclusion of trees, also one of the eco- observations of the tree canopy at the field and complemented system services. The response of rainfall interception according with leaf area index (LAI) measurements, using LAI-2200c to various influencing variables, type of rainfall events, and Plant Canopy Analyzer (LI-COR). In general, the leafless phe- phenoseasons has been analysed; however, the process of rain- noseason was observed between October and April, when LAI fall interception associated with dry and wet periods has been was on average 0.8 and the canopy storage capacity was 1.1 neglected so far. As numerous researchers have observed the mm. The leafed phenoseason was observed between April and relationship between wet and dry periods and vegetation re- October, when LAI was equal to 2.6 and the canopy storage sponse to various natural processes, the main objective of the capacity increased to 3.5 mm. The group of the trees on the presented analysis is to investigate a possible influence of gen- northern part of the plot are pine trees (Pinus nigra Arnold). eral weather conditions (e.g., wet and dry periods) on through- They are on average 12.6 m high, have an average diameter at fall, stemflow, and rainfall interception. Extreme weather breast height of 19 cm, and a total projected crown area of 22.7 events are becoming more frequent due to climate change and m . The bark surface is rough, the bark itself is thick and more the differences in water balance between dry and wet periods absorbent with an estimated storage capacity of 3.5 mm. The are increasing. As a result, the connections between climate branches are inclined downwards. As pine is a coniferous tree variables and individual interception processes as well as the species, phenoseasons are not influencing the canopy character- processes of the hydrological cycle are also different. There are istics to such an extent as in the case of birch trees. However, not many studies with data sets long enough to capture wet and LAI in winter is 3.4 and the canopy storage capacity was esti- dry periods, therefore this is one of the important advantages of mated to be 2.7 mm, while in the summer time, LAI is 4.3 and this study. Two statistical methods, namely boosted regression the canopy storage capacity 2.9 mm. trees and random forest, were used to evaluate the influence of meteorological variables on rainfall partitioning components Measurements during wet, dry, and average years. Such statistical methods are seldom used for analysis of rainfall interception data, although The components of rainfall partitioning have been measured the application of such methods can give us a new, different at the study plot since the beginning of 2014 (Zabret and Šraj, insight into the data and the connections between them. 2021; Zabret et al., 2018). Measurements of throughfall and Additionally, the study of different tree species is very im- stemflow were performed under both groups of trees, while portant in the field of interception, as these results cannot be rainfall in the open was measured on the clearing at the study generalized. plot and at the nearby rooftop (Zabret, 2013; Zabret and Šraj 2019a; Zabret and Šraj, 2021). Values of other meteorological MATERIAL AND METHODS characteristics (wind speed and direction, air temperature and Study site humidity) were obtained from the Ljubljana Bežigrad meteoro- logical station (ARSO, 2020), which is because of its location The study site is located in the outskirt of the city of representative for the whole Ljubljana basin (Nadbath, 2008). Ljubljana, Slovenia (46.04° N, 14.49° E). The area has typical Measurements of throughfall were performed both automati- sub-alpine climate with well-defined seasons and is character- cally and manually. Under each group of trees there were two ized by Temperate oceanic climate (Cfb) according to the Kö- fixed steel trough gauges (0.75 m ) positioned from the tree ppen Climate Classification. The long-term analysis of the trunk towards the edge of the canopy. One was equipped with a meteorological data was prepared taking into account the data tipping bucket flow gauge (Unidata 6506G, 50 mL/tip) and a collected at the Ljubljana Bežigrad meteorological station be- data logger (Onset HOBO Event), while the other one was tween years 1986 and 2016 (ARSO, 2020). The average air connected to 10 L and 50 L polyethylene containers, which temperature for the area was equal to 10.5 °C. Generally, the were manually emptied after each event. Under each group of lowest temperatures are observed during January (–0.1 °C on trees there were also 10 funnel-type gauges (78.5 cm , 1-L average), while the warmest is July (20.8 °C on average). The capacity), manually emptied after each event and occasionally average long-term air temperature in winter was 0.8 °C, in moved under the trees to capture the spatial variability of spring and autumn 10.7 °C, and in summer 19.9 °C. The aver- throughfall. These collectors were moved after every 20 events age amount of rainfall delivered per year in the analysed period in a random pattern under the canopy. Throughfall values used was 1355 mm. The driest year was observed to be 2011, char- in the analysis were determined as the weighted average ac- acterized by 998 mm of rainfall, while the wettest year was cording to all the collectors’ area used. 2014, delivering 1851 mm of rainfall in total. The most rainfall Stemflow was measured per one tree from each group. The is in general delivered during the autumn months (around 30% halved rubber collar was spirally wrapped around the tree trunk of total yearly rainfall), while winter is the driest period, also and attached with silicone and nails. In case of a pine tree the because snow precipitation is observed instead of rainfall in the water was collected in a manually read 1-L container at the colder part of the year. bottom of the tree, which was emptied at the same time as the The study plot is part of a small urban park, located between throughfall collectors. In case of a birch tree, the stemflow was educational and business buildings. The research plot itself automatically recorded, as the hose from the collar was con- spans over 600 m and is covered with regularly mowed grass. nected to a tipping bucket flow gauge (Onset RG2-M, 0.2 In its western part there are two separated groups of trees, while mm/tip) and a data logger (Onset HOBO Event). 457 Katarina Zabret, Mojca Šraj Rainfall was measured at two locations, at the clearing ap- similar statistical methods, namely general boosted regression proximately 10 m from the nearest tree canopy and at the near- trees (BRT) and random forest (RF). Both models are based on by rooftop, approximately 45 m from the treetops. Rainfall at the method of the regression trees, however the way of upgrad- the clearing was measured with a tipping bucket rain gauge ing them differs for each method. Two methods were selected (Onset RG2-M, 0.2 mm/tip), connected to the data logger (On- for the analysis as a combination of several methods allows the set HOBO Event). Rainfall on the rooftop was measured with a verification of the results of an individual method and enables a disdrometer (OTT Parsivel), enabling also measurements of broader interpretation of the results. The regression tree model rainfall microstructure, i.e. raindrop diameter, raindrop veloci- is designed by repeating the divisions of the influential varia- ty, and the number of raindrops. The measuring area of the bles and by adapting a simple prediction model for the target disdrometer is 54 cm and the measured data are allocated to variable within each division. The result of the division process one of the 32 drop diameter classes (ranging from 0.312 mm to is shown graphically with a decision or regression tree (Loh, 24.5 mm) and 32 velocity classes (ranging from 0.05 m/s to 2011; Zabret et al., 2018). As a target variable, throughfall 20.8 m/s). The drop diameters smaller than 0.312 mm were (TF), stemflow (SF), and rainfall interception (I) were set. Each assigned to the smallest drop diameter class, as they are outside model was run six times per observed year (namely, 2014, the device’s measurement range. The recorded time series data 2015, and 2016), once per each target variable, taking into from the rain gauge and the disdrometer were used to identify account all influential variables and also the variables without the rainfall events (separated with at least a 4-hour dry period) data on the rainfall microstructure due to the longer period and their characteristics (duration and intensity). The 4-hour without available data (year 2015 was excluded). The influen- dry period was selected to divide the events based on the obser- tial variables included in the analysis (Table 1) were the total vations of the rainfall and throughfall dynamics at the field, as rainfall amount per event (Pa), the average rainfall event inten- during the wetter time of the year throughfall lasted for quite sity (Pi), the total duration of the rainfall event (Pd), the average some time after the cessation of the rainfall. Shorter rainfall air temperature (T), and the vapour pressure deficit (VPD) interruptions were captured as part of the defined events. The during an event, the average wind speed (Ws) and the direction dry period was defined with an accuracy of 0.2 mm of rainfall (Wd) per event, the dry period duration before a rainfall event (equal to the volume of the rain gauge tipping bucket). (DryP), the time when an event occurred, namely during the The tree characteristics were determined in individual sur- day, the night, or both (DN), the phenoseason (Feno), the aver- veys. The photographs of the trees were taken at a required age raindrop diameter (DropD), the velocity (DropV) per event, distance to avoid deformation of proportions and were used to the median volume diameter of an event’s raindrops (MVD), determine the tree height, the area of the projected canopy, and and the number of raindrops delivered per event (DropNr). the branch inclination. The diameter at breast height was calcu- lated from the measured perimeter of the stem. The bark stor- Table 1. Influential variables included in the analysis. age capacity was determined from the bark samples, extracted using a steel hole puncher, according to the procedure described Variable Abbreviation Unit by Perez-Harguindeguy et al. (2013). Phenoseasons were de- Rainfall amount per event P mm termined based on the regular measurements of LAI, performed Average rainfall event intensity per P mm/h event with LAI-2200c Plant Canopy Analyzer (LI-COR) following Total duration of the rainfall event P h the protocol for isolated trees (Li-COR, 2015). The canopy d storage capacity was calculated from the observed rainfall and Average air temperature during the T °C throughfall data according to the Leyton graphical method event Average vapour pressure deficit VPD kPa (Leyton et al., 1967). during the event Average wind speed during the event Ws m/s Data analysis Average wind direction during the W ° event Measured data of rainfall precipitation (P), throughfall (TF), Dry period duration before the DryP h and stemflow (SF), collected in years 2014, 2015, and 2016, rainfall event were used in the analysis. Based on these data, the third com- Time when the event occurred, name- DN – ponent of rainfall partitioning, i.e. rainfall interception (I), was ly during the day, the night, or both calculated for each event: Phenoseason Feno – Average raindrops diameter of the DropD mm I = P – TF – SF (1) drops, observed during the event Average raindrops velocity of the DropV m/s In the selected period, 413 rainfall events were observed in drops, observed during the event total, but not all of them were included in the analysis. Snow Median volume diameter of an MVD mm and sleet events were excluded in the initial phase, while during event’s raindrops the further preparation of the data, the events without complete Number of raindrops delivered per DropNr – time series on rainfall, throughfall, and stemflow due to clog- event ging of the measurement equipment were also excluded. There- fore 365 rainfall events were taken into account in the analysis, capturing 86% of the total rainfall, delivered in the analysed The BRT method combines two algorithms, regression trees period. Additionally, the disdrometer was not operational due to and boosting (Elith et al., 2008), which improve the efficiency a software error for a longer time period during 2015. There- of an individual model and provide a better understanding of fore, rainfall microstructure data were not included in the anal- the results with additional factors. Boosting is based on the ysis for this year. assumption that the average of many raw predictions, which are For the selected rainfall events, the influence of the variables upgraded after every single repetition, will result in a better describing general weather conditions was evaluated using two final model. The sequential approach of the step-by-step meth- 458 Relation of influencing variables and weather conditions on rainfall partitioning by birch and pine trees od iteratively adjusts and improves the model based on a set of age annual precipitation, 2014 was recognised as a wet, 2015 as training data (Elith et al., 2008). Due to the larger number of a dry, and 2016 as an average year. During 2014 we registered model runs, it is also possible to estimate the impact of an indi- 167 events, delivering 1575 mm of rainfall. For this year, the vidual variable on the design of the model and thus on the total rainfall amount (1841 mm) was 36% larger than the aver- target variable. Friedman (2001) presented an equation that can age long-term yearly rainfall amount of 1355 mm measured at be used to estimate the relative influence (RI) of each variable the Ljubljana-Bežigrad meteorological station. On the contrary, included in the BRT model. The RI is based on how many in 2015, we recorded 85 events, delivering 931 mm of rainfall. times a variable has been selected in the model to divide the The total delivered rainfall (1106 mm) was 18% smaller than regression tree. The number of selections is weighted by the the long-term average rainfall amount per year (1355 mm). square of the model improvement rate as a result of each split Furthermore, the year 2016 was similar to an average one, as and expressed as an average with respect to all generated re- we observed 113 rainfall events delivering 1139 mm of rainfall. gression trees (Friedman and Meulman, 2003). The RI is ad- Through the entire year, 1317 mm of rainfall was measured, justed so that the sum of the RI values of all considered varia- which is comparable to a long-term average precipitation of bles equals 100, making the higher values directly indicating a 1355 mm at the Ljubljana-Bežigrad meteorological station. greater influence of the variable. Although during the dry year 2014 the largest number of the The BRT models were implemented using the “gbm” pack- rainfall events were recorded, they on average delivered the age (Ridgeway, 2020) in R software (R core team, 2020). In the smallest amount of rainfall per event (9.4 mm) and on average initial phase we determined the arguments of the model, using lasted for the shortest time (5.7 h), but were on average the 75% of the whole data set for training and 25% of the data for most intense (2.1 mm/h) (Figure 1). The average rainfall inten- testing of the model, implementing 50 iterations for each set of sity and duration of rainfall events during the years 2015 and the arguments and calculating the RMSE value of predictions 2016 were similar (average intensity of 1.4 mm/h and 1.5 from all iterations. When adjusting the model, various number mm/h, respectively and average duration of 8.0 h and 8.1 h, of regression trees (15000, 1500, and 500) and values of the respectively); however, the events in the dry year 2015 deliv- shrinkage parameter (0.001, 0.01, 0.05) were applied. According ered on average more rainfall (11.0 mm) than the events in the to the results, the final BRT models were estimated, taking into average year 2016 (10.1 mm per event on average). account the Gaussian distribution, 1500 trees, a shrinkage pa- Comparing the climate conditions in the considered years rameter of 0.01, and 5 cross-validation folds. only slight differences were observed for the wind characteris- Random Forest (RF) is an ensemble-learning algorithm, tics, vapour pressure deficit, and air temperature. However, a which merges the concepts of regression trees and bagging noticeably shorter dry period between the events was observed (Breiman, 2001). Bagging is a procedure enabling growing of in the wet year 2014 (40 h on average) comparing to the years regression trees from different subsets in order to avoid highly 2015 and 2016 (58 h and 56 h, respectively). The rainfall events correlated predictors. This algorithm relies on random selection characteristics in the considered years also differ according to of trees to describe the reliable relationship between the target the rainfall microstructure. The size of the rainfall drops was and the influential variables. Cases are randomly selected from significantly different (p < 0.001) during the wet year 2014 a data set, a random sample is used to design an individual comparing to the years 2015 and 2016, as in the year 2014 an regression tree, and predictions are formed for the remaining average raindrop diameter was equal to 0.85 mm and MVD was cases. The model repeats this process several times. Random- equal to 1.79 mm, while during the years 2015 and 2016 the ness is additionally ensured by imposing different randomly drop diameter on average accounted to 0.67 mm and 0.62 mm selected sets of influential variables on each division. This is and MVD to 1.51 and 1.44 mm, respectively. However, the possible due to random and repeated selection of individual larger raindrops resulted in the smaller number of drops per target values and influential variables (Breiman, 2001). For event, as the lowest number of raindrops was on average de- each variable the variable importance measure is also estimated tected in the wet year 2014 (Figure 1). (Breiman et al., 2018). The variable importance gives the total The values of rainfall partitioning components were quite decrease in node impurities from splitting on the variable, aver- similar for the years 2014 and 2016, while some deviations are aged over all trees. In case of regression, as presented here, it is observed for the values measured in 2015, when higher values measured by the residual sum of squares. of throughfall and stemflow proportions according to the rain- The RF models were built in R software (R core team, fall in the open were observed (Figure 2). In general, over all 2020), using package “RandomForest” (Breiman et al., 2018). three observed years, throughfall under the birch tree was on In the first phase of the model establishment, we divided the average equal to 53% (± 34%), average stemflow was 1.2% (± data set into a training (75%) and test (25%) set. The model 2.5%), and average rainfall interception was 46% (± 35%). arguments were selected one by one, applying numerous itera- Throughfall under the pine tree was on average lower than tions for each of the 30 models. For the number of variables under the birch tree, resulting in 27% (± 26%) of rainfall in the randomly sampled as candidates at each split (mtry), the values open, while stemflow accounted for only 0.03% (± 0.10%) and between 10 and 40 were tested, using the “tune” function. The the rainfall interception by the pine tree on average presented maximum number of terminal nodes of the trees (maxnode) 73% (± 26%) of rainfall in the open. was applied for the values between 5 and 30, while the number of the trees to grow (n.trees) was tested for values between 250 Influence of the rainfall event characteristics on throughfall and 5000. For these two arguments the best value was selected according to the RMSE and R values of the iteration results. Both of the applied models, namely BRT and RF, indicate that throughfall under the birch trees is influenced by the larger RESULTS number of variables than throughfall under the pine trees, re- gardless the year (Figure 3). Throughfall (TF) under the birch The analysed data on rainfall partitioning were collected trees in the wet year 2014 was the most dependent on the rain- during the years 2014, 2015, and 2016. These years were hy- fall amount (P ) and intensity (P ), rainfall duration (P ), and a i d drologically quite distinct, as according to the long-term aver- the average vapour pressure deficit (VPD) during the rainfall 459 Katarina Zabret, Mojca Šraj Fig. 1. Boxplots of considered rainfall event characteristics for each analysed year. Fig. 2. Measured throughfall (TF) and stemflow (SF) by birch and pine trees per rainfall event according to the observed year. event (Figure 3). Rainfall amount and intensity demonstrated wind speed according to the RF model. In case of data for 2016, between 18% and 20% of relative influence (RI) each by both the inclusion of rainfall microstructure variables does not affect applied methods, while RI for the first four most influential the order of the influencing factors (Figure 3). As the most in- variables exceeded 60% in total. However, when taking into fluencing variable, the rainfall amount is still recognised by both account also the variables describing the rainfall microstructure, applied models, however the second most influencing variable, the number of raindrops (DropNr) became the most influencing having a similar value of RI, is the number of raindrops. In this variable, indicating the amount of throughfall by birch in the case both variables together represent 45% and 60% of RI ac- wet year 2014. cording to the BRT and RF model, respectively. For the throughfall under the birch trees during the dry year The number of influencing variables according to the domi- 2015, both models assigned a similar relative influence of al- nant value of the relative influence in the case of throughfall most 30% to rainfall intensity, indicating this variable as the under the pine tree is more straightforward (Figure 3). Rainfall most significant in addition to the rainfall amount. The BRT amount was recognized to be the most influencing variable model also recognized air temperature and vapour pressure regardless the year, with an average RI between 43% (RF for deficit as the influential variables with RI of 10%, while the 2014) and 82% (RF for 2016). Both models also recognized the random forest model assigned more than 8% of RI to rainfall influence of rainfall intensity and duration on throughfall by duration and wind speed (W ) (Figure 3). pine trees in 2014, while in 2015, more than 8% of RI was The data collected during the average year 2016 showed a assigned to wind speed. In 2016, in addition to the rainfall significant influence of the rainfall amount only, as it represent- duration air temperature was the second most influencing vari- ed almost 40% of RI according to the BRT model and more able with RI larger than 5%. None of the rainfall microstructure than half of the total RI expressed by RF model. More than 9% variables exceeded more than 6% of RI, regardless the applied of RI was assigned also to wind speed and vapour pressure model or the year observed in case of throughfall under the pine deficit according to the BRT method and to air temperature and trees. 460 Relation of influencing variables and weather conditions on rainfall partitioning by birch and pine trees Fig. 3. Relative influence (RI) of the considered variables for throughfall (TF) by the birch and pine trees according to the observed years, evaluated by the boosted regression trees (BRT) and random forest (RF) models. Influence of the rainfall event characteristics on stemflow encing variables, combined with the number of raindrops. Stemflow in the wet year 2014 is still the most influenced by Similarly as throughfall, stemflow is in general the most in- the rainfall amount, while the number of raindrops and MVD fluenced by the rainfall amount (Figure 4). Stemflow (SF) by were also recognized as more influential. However, for the birch tree was the most characterized by the rainfall amount stemflow in 2016, the number of raindrops together with the regardless the year as the RI for this variable ranged between rainfall amount and duration were recognized as the variables 35% (RF, year 2014) and 61% (RF, year 2015). Stemflow by with the highest RI (together accounting for 59% according to the birch tree in the wet year 2014 and the average year 2016 the BRT and 82% according to the RF model). was also highly influenced by the rainfall duration, which had The amount of stemflow by the birch trees was similarly in- the second highest RI in both years, regardless the model used. fluenced during the years 2014 and 2016, however for the pine On the contrary, in the dry year 2015 stemflow by the birch trees similarities can be observed between the years 2015 and tree was affected by a larger number of variables (Figure 4). 2016 (Figure 4). Stemflow by the pine trees during 2014 was The BRT model indicated that in addition to the rainfall the most influenced by wind direction, followed by the rainfall amount, stemflow by the birch tree is also influenced by rainfall amount. In case of the BRT model these two variables resulted intensity, wind speed, vapour pressure deficit, and rainfall in RI of 77%, while in case of the RF model, the influencing duration, as RI for all mentioned variables was larger than 9% variables with RI of more than 10% are also vapour pressure (Figure 4). However, according to the RF model, the value of deficit, wind speed, and air temperature. RI higher than 10% was estimated for the dry period duration When also including the rainfall microstructure variables, and air temperature. the influence of wind direction is minimized, as rainfall When taking into account also the rainfall microstructure amount, duration, and the number of raindrops in combination characteristics, the rainfall amount is still one of the most influ- with MVD (estimated by the BRT model) and air temperature 461 Katarina Zabret, Mojca Šraj Fig. 4. Relative influence (RI) of the considered variables for stemflow (SF) by the birch and pine trees according to the observed years, evaluated by the boosted regression trees (BRT) and random forest (RF) models. (estimated by the RF model) resulted in RI higher than 50%. Rainfall interception by birch and pine trees is the most in- Stemflow by the pine trees in 2015 and 2016 is significantly fluenced by the rainfall amount, which has the highest values of influenced by the rainfall amount and intensity, as regardless RI according to both models. In case of the birch trees the val- the model or the year, these two variables present between 64% ues of RI for the amount of rainfall ranged between 22% and and 85% of RI. The substantial influence of the rainfall amount 63%, while in case of the pine trees they were even higher, and intensity is also retained when introducing the rainfall ranging from 47% to 83%. Comparing these values to RI esti- microstructure influence. In this case, as the second most influ- mated for throughfall, the values were a bit larger in case of the encing variable with RI larger than 10% both models recog- birch trees, while for the pine trees they were kept in a similar nised MVD. range. Rainfall interception of the birch trees was in the wet year Influence of rainfall event characteristics on rainfall 2014 also significantly influenced by the rainfall duration and interception intensity, while in the dry year 2015 it was mainly influenced by rainfall intensity and in the average year 2016 by vapour Rainfall interception (I) is calculated as the difference be- pressure deficit (according to the BRT model) and air tempera- tween the measured values, i.e. rainfall amount in the open, ture (according to the RF model). In case of the pine trees the throughfall, and stemflow (Eq. 1). Therefore, as the amount of results were also very similar to the ones for the throughfall; in throughfall is much larger than stemflow, this is the value that 2015 and 2016 only the rainfall amount played a significant mainly determines the proportion of intercepted rainfall, result- role in the process of rainfall interception, while in the wet year ing in similarly evaluated influencing variables as throughfall 2014 also rainfall intensity and duration demonstrated RI values (Figure 3). larger than 10% (Figure 5). 462 Relation of influencing variables and weather conditions on rainfall partitioning by birch and pine trees Fig. 5. Relative influence (RI) of the considered variables for rainfall interception (I) by the birch and pine trees according to the observed years, evaluated by the boosted regression trees (BRT) and random forest (RF) models. The results of both applied models considering also the rain- account results of the BRT instead of the RF model. Thus, the fall microstructure are also similar to the results of throughfall combined analysis of the two methods allows for a more com- data analysis (Figure 3). In case of the birch trees, the number prehensive evaluation of the results, as the RF model indicates of drops was recognised as a variable with the highest influence the most influencing variables, while the BRT model highlights among the newly introduced variables, while in case of the pine also the other possible variables with meaningful influence. trees for none of these variables the estimated RI exceeded 6% The results demonstrate that throughfall, stemflow, and rain- (Figure 5). fall interception by birch and pine trees were the most influ- enced by the amount of rainfall, which has been repeatedly DISCUSSION recognized as the factor most influencing the rainfall partition- ing components in general also in other studies (e.g., Levia and Although the two methods are very similar as they are both Germer, 2015; Staelens et al., 2008; Su et al., 2019; Zabret et based on the principle of regression trees, there is one main al., 2018). In case of both considered tree species, rainfall dura- difference if we consider the method associated with the regres- tion seems to play an important role mainly during the wet year sion trees (boosting and bagging). This is also reflected in the 2014, while rainfall intensity had a significant influence on estimation of the most influential variables and their RI values. rainfall partitioning by birch trees during the dry year 2015. A comparison of the results by the two models shows that in This observation seems to correlate well with the results pre- general, the RI values of the variables estimated by the RF sented by Mużyło et al. (2012), who observed a significant model are higher than those estimated by the BRT model (Fig- influence of rainfall duration on throughfall in a deciduous ures 3–5). Therefore, the number of the variables for which the forest, especially during the leafless season. The leafless season RI value exceeds the threshold value is larger when taking into is usually characterized by more precipitation and generally 463 Katarina Zabret, Mojca Šraj wetter months, which may be equivalent to the hydrologically The relative influence, estimated by the BRT and RF mod- wetter year of 2014, in which a more pronounced influence of els, shows that throughfall under the birch trees is determined rainfall duration was observed in this study (Figures 3–5). by a larger number of influencing variables. In addition to the As the wetter year 2014 can be correlated with the wetter rainfall amount, duration, and intensity, also air temperature leafless phenoseason, the drier year 2015 is expected to be and vapour pressure deficit (VPD) were assigned with values of associated with the drier leafed period. Therefore, the influence RI larger than 8%. Air temperature and VPD are closely con- of rainfall intensity on rainfall partitioning in the drier year nected to the season of the year, corresponding also to the phe- 2015 is initially unexpected. Rainfall intensity was actually noseasons, and are especially significant for a deciduous birch recognized as one of the most influential variables in previous trees (Zabret et al., 2018). Therefore, the significant RI values studies, but its effect was observed for winter throughfall (Xiao of air temperature and VPD may indirectly indicate the influ- et al., 2000), rainfall interception in the leafless period (Zabret ence of phenoseasons on throughfall by birch, which is larger in et al., 2018), and rainfall interception in a wet year (Zabret and the leafless period, characterized by lower air temperature and Šraj, 2019b). However, for a beech tree, Staelens et al. (2008) lower VPD values (Andre et al., 2008; Brasil et al., 2020; reported significant influence of rainfall intensity on stemflow, Mużyło et al., 2012; Šraj et al., 2008; Zabret and Šraj, 2018; especially during the leafed period resulting in a decrease in the Zabret et al., 2021). However, the relation between the influ- stemflow amount due to splashing of droplets intercepted by ence of phenoseasons and meteorological variables on rainfall the canopy and forming throughfall instead of stemflow. Addi- partitioning has already been recognized as a very complex one tionally, a more evident influence of rainfall intensity was (e.g., Andre et al., 2008; Mużyło et al., 2012; Zabret and Šraj, estimated by both applied models for birch rather than for pine 2021). When analysing the influence of air temperature and trees (Figures 3–5). A different influence of rainfall intensity on VPD on throughfall by the birch tree, the results are similar tree species with distinct vegetation properties was already among the years (Figure 6). Throughfall is in general decreas- observed in other analyses (e.g., Sadeghi et al., 2020; Siegert ing with increasing air temperature, which was observed also and Levia, 2014; Zabret et al., 2018). Birch trees have a by Staelens et al. (2008). Warmer months of the year are also smoother bark surface and more flexible leaves compared to the characterized with a fully leafed canopy, also decreasing the rougher and more absorbent bark of pine trees and its compact throughfall, while a higher air temperature increases the evapo- needles, therefore the process of splashing of intercepted drop- ration, which may also lead to a decrease in throughfall (Šraj et lets may be more intense in the canopy of the birch trees. al., 2008; Xiao et al., 2000). However, the response of Fig. 6. Partial dependence plots of the influence of air temperature (T) and vapour pressure deficit (VPD) on throughfall (TF) by birch trees during the considered years. 464 Relation of influencing variables and weather conditions on rainfall partitioning by birch and pine trees throughfall according the VPD values is similar for the years The variables with the highest influence expressed by both 2014 and 2015, as lower VPD up to 1.5 kPa increases through- models were the rainfall amount and the number of raindrops. fall under the birch tree, while larger values of VPD decrease The comparison of the influential variables indicates to some the amount of throughfall (Figure 6). The data collected in 2016 extent the correlation between the wet period and the leafless show a bit different response of throughfall according to the season, as well as between the dry period and the leafed season. VPD values up to 2 kPa, while larger VPD values decrease For example, rainfall duration had a high relative influence on throughfall under birch trees as well. rainfall partitioning by both tree species mainly in the wet year Pine tree’s stemflow was influenced by more variables com- 2014, while researchers already reported its influence in the pared to throughfall. In addition to the rainfall amount, which leafless season. Stemflow by birch trees was also strongly mainly determines stemflow under both tree species, rainfall influenced by air temperature and vapour pressure deficit, intensity and rainfall duration also had an important influence which are dependent on the season of the year, which is also on stemflow under the pine and birch trees, respectively. The consistent with the phenoseason. However, the results of the significance of rainfall duration on stemflow by the birch tree models also indicate significant differences in the response of was also recognised in the analysis of the stemflow response the two tree species. The influence of rainfall intensity, the (Zabret and Šraj, 2021), as well as in the analysis of predicting number of raindrops, and the median volume diameter was more the stemflow of a birch tree using the regression trees (Zabret et pronounced in the case of the birch trees, while it was negligible al., 2018). Although rainfall intensity was recognised as a less in the case of the pine trees. This observation coincides with the influential variable in the case of throughfall under the pine conclusions of previous studies, i.e. that raindrops behave trees, it seems to have a larger influence on its stemflow. Re- differently when interacting with needles or leaves. sults of both implemented models (BRT and RF) indicate RI The presented analysis mainly confirms all previous obser- values larger than 15% for rainfall intensity in the years 2015 vations made by other researchers about the different influences and 2016 (Figure 4). This is consistent with the results of the on the rainfall partitioning process by distinct tree species. BRT model applied to stemflow data of a leafed phenoseason However, a new insight into the impact of wet and dry period is (Zabret et al., 2018), which also indicates some similarities in presented, indicating that during a longer wet period the trees meteorological influences during the leafed phenoseason and behave similarly as in the leafless period and during the longer dry period the rainfall interception process is similar as that the drier hydrological year. in Introduction of variables specifying the rainfall microstruc- the leafed period. Nevertheless, additional analysis, taking into ture into the analysis expressed significant influence of the account multiple wet and dry periods as well as data for these number of rain drops on throughfall and rainfall interception by periods for other tree species and other locations with different the birch trees. The number of raindrops as well as the mean microclimatic characteristics, should be implemented in order volume diameter (MVD) were estimated to have considerable to understand this aspect in more detail. influence on stemflow in case of both considered tree species. However, no influence of these variables on throughfall under Acknowledgements. The work was founded by the Slovenian the pine trees was observed, as throughfall under the pine trees Research Agency (ARRS) through research program P2-0180. was still the most influenced by the amount of rainfall, which provided more than half of the RI according to the other con- REFERENCES sidered variables. A more noticeable influence of rainfall mi- crostructure on throughfall by birch than by pine trees was also Andre, F., Jonard, M., Ponette, Q., 2008. Influence of species confirmed in previous study (Zabret et al., 2018), in which a and rain event characteristics on stemflow volume in a tem- similar comparison of data with BRT models per phenoseason perate mixed oak–beech stand. Hydrol. Process., 22, 4455– showed the influence of MVD on throughfall by pine trees only 4466. in the leafless season, while the number of raindrops had a ARSO, 2020. Measurements archive. http://www.meteo.si/met/ significant role in regulating throughfall under the birch trees sl/archive/ (Accessed 5 May 2021). regardless the phenoseason. This might be connected to the Bezak, N., Mikoš, M., 2014. Estimation of design floods using distinct characteristics of the foliage of the considered tree univariate and multivariate flood frequency approach with species, i.e. leaves of the birch trees and needles of the pine regard to one wet year. Acta hydrotechnica, 27, 103–117. trees. The different interaction of needles and leaves to the rain Brasil, J.B., de Andrade, E.M., de Queiroz Palácio, H.A., dos drops and their characteristics has been already reported by Santos, J.C.N., Medeiros, P.H.A., 2020. Temporal variabil- other researchers (e.g., Holder, 2013; Nanko et al., 2016; Zabret ity of throughfall as a function of the canopy development et al., 2017; Zabret et al., 2018). stage: from seasonal to intra-event scale. Hydrol. Sci. J., 65, 1640–1651. CONCLUSIONS Breiman, L., 2001. Random Forests. Mach. Learn., 45, 5–32. Breiman, L., Cutler, A., Liaw, A., Wiener, M., 2018. Package The rainfall partitioning process is part of the hydrological ‘RandomForest’. https://cran.r-project.org/web/packages/ cycle, for which changes are expected due to more pronounced randomForest/randomForest.pdf (Accessed 25 March 2021) precipitation patterns, resulting in more intense wet and dry Dohnal, M., Černý, T., Votrubová, J., Tesař, M., 2014. Rainfall periods. Accordingly, the influence of meteorological factors interception and spatial variability of throughfall in spruce on throughfall, stemflow, and rainfall interception during a wet, stand. J. Hydrol. Hydromech., 62, 277–284. a dry, and an average year was analysed. Two similar statistical Elith, J., Leathwick, J.R., Hastie, T., 2008. A working guide to methods based on the regression tree approach were applied, boosted regression trees. J. Anim. Ecol., 77, 802–813. namely boosted regression trees and random forest. The com- Friedman, J.H., 2001. Greedy function approximation: a gradi- parison of the results showed that the methods are complemen- ent boosting machine. Ann. Stat., 29, 1189–1232. tary, since the BRT model indicates numerous variables with Friedman, J.H., Meulman, J.J., 2003. Multiple additive regres- relevant influence and the RF model highlights the variables sion trees with application in epidemiology. Stat. Med., 22, with the highest influence. 1365–1381. 465 Katarina Zabret, Mojca Šraj Gao, S., Zhou, T., Yi, C., Shi, P., Fang, W., Liu, R., Liang, E., Partitioning by Vegetation: A Global Synthesis. Springer Camarero, J.J., 2020. Asymmetric impacts of dryness and Nature, Berlin, Germany, pp. 49–70. wetness on tree growth and forest coverage. Agr. Forest. Schooling, J.T., Carlyle-Moses, D.E., 2015. The influence of Meteorol., 288–289, 107980. rainfall depth class and deciduous tree traits on stemflow Hao, Y., Wang, Y., Mei, X., Huang, X., Cui, X., Zhou, X., Niu, production in an urban park. Urban Ecosyst., 18, 1261–1284. H., 2008. CO2, H2O and energy exchange of an Inner Mon- Siegert, C.M., Levia, D.F., 2014. Seasonal and meteorological golia steppe ecosystem during a dry and wet year. Acta Oec- effects on differential stemflow funneling ratios for two de- ologica, 33, 133–143. ciduous tree species. J. Hydrol., 519, 446–454. Holder, C.D., 2013. Effects of leaf hydrophobicity and water Staelens, J., De Schrijver, A., Verheyen, K., Verhoest, N.E.C., droplet retention on canopy storage capacity. Ecohydrology, 2008. Rainfall partitioning into throughfall, stemflow, and 6, 483–490. interception within a single beech (Fagus sylvatica L.) cano- Honda, E.A., Mendonça, A.H., Durigan, G., 2014. Factors py: influence of foliation, rain event characteristics, and me- affecting the stemflow of trees in the Brazilian Cerrado. teorology. Hydrol. Process., 22, 33–45. Ecohydrology, 8, 1351–1362. Su, L., Xie, Z., Xu, W., Zhao, C., 2019. Variability of through- Hungate, B., Hampton, H., 2012. Valuing ecosystems for cli- fall quantity in a mixed evergreen-deciduous broadleaved mate. Nat. Clim. Change, 2, 151–152. forest in central China. J. Hydrol. Hydromech., 67, 225–231. Inglezakis, V.J., Poulopoulos, S.G., Arkhangelsky, E., Zorpas, Šraj, M., Brilly, M., Mikoš, M., 2008. Rainfall interception by A.A., Menegaki, A.N., 2016. Aquatic environment. In: Pou- two deciduous Mediterranean forests of contrasting stature lopoulos, S., Inglezakis, V. (Eds.): Environment and Devel- in Slovenia. Agr. Forest. Meteorol., 148, 121–134. opment: Basic Principles, Human Activities, and Environ- Xiao, Q., McPherson, E.G., Ustin, S.L., Grismer, M.E., Simp- mental Implications. Elsevier, pp. 137–212. son, J.R., 2000. Winter rainfall interception by two mature Klamerus-Iwan A., Link T.E., Keim R.F., Van Stan, J.T., 2020. open-grown trees in Davis, California. Hydrol. Process., 14, Storage and routing of precipitation through canopies. In: 763–784. Van Stan, J T., Gutmann, E., Friesen, J. (Eds.): Precipitation Xu, Z., Feng, Z., Zhao, C., Zheng, J., Yang, J., Tian, F., Peng, Partitioning by Vegetation: A Global Synthesis. Springer H., Wang, C., Peng, S., Sher, H., 2013. The canopy rainfall Nature, Berlin, Germany, pp. 17–34. interception in actual and potential distribution of Qinghai Levia, D.F., Germer, S., 2015. A review of stemflow generation spruce (Picea crassifolia) forest. J. Hydrol. Hydromech., 61, dynamics and stemflow-environment interactions in forests 64–72. and shrublands. Rev. Geophys., 53, 673–714. Xu, L., Cao, G., Wang, Y., Hao, J., Wang, Y., Yu, P., Liu, Z., Leyton, L., Reynolds, E.R.C., Thompson, F.B., 1967. Rainfall Xiong, W., Wang, X., 2020. Components of stand water bal- interception in forest and moorland. In: Sopper, W.E., Lull, ance of a larch plantation after thinning during the extremely H.W. (Eds.): Forest Hydrology. Pergamon, Oxford, pp. 163– wet and dry years in the Loess Plateau, China. Global Eco. 178. Conser., 24, e01307. Loh, W., 2011. Classification and regression trees. Data Min. Yue, K., De Frenne, P., Fornara, D.A., Van Meerbeek, K., Li, Knowl. Disc., 1, 14–23. W., Peng, X., Ni, X., Peng, Y., Wu, F., Yang, Y., Peñuelas, Mużyło, A., Llorens, P., Domingo, F., 2012. Rainfall partition- J., 2021. Global patterns and drivers of rainfall partitioning ing in a deciduous forest plot in leafed and leafless periods. by trees and shrubs. Glob. Change. Biol., 27, 3350–3357. Ecohydrology, 5, 759–767. Zabret, K., 2013. The influence of tree characteristics on rainfall Nadbath, M., 2008. Meteorological station Ljubljana Bežigrad. interception. Acta Hydrotech., 26, 99–116. (In Slovenian.) Naše okolje 15, 1. (In Slovenian.) Zabret, K., Rakovec, J., Mikoš, M., Šraj, M., 2017. Influence of Nanko, K., Hudson, S.A., Levia, D.F., 2016. Differences in raindrop size distribution on throughfall dynamics under throughfall drop size distributions in the presence and ab- pine and birch trees at the rainfall event level. Atmosphere, sence of foliage. Hydrolog. Sci. J., 61, 620–627. 8, 240. Peng, Y., Chen, L., Tian, J., Sun, B., Jiang, C., Lu, Y., Shang, Zabret, K., Rakovec, J., Šraj, M., 2018. Influence of meteoro- J., 2021. Ecosystem services help alleviate the intensity of logical variables on rainfall partitioning for deciduous and dryness/wetness. Global Ecol. Conser., 27, e01581. coniferous tree species in urban area. J. Hydrol., 558, 29–41. Perez-Harguindeguy, N., Diaz, S., Garnier, E. et al. 2013. New Zabret, K., Šraj, M., 2019a. Evaluating the influence of rain event handbook for standardized measurement of plant functional characteristics on rainfall interception by urban trees using traits worldwide. Aust. J. Bot., 61, 167–234 multiple correspondence analysis. Water, 11, 2659. R core team, 2020. R: A language and environment for statisti- Zabret, K., Šraj, M., 2019b. Rainfall interception by urban trees cal computing. R Foundation for Statistical Computing, Vi- and their impact on potential surface runoff. Clean Soil, Air, enna, Austria. http://www.R-project.org/ (Accessed 20 Au- Water, 47, 8, 1800327. gust 2020) Zabret, K., Šraj, M., 2021. How characteristics of a rainfall Ridgeway, G., 2020. Generalized Boosted Regression Models. event and the meteorological conditions determine the de- https://cran.r-project.org/web/packages/gbm/gbm.pdf (Ac- velopment of stemflow: A case study of a birch tree. Front. cessed 10 August 2020) Front. For. Glob. Change, 4, 663100. Sadeghi, S.M.M., Gordon, D.A., Van Stan, J.T., 2020. A global synthesis of throughfall and stemflow hydrometeorology. In: Received 28 May 2021 Van Stan, J T., Gutmann, E., Friesen, J. (Eds.): Precipitation Accepted 5 August 2021 http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Journal of Hydrology and Hydromechanics de Gruyter

Relation of influencing variables and weather conditions on rainfall partitioning by birch and pine trees

Loading next page...
 
/lp/de-gruyter/relation-of-influencing-variables-and-weather-conditions-on-rainfall-9mDLwfToA4
Publisher
de Gruyter
Copyright
© 2021 Katarina Zabret et al., published by Sciendo
ISSN
0042-790X
eISSN
1338-4333
DOI
10.2478/johh-2021-0023
Publisher site
See Article on Publisher Site

Abstract

J. Hydrol. Hydromech., 69, 2021, 4, 456–466 ©2021. This is an open access article distributed DOI: 10.2478/johh-2021-0023 under the Creative Commons Attribution ISSN 1338-4333 NonCommercial-NoDerivatives 4.0 License Relation of influencing variables and weather conditions on rainfall partitioning by birch and pine trees 1, 2 1* Katarina Zabret , Mojca Šraj University of Ljubljana, Faculty of Civil and Geodetic Engineering, Jamova 2, 1000 Ljubljana, Slovenia. Institute for Water of the Republic of Slovenia, Einspielerjeva 6, 1000 Ljubljana, Slovenia. Corresponding author. Tel.: +386 1 4768 684. E-mail: mojca.sraj@fgg.uni-lj.si Abstract: General weather conditions may have a strong influence on the individual elements of the hydrological cycle, an important part of which is rainfall interception. The influence of general weather conditions on this process was analysed, evaluating separately the influence of various variables on throughfall, stemflow, and rainfall interception for a wet (2014), a dry (2015), and an average (2016) year. The analysed data were measured for the case of birch and pine trees at a study site in the city of Ljubljana, Slovenia. The relationship between the components of rainfall partitioning and the influential variables for the selected years was estimated using two statistical models, namely boosted regression trees and random forest. The results of both implemented models complemented each other well, as both indicated the rainfall amount and the number of raindrops as the most influential variables. During the wet year 2014 rainfall duration seems to play an important role, correlating with the previously observed influence of the variables during the wetter leafless period. Similarly, during the dry year 2015, rainfall intensity had a significant influence on rainfall partitioning by the birch tree, again corresponding to the influences observed during the drier leafed period. Keywords: Throughfall; Stemflow; Rainfall interception; Rainfall microstructure; Boosted regression trees; Random forest. INTRODUCTION ecosystems, for which Hao et al. (2008) reported that both timing and frequency of rainfall events during the growing The hydrological cycle is altering due to climate change, as season significantly alter the capacity of steppe vegetation to differences in global redistribution of precipitation and varia- uptake CO . tions in seasonal precipitation patterns are observed (Inglezakis Forest ecosystems and trees in general also significantly in- et al., 2016). This results in a significant reduction of precipita- fluence the hydrological cycle through the process of rainfall tion in some parts of the world, while major variations in the interception (Dohnal et al., 2014; Klamerus-Iwan et al., 2020; timing and amount of precipitation per dry and wet season are Xu et al., 2013). Precipitation reaching the vegetation surface is expected elsewhere (Peng et al., 2021). The pronounced differ- distributed among the intercepted rainfall, which is captured by ences between the wet and dry periods significantly alter the the canopy and eventually evaporates back into to the atmos- water yield and the local water balance, the ecosystem services, phere, throughfall, which is described as the precipitation the water availability for vegetation, leading to changed occur- reaching the ground by dripping from the canopy or falling rences of floods and droughts (Bezak and Mikoš, 2014; Hun- directly to the ground through the gaps in the foliage, and gate and Hampton, 2012; Xu et al., 2020). stemflow, presenting the water flowing to the ground down the In the context of climate change, the relationship between branches and stems (Levia and Germer, 2015; Sadeghi et al., the water balance and vegetation in dry and wet periods is 2020; Staelens et al., 2008; Xiao et al., 2000; Yue et al., 2021; increasingly recognized. In this aspect, various influences of Zabret et al., 2018). Rainfall interception is influenced by vege- different vegetation systems were studied. Vegetation is an tation and meteorological characteristics. Vegetation character- important component, determining the ecosystem services, istics considered are mainly tree characteristics, such as the tree which were recognised to help mitigate the intensity of ex- height and surface area (e.g., projected tree canopy), smooth- tremely dry and wet conditions expected in the future (Peng et ness and absorbance of the bark, leaf area index, canopy cover- al., 2021). An important contribution to the ecosystem services age, and canopy storage capacity (Dohnal et al., 2014; Klamer- is also presented by the forest ecosystem affecting the global us-Iwan et al., 2020; Xu et al., 2013; Zabret, 2013). According carbon budget. The different response of a forest ecosystem in to the differences among the tree species, the different response wet and dry periods was analysed by Xiao et al. (2020), who of rainfall partitioning was analysed (Honda et al., 2014; concluded that in the dry season the precipitation generated Schooling and Carlyle-Moses, 2015). As characteristics of significantly positive effects to the cumulative CO emissions, some tree species (e.g., deciduous trees) are substantially influ- while the soil respiration rate was mainly influenced by the fine enced by the phenoseasons (presence and absence of leaves in root biomass regardless the season. An analysis of historical the tree canopy), the rainfall partitioning in leafed and leafless data from the tree rings was performed by Gao et al. (2020), period has also been frequently studied, mainly in relation to who observed that the growth of trees was improved by wet- the meteorological conditions (Brasil et al., 2020; Levia and ness, suggesting that tree growth is more sensitive to wetness Germer, 2015; Mużyło et al., 2012; Su et al., 2019; Zabret et than the forest coverage. Wetter conditions may, on the contra- al., 2018). Meteorological characteristics on the contrary ex- ry, reduce the carbon flux and evapotranspiration in steppe plain the characteristics of rainfall events, for example the 456 Relation of influencing variables and weather conditions on rainfall partitioning by birch and pine trees rainfall amount, duration and intensity, air temperature and in the east side there is a clearing. One group of trees in the humidity, vapour pressure deficit and wind conditions (Andre southern part consists of birch trees (Betula pendula Roth.), et al., 2008; Staelens et al., 2008; Zabret and Šraj, 2019a). which are on average 15.7 m high and have a total projected Although meteorological conditions are significantly associated crown area of 17.9 m and a diameter at breast height of 17.9 with dry and wet periods, which influence the hydrological cm. Their branches grow upwards, and its bark is smooth and cycle, the influence of these two water-related conditions has thin with a bark storage capacity estimated to be 0.7 mm (Za- been so far overlooked in the analysis of rainfall interception. bret and Šraj, 2021). Birch is a deciduous tree species with Rainfall interception is an important part of the hydrological distinct phenoseasons, which were determined according to the cycle and is, due to the inclusion of trees, also one of the eco- observations of the tree canopy at the field and complemented system services. The response of rainfall interception according with leaf area index (LAI) measurements, using LAI-2200c to various influencing variables, type of rainfall events, and Plant Canopy Analyzer (LI-COR). In general, the leafless phe- phenoseasons has been analysed; however, the process of rain- noseason was observed between October and April, when LAI fall interception associated with dry and wet periods has been was on average 0.8 and the canopy storage capacity was 1.1 neglected so far. As numerous researchers have observed the mm. The leafed phenoseason was observed between April and relationship between wet and dry periods and vegetation re- October, when LAI was equal to 2.6 and the canopy storage sponse to various natural processes, the main objective of the capacity increased to 3.5 mm. The group of the trees on the presented analysis is to investigate a possible influence of gen- northern part of the plot are pine trees (Pinus nigra Arnold). eral weather conditions (e.g., wet and dry periods) on through- They are on average 12.6 m high, have an average diameter at fall, stemflow, and rainfall interception. Extreme weather breast height of 19 cm, and a total projected crown area of 22.7 events are becoming more frequent due to climate change and m . The bark surface is rough, the bark itself is thick and more the differences in water balance between dry and wet periods absorbent with an estimated storage capacity of 3.5 mm. The are increasing. As a result, the connections between climate branches are inclined downwards. As pine is a coniferous tree variables and individual interception processes as well as the species, phenoseasons are not influencing the canopy character- processes of the hydrological cycle are also different. There are istics to such an extent as in the case of birch trees. However, not many studies with data sets long enough to capture wet and LAI in winter is 3.4 and the canopy storage capacity was esti- dry periods, therefore this is one of the important advantages of mated to be 2.7 mm, while in the summer time, LAI is 4.3 and this study. Two statistical methods, namely boosted regression the canopy storage capacity 2.9 mm. trees and random forest, were used to evaluate the influence of meteorological variables on rainfall partitioning components Measurements during wet, dry, and average years. Such statistical methods are seldom used for analysis of rainfall interception data, although The components of rainfall partitioning have been measured the application of such methods can give us a new, different at the study plot since the beginning of 2014 (Zabret and Šraj, insight into the data and the connections between them. 2021; Zabret et al., 2018). Measurements of throughfall and Additionally, the study of different tree species is very im- stemflow were performed under both groups of trees, while portant in the field of interception, as these results cannot be rainfall in the open was measured on the clearing at the study generalized. plot and at the nearby rooftop (Zabret, 2013; Zabret and Šraj 2019a; Zabret and Šraj, 2021). Values of other meteorological MATERIAL AND METHODS characteristics (wind speed and direction, air temperature and Study site humidity) were obtained from the Ljubljana Bežigrad meteoro- logical station (ARSO, 2020), which is because of its location The study site is located in the outskirt of the city of representative for the whole Ljubljana basin (Nadbath, 2008). Ljubljana, Slovenia (46.04° N, 14.49° E). The area has typical Measurements of throughfall were performed both automati- sub-alpine climate with well-defined seasons and is character- cally and manually. Under each group of trees there were two ized by Temperate oceanic climate (Cfb) according to the Kö- fixed steel trough gauges (0.75 m ) positioned from the tree ppen Climate Classification. The long-term analysis of the trunk towards the edge of the canopy. One was equipped with a meteorological data was prepared taking into account the data tipping bucket flow gauge (Unidata 6506G, 50 mL/tip) and a collected at the Ljubljana Bežigrad meteorological station be- data logger (Onset HOBO Event), while the other one was tween years 1986 and 2016 (ARSO, 2020). The average air connected to 10 L and 50 L polyethylene containers, which temperature for the area was equal to 10.5 °C. Generally, the were manually emptied after each event. Under each group of lowest temperatures are observed during January (–0.1 °C on trees there were also 10 funnel-type gauges (78.5 cm , 1-L average), while the warmest is July (20.8 °C on average). The capacity), manually emptied after each event and occasionally average long-term air temperature in winter was 0.8 °C, in moved under the trees to capture the spatial variability of spring and autumn 10.7 °C, and in summer 19.9 °C. The aver- throughfall. These collectors were moved after every 20 events age amount of rainfall delivered per year in the analysed period in a random pattern under the canopy. Throughfall values used was 1355 mm. The driest year was observed to be 2011, char- in the analysis were determined as the weighted average ac- acterized by 998 mm of rainfall, while the wettest year was cording to all the collectors’ area used. 2014, delivering 1851 mm of rainfall in total. The most rainfall Stemflow was measured per one tree from each group. The is in general delivered during the autumn months (around 30% halved rubber collar was spirally wrapped around the tree trunk of total yearly rainfall), while winter is the driest period, also and attached with silicone and nails. In case of a pine tree the because snow precipitation is observed instead of rainfall in the water was collected in a manually read 1-L container at the colder part of the year. bottom of the tree, which was emptied at the same time as the The study plot is part of a small urban park, located between throughfall collectors. In case of a birch tree, the stemflow was educational and business buildings. The research plot itself automatically recorded, as the hose from the collar was con- spans over 600 m and is covered with regularly mowed grass. nected to a tipping bucket flow gauge (Onset RG2-M, 0.2 In its western part there are two separated groups of trees, while mm/tip) and a data logger (Onset HOBO Event). 457 Katarina Zabret, Mojca Šraj Rainfall was measured at two locations, at the clearing ap- similar statistical methods, namely general boosted regression proximately 10 m from the nearest tree canopy and at the near- trees (BRT) and random forest (RF). Both models are based on by rooftop, approximately 45 m from the treetops. Rainfall at the method of the regression trees, however the way of upgrad- the clearing was measured with a tipping bucket rain gauge ing them differs for each method. Two methods were selected (Onset RG2-M, 0.2 mm/tip), connected to the data logger (On- for the analysis as a combination of several methods allows the set HOBO Event). Rainfall on the rooftop was measured with a verification of the results of an individual method and enables a disdrometer (OTT Parsivel), enabling also measurements of broader interpretation of the results. The regression tree model rainfall microstructure, i.e. raindrop diameter, raindrop veloci- is designed by repeating the divisions of the influential varia- ty, and the number of raindrops. The measuring area of the bles and by adapting a simple prediction model for the target disdrometer is 54 cm and the measured data are allocated to variable within each division. The result of the division process one of the 32 drop diameter classes (ranging from 0.312 mm to is shown graphically with a decision or regression tree (Loh, 24.5 mm) and 32 velocity classes (ranging from 0.05 m/s to 2011; Zabret et al., 2018). As a target variable, throughfall 20.8 m/s). The drop diameters smaller than 0.312 mm were (TF), stemflow (SF), and rainfall interception (I) were set. Each assigned to the smallest drop diameter class, as they are outside model was run six times per observed year (namely, 2014, the device’s measurement range. The recorded time series data 2015, and 2016), once per each target variable, taking into from the rain gauge and the disdrometer were used to identify account all influential variables and also the variables without the rainfall events (separated with at least a 4-hour dry period) data on the rainfall microstructure due to the longer period and their characteristics (duration and intensity). The 4-hour without available data (year 2015 was excluded). The influen- dry period was selected to divide the events based on the obser- tial variables included in the analysis (Table 1) were the total vations of the rainfall and throughfall dynamics at the field, as rainfall amount per event (Pa), the average rainfall event inten- during the wetter time of the year throughfall lasted for quite sity (Pi), the total duration of the rainfall event (Pd), the average some time after the cessation of the rainfall. Shorter rainfall air temperature (T), and the vapour pressure deficit (VPD) interruptions were captured as part of the defined events. The during an event, the average wind speed (Ws) and the direction dry period was defined with an accuracy of 0.2 mm of rainfall (Wd) per event, the dry period duration before a rainfall event (equal to the volume of the rain gauge tipping bucket). (DryP), the time when an event occurred, namely during the The tree characteristics were determined in individual sur- day, the night, or both (DN), the phenoseason (Feno), the aver- veys. The photographs of the trees were taken at a required age raindrop diameter (DropD), the velocity (DropV) per event, distance to avoid deformation of proportions and were used to the median volume diameter of an event’s raindrops (MVD), determine the tree height, the area of the projected canopy, and and the number of raindrops delivered per event (DropNr). the branch inclination. The diameter at breast height was calcu- lated from the measured perimeter of the stem. The bark stor- Table 1. Influential variables included in the analysis. age capacity was determined from the bark samples, extracted using a steel hole puncher, according to the procedure described Variable Abbreviation Unit by Perez-Harguindeguy et al. (2013). Phenoseasons were de- Rainfall amount per event P mm termined based on the regular measurements of LAI, performed Average rainfall event intensity per P mm/h event with LAI-2200c Plant Canopy Analyzer (LI-COR) following Total duration of the rainfall event P h the protocol for isolated trees (Li-COR, 2015). The canopy d storage capacity was calculated from the observed rainfall and Average air temperature during the T °C throughfall data according to the Leyton graphical method event Average vapour pressure deficit VPD kPa (Leyton et al., 1967). during the event Average wind speed during the event Ws m/s Data analysis Average wind direction during the W ° event Measured data of rainfall precipitation (P), throughfall (TF), Dry period duration before the DryP h and stemflow (SF), collected in years 2014, 2015, and 2016, rainfall event were used in the analysis. Based on these data, the third com- Time when the event occurred, name- DN – ponent of rainfall partitioning, i.e. rainfall interception (I), was ly during the day, the night, or both calculated for each event: Phenoseason Feno – Average raindrops diameter of the DropD mm I = P – TF – SF (1) drops, observed during the event Average raindrops velocity of the DropV m/s In the selected period, 413 rainfall events were observed in drops, observed during the event total, but not all of them were included in the analysis. Snow Median volume diameter of an MVD mm and sleet events were excluded in the initial phase, while during event’s raindrops the further preparation of the data, the events without complete Number of raindrops delivered per DropNr – time series on rainfall, throughfall, and stemflow due to clog- event ging of the measurement equipment were also excluded. There- fore 365 rainfall events were taken into account in the analysis, capturing 86% of the total rainfall, delivered in the analysed The BRT method combines two algorithms, regression trees period. Additionally, the disdrometer was not operational due to and boosting (Elith et al., 2008), which improve the efficiency a software error for a longer time period during 2015. There- of an individual model and provide a better understanding of fore, rainfall microstructure data were not included in the anal- the results with additional factors. Boosting is based on the ysis for this year. assumption that the average of many raw predictions, which are For the selected rainfall events, the influence of the variables upgraded after every single repetition, will result in a better describing general weather conditions was evaluated using two final model. The sequential approach of the step-by-step meth- 458 Relation of influencing variables and weather conditions on rainfall partitioning by birch and pine trees od iteratively adjusts and improves the model based on a set of age annual precipitation, 2014 was recognised as a wet, 2015 as training data (Elith et al., 2008). Due to the larger number of a dry, and 2016 as an average year. During 2014 we registered model runs, it is also possible to estimate the impact of an indi- 167 events, delivering 1575 mm of rainfall. For this year, the vidual variable on the design of the model and thus on the total rainfall amount (1841 mm) was 36% larger than the aver- target variable. Friedman (2001) presented an equation that can age long-term yearly rainfall amount of 1355 mm measured at be used to estimate the relative influence (RI) of each variable the Ljubljana-Bežigrad meteorological station. On the contrary, included in the BRT model. The RI is based on how many in 2015, we recorded 85 events, delivering 931 mm of rainfall. times a variable has been selected in the model to divide the The total delivered rainfall (1106 mm) was 18% smaller than regression tree. The number of selections is weighted by the the long-term average rainfall amount per year (1355 mm). square of the model improvement rate as a result of each split Furthermore, the year 2016 was similar to an average one, as and expressed as an average with respect to all generated re- we observed 113 rainfall events delivering 1139 mm of rainfall. gression trees (Friedman and Meulman, 2003). The RI is ad- Through the entire year, 1317 mm of rainfall was measured, justed so that the sum of the RI values of all considered varia- which is comparable to a long-term average precipitation of bles equals 100, making the higher values directly indicating a 1355 mm at the Ljubljana-Bežigrad meteorological station. greater influence of the variable. Although during the dry year 2014 the largest number of the The BRT models were implemented using the “gbm” pack- rainfall events were recorded, they on average delivered the age (Ridgeway, 2020) in R software (R core team, 2020). In the smallest amount of rainfall per event (9.4 mm) and on average initial phase we determined the arguments of the model, using lasted for the shortest time (5.7 h), but were on average the 75% of the whole data set for training and 25% of the data for most intense (2.1 mm/h) (Figure 1). The average rainfall inten- testing of the model, implementing 50 iterations for each set of sity and duration of rainfall events during the years 2015 and the arguments and calculating the RMSE value of predictions 2016 were similar (average intensity of 1.4 mm/h and 1.5 from all iterations. When adjusting the model, various number mm/h, respectively and average duration of 8.0 h and 8.1 h, of regression trees (15000, 1500, and 500) and values of the respectively); however, the events in the dry year 2015 deliv- shrinkage parameter (0.001, 0.01, 0.05) were applied. According ered on average more rainfall (11.0 mm) than the events in the to the results, the final BRT models were estimated, taking into average year 2016 (10.1 mm per event on average). account the Gaussian distribution, 1500 trees, a shrinkage pa- Comparing the climate conditions in the considered years rameter of 0.01, and 5 cross-validation folds. only slight differences were observed for the wind characteris- Random Forest (RF) is an ensemble-learning algorithm, tics, vapour pressure deficit, and air temperature. However, a which merges the concepts of regression trees and bagging noticeably shorter dry period between the events was observed (Breiman, 2001). Bagging is a procedure enabling growing of in the wet year 2014 (40 h on average) comparing to the years regression trees from different subsets in order to avoid highly 2015 and 2016 (58 h and 56 h, respectively). The rainfall events correlated predictors. This algorithm relies on random selection characteristics in the considered years also differ according to of trees to describe the reliable relationship between the target the rainfall microstructure. The size of the rainfall drops was and the influential variables. Cases are randomly selected from significantly different (p < 0.001) during the wet year 2014 a data set, a random sample is used to design an individual comparing to the years 2015 and 2016, as in the year 2014 an regression tree, and predictions are formed for the remaining average raindrop diameter was equal to 0.85 mm and MVD was cases. The model repeats this process several times. Random- equal to 1.79 mm, while during the years 2015 and 2016 the ness is additionally ensured by imposing different randomly drop diameter on average accounted to 0.67 mm and 0.62 mm selected sets of influential variables on each division. This is and MVD to 1.51 and 1.44 mm, respectively. However, the possible due to random and repeated selection of individual larger raindrops resulted in the smaller number of drops per target values and influential variables (Breiman, 2001). For event, as the lowest number of raindrops was on average de- each variable the variable importance measure is also estimated tected in the wet year 2014 (Figure 1). (Breiman et al., 2018). The variable importance gives the total The values of rainfall partitioning components were quite decrease in node impurities from splitting on the variable, aver- similar for the years 2014 and 2016, while some deviations are aged over all trees. In case of regression, as presented here, it is observed for the values measured in 2015, when higher values measured by the residual sum of squares. of throughfall and stemflow proportions according to the rain- The RF models were built in R software (R core team, fall in the open were observed (Figure 2). In general, over all 2020), using package “RandomForest” (Breiman et al., 2018). three observed years, throughfall under the birch tree was on In the first phase of the model establishment, we divided the average equal to 53% (± 34%), average stemflow was 1.2% (± data set into a training (75%) and test (25%) set. The model 2.5%), and average rainfall interception was 46% (± 35%). arguments were selected one by one, applying numerous itera- Throughfall under the pine tree was on average lower than tions for each of the 30 models. For the number of variables under the birch tree, resulting in 27% (± 26%) of rainfall in the randomly sampled as candidates at each split (mtry), the values open, while stemflow accounted for only 0.03% (± 0.10%) and between 10 and 40 were tested, using the “tune” function. The the rainfall interception by the pine tree on average presented maximum number of terminal nodes of the trees (maxnode) 73% (± 26%) of rainfall in the open. was applied for the values between 5 and 30, while the number of the trees to grow (n.trees) was tested for values between 250 Influence of the rainfall event characteristics on throughfall and 5000. For these two arguments the best value was selected according to the RMSE and R values of the iteration results. Both of the applied models, namely BRT and RF, indicate that throughfall under the birch trees is influenced by the larger RESULTS number of variables than throughfall under the pine trees, re- gardless the year (Figure 3). Throughfall (TF) under the birch The analysed data on rainfall partitioning were collected trees in the wet year 2014 was the most dependent on the rain- during the years 2014, 2015, and 2016. These years were hy- fall amount (P ) and intensity (P ), rainfall duration (P ), and a i d drologically quite distinct, as according to the long-term aver- the average vapour pressure deficit (VPD) during the rainfall 459 Katarina Zabret, Mojca Šraj Fig. 1. Boxplots of considered rainfall event characteristics for each analysed year. Fig. 2. Measured throughfall (TF) and stemflow (SF) by birch and pine trees per rainfall event according to the observed year. event (Figure 3). Rainfall amount and intensity demonstrated wind speed according to the RF model. In case of data for 2016, between 18% and 20% of relative influence (RI) each by both the inclusion of rainfall microstructure variables does not affect applied methods, while RI for the first four most influential the order of the influencing factors (Figure 3). As the most in- variables exceeded 60% in total. However, when taking into fluencing variable, the rainfall amount is still recognised by both account also the variables describing the rainfall microstructure, applied models, however the second most influencing variable, the number of raindrops (DropNr) became the most influencing having a similar value of RI, is the number of raindrops. In this variable, indicating the amount of throughfall by birch in the case both variables together represent 45% and 60% of RI ac- wet year 2014. cording to the BRT and RF model, respectively. For the throughfall under the birch trees during the dry year The number of influencing variables according to the domi- 2015, both models assigned a similar relative influence of al- nant value of the relative influence in the case of throughfall most 30% to rainfall intensity, indicating this variable as the under the pine tree is more straightforward (Figure 3). Rainfall most significant in addition to the rainfall amount. The BRT amount was recognized to be the most influencing variable model also recognized air temperature and vapour pressure regardless the year, with an average RI between 43% (RF for deficit as the influential variables with RI of 10%, while the 2014) and 82% (RF for 2016). Both models also recognized the random forest model assigned more than 8% of RI to rainfall influence of rainfall intensity and duration on throughfall by duration and wind speed (W ) (Figure 3). pine trees in 2014, while in 2015, more than 8% of RI was The data collected during the average year 2016 showed a assigned to wind speed. In 2016, in addition to the rainfall significant influence of the rainfall amount only, as it represent- duration air temperature was the second most influencing vari- ed almost 40% of RI according to the BRT model and more able with RI larger than 5%. None of the rainfall microstructure than half of the total RI expressed by RF model. More than 9% variables exceeded more than 6% of RI, regardless the applied of RI was assigned also to wind speed and vapour pressure model or the year observed in case of throughfall under the pine deficit according to the BRT method and to air temperature and trees. 460 Relation of influencing variables and weather conditions on rainfall partitioning by birch and pine trees Fig. 3. Relative influence (RI) of the considered variables for throughfall (TF) by the birch and pine trees according to the observed years, evaluated by the boosted regression trees (BRT) and random forest (RF) models. Influence of the rainfall event characteristics on stemflow encing variables, combined with the number of raindrops. Stemflow in the wet year 2014 is still the most influenced by Similarly as throughfall, stemflow is in general the most in- the rainfall amount, while the number of raindrops and MVD fluenced by the rainfall amount (Figure 4). Stemflow (SF) by were also recognized as more influential. However, for the birch tree was the most characterized by the rainfall amount stemflow in 2016, the number of raindrops together with the regardless the year as the RI for this variable ranged between rainfall amount and duration were recognized as the variables 35% (RF, year 2014) and 61% (RF, year 2015). Stemflow by with the highest RI (together accounting for 59% according to the birch tree in the wet year 2014 and the average year 2016 the BRT and 82% according to the RF model). was also highly influenced by the rainfall duration, which had The amount of stemflow by the birch trees was similarly in- the second highest RI in both years, regardless the model used. fluenced during the years 2014 and 2016, however for the pine On the contrary, in the dry year 2015 stemflow by the birch trees similarities can be observed between the years 2015 and tree was affected by a larger number of variables (Figure 4). 2016 (Figure 4). Stemflow by the pine trees during 2014 was The BRT model indicated that in addition to the rainfall the most influenced by wind direction, followed by the rainfall amount, stemflow by the birch tree is also influenced by rainfall amount. In case of the BRT model these two variables resulted intensity, wind speed, vapour pressure deficit, and rainfall in RI of 77%, while in case of the RF model, the influencing duration, as RI for all mentioned variables was larger than 9% variables with RI of more than 10% are also vapour pressure (Figure 4). However, according to the RF model, the value of deficit, wind speed, and air temperature. RI higher than 10% was estimated for the dry period duration When also including the rainfall microstructure variables, and air temperature. the influence of wind direction is minimized, as rainfall When taking into account also the rainfall microstructure amount, duration, and the number of raindrops in combination characteristics, the rainfall amount is still one of the most influ- with MVD (estimated by the BRT model) and air temperature 461 Katarina Zabret, Mojca Šraj Fig. 4. Relative influence (RI) of the considered variables for stemflow (SF) by the birch and pine trees according to the observed years, evaluated by the boosted regression trees (BRT) and random forest (RF) models. (estimated by the RF model) resulted in RI higher than 50%. Rainfall interception by birch and pine trees is the most in- Stemflow by the pine trees in 2015 and 2016 is significantly fluenced by the rainfall amount, which has the highest values of influenced by the rainfall amount and intensity, as regardless RI according to both models. In case of the birch trees the val- the model or the year, these two variables present between 64% ues of RI for the amount of rainfall ranged between 22% and and 85% of RI. The substantial influence of the rainfall amount 63%, while in case of the pine trees they were even higher, and intensity is also retained when introducing the rainfall ranging from 47% to 83%. Comparing these values to RI esti- microstructure influence. In this case, as the second most influ- mated for throughfall, the values were a bit larger in case of the encing variable with RI larger than 10% both models recog- birch trees, while for the pine trees they were kept in a similar nised MVD. range. Rainfall interception of the birch trees was in the wet year Influence of rainfall event characteristics on rainfall 2014 also significantly influenced by the rainfall duration and interception intensity, while in the dry year 2015 it was mainly influenced by rainfall intensity and in the average year 2016 by vapour Rainfall interception (I) is calculated as the difference be- pressure deficit (according to the BRT model) and air tempera- tween the measured values, i.e. rainfall amount in the open, ture (according to the RF model). In case of the pine trees the throughfall, and stemflow (Eq. 1). Therefore, as the amount of results were also very similar to the ones for the throughfall; in throughfall is much larger than stemflow, this is the value that 2015 and 2016 only the rainfall amount played a significant mainly determines the proportion of intercepted rainfall, result- role in the process of rainfall interception, while in the wet year ing in similarly evaluated influencing variables as throughfall 2014 also rainfall intensity and duration demonstrated RI values (Figure 3). larger than 10% (Figure 5). 462 Relation of influencing variables and weather conditions on rainfall partitioning by birch and pine trees Fig. 5. Relative influence (RI) of the considered variables for rainfall interception (I) by the birch and pine trees according to the observed years, evaluated by the boosted regression trees (BRT) and random forest (RF) models. The results of both applied models considering also the rain- account results of the BRT instead of the RF model. Thus, the fall microstructure are also similar to the results of throughfall combined analysis of the two methods allows for a more com- data analysis (Figure 3). In case of the birch trees, the number prehensive evaluation of the results, as the RF model indicates of drops was recognised as a variable with the highest influence the most influencing variables, while the BRT model highlights among the newly introduced variables, while in case of the pine also the other possible variables with meaningful influence. trees for none of these variables the estimated RI exceeded 6% The results demonstrate that throughfall, stemflow, and rain- (Figure 5). fall interception by birch and pine trees were the most influ- enced by the amount of rainfall, which has been repeatedly DISCUSSION recognized as the factor most influencing the rainfall partition- ing components in general also in other studies (e.g., Levia and Although the two methods are very similar as they are both Germer, 2015; Staelens et al., 2008; Su et al., 2019; Zabret et based on the principle of regression trees, there is one main al., 2018). In case of both considered tree species, rainfall dura- difference if we consider the method associated with the regres- tion seems to play an important role mainly during the wet year sion trees (boosting and bagging). This is also reflected in the 2014, while rainfall intensity had a significant influence on estimation of the most influential variables and their RI values. rainfall partitioning by birch trees during the dry year 2015. A comparison of the results by the two models shows that in This observation seems to correlate well with the results pre- general, the RI values of the variables estimated by the RF sented by Mużyło et al. (2012), who observed a significant model are higher than those estimated by the BRT model (Fig- influence of rainfall duration on throughfall in a deciduous ures 3–5). Therefore, the number of the variables for which the forest, especially during the leafless season. The leafless season RI value exceeds the threshold value is larger when taking into is usually characterized by more precipitation and generally 463 Katarina Zabret, Mojca Šraj wetter months, which may be equivalent to the hydrologically The relative influence, estimated by the BRT and RF mod- wetter year of 2014, in which a more pronounced influence of els, shows that throughfall under the birch trees is determined rainfall duration was observed in this study (Figures 3–5). by a larger number of influencing variables. In addition to the As the wetter year 2014 can be correlated with the wetter rainfall amount, duration, and intensity, also air temperature leafless phenoseason, the drier year 2015 is expected to be and vapour pressure deficit (VPD) were assigned with values of associated with the drier leafed period. Therefore, the influence RI larger than 8%. Air temperature and VPD are closely con- of rainfall intensity on rainfall partitioning in the drier year nected to the season of the year, corresponding also to the phe- 2015 is initially unexpected. Rainfall intensity was actually noseasons, and are especially significant for a deciduous birch recognized as one of the most influential variables in previous trees (Zabret et al., 2018). Therefore, the significant RI values studies, but its effect was observed for winter throughfall (Xiao of air temperature and VPD may indirectly indicate the influ- et al., 2000), rainfall interception in the leafless period (Zabret ence of phenoseasons on throughfall by birch, which is larger in et al., 2018), and rainfall interception in a wet year (Zabret and the leafless period, characterized by lower air temperature and Šraj, 2019b). However, for a beech tree, Staelens et al. (2008) lower VPD values (Andre et al., 2008; Brasil et al., 2020; reported significant influence of rainfall intensity on stemflow, Mużyło et al., 2012; Šraj et al., 2008; Zabret and Šraj, 2018; especially during the leafed period resulting in a decrease in the Zabret et al., 2021). However, the relation between the influ- stemflow amount due to splashing of droplets intercepted by ence of phenoseasons and meteorological variables on rainfall the canopy and forming throughfall instead of stemflow. Addi- partitioning has already been recognized as a very complex one tionally, a more evident influence of rainfall intensity was (e.g., Andre et al., 2008; Mużyło et al., 2012; Zabret and Šraj, estimated by both applied models for birch rather than for pine 2021). When analysing the influence of air temperature and trees (Figures 3–5). A different influence of rainfall intensity on VPD on throughfall by the birch tree, the results are similar tree species with distinct vegetation properties was already among the years (Figure 6). Throughfall is in general decreas- observed in other analyses (e.g., Sadeghi et al., 2020; Siegert ing with increasing air temperature, which was observed also and Levia, 2014; Zabret et al., 2018). Birch trees have a by Staelens et al. (2008). Warmer months of the year are also smoother bark surface and more flexible leaves compared to the characterized with a fully leafed canopy, also decreasing the rougher and more absorbent bark of pine trees and its compact throughfall, while a higher air temperature increases the evapo- needles, therefore the process of splashing of intercepted drop- ration, which may also lead to a decrease in throughfall (Šraj et lets may be more intense in the canopy of the birch trees. al., 2008; Xiao et al., 2000). However, the response of Fig. 6. Partial dependence plots of the influence of air temperature (T) and vapour pressure deficit (VPD) on throughfall (TF) by birch trees during the considered years. 464 Relation of influencing variables and weather conditions on rainfall partitioning by birch and pine trees throughfall according the VPD values is similar for the years The variables with the highest influence expressed by both 2014 and 2015, as lower VPD up to 1.5 kPa increases through- models were the rainfall amount and the number of raindrops. fall under the birch tree, while larger values of VPD decrease The comparison of the influential variables indicates to some the amount of throughfall (Figure 6). The data collected in 2016 extent the correlation between the wet period and the leafless show a bit different response of throughfall according to the season, as well as between the dry period and the leafed season. VPD values up to 2 kPa, while larger VPD values decrease For example, rainfall duration had a high relative influence on throughfall under birch trees as well. rainfall partitioning by both tree species mainly in the wet year Pine tree’s stemflow was influenced by more variables com- 2014, while researchers already reported its influence in the pared to throughfall. In addition to the rainfall amount, which leafless season. Stemflow by birch trees was also strongly mainly determines stemflow under both tree species, rainfall influenced by air temperature and vapour pressure deficit, intensity and rainfall duration also had an important influence which are dependent on the season of the year, which is also on stemflow under the pine and birch trees, respectively. The consistent with the phenoseason. However, the results of the significance of rainfall duration on stemflow by the birch tree models also indicate significant differences in the response of was also recognised in the analysis of the stemflow response the two tree species. The influence of rainfall intensity, the (Zabret and Šraj, 2021), as well as in the analysis of predicting number of raindrops, and the median volume diameter was more the stemflow of a birch tree using the regression trees (Zabret et pronounced in the case of the birch trees, while it was negligible al., 2018). Although rainfall intensity was recognised as a less in the case of the pine trees. This observation coincides with the influential variable in the case of throughfall under the pine conclusions of previous studies, i.e. that raindrops behave trees, it seems to have a larger influence on its stemflow. Re- differently when interacting with needles or leaves. sults of both implemented models (BRT and RF) indicate RI The presented analysis mainly confirms all previous obser- values larger than 15% for rainfall intensity in the years 2015 vations made by other researchers about the different influences and 2016 (Figure 4). This is consistent with the results of the on the rainfall partitioning process by distinct tree species. BRT model applied to stemflow data of a leafed phenoseason However, a new insight into the impact of wet and dry period is (Zabret et al., 2018), which also indicates some similarities in presented, indicating that during a longer wet period the trees meteorological influences during the leafed phenoseason and behave similarly as in the leafless period and during the longer dry period the rainfall interception process is similar as that the drier hydrological year. in Introduction of variables specifying the rainfall microstruc- the leafed period. Nevertheless, additional analysis, taking into ture into the analysis expressed significant influence of the account multiple wet and dry periods as well as data for these number of rain drops on throughfall and rainfall interception by periods for other tree species and other locations with different the birch trees. The number of raindrops as well as the mean microclimatic characteristics, should be implemented in order volume diameter (MVD) were estimated to have considerable to understand this aspect in more detail. influence on stemflow in case of both considered tree species. However, no influence of these variables on throughfall under Acknowledgements. The work was founded by the Slovenian the pine trees was observed, as throughfall under the pine trees Research Agency (ARRS) through research program P2-0180. was still the most influenced by the amount of rainfall, which provided more than half of the RI according to the other con- REFERENCES sidered variables. A more noticeable influence of rainfall mi- crostructure on throughfall by birch than by pine trees was also Andre, F., Jonard, M., Ponette, Q., 2008. Influence of species confirmed in previous study (Zabret et al., 2018), in which a and rain event characteristics on stemflow volume in a tem- similar comparison of data with BRT models per phenoseason perate mixed oak–beech stand. Hydrol. Process., 22, 4455– showed the influence of MVD on throughfall by pine trees only 4466. in the leafless season, while the number of raindrops had a ARSO, 2020. Measurements archive. http://www.meteo.si/met/ significant role in regulating throughfall under the birch trees sl/archive/ (Accessed 5 May 2021). regardless the phenoseason. This might be connected to the Bezak, N., Mikoš, M., 2014. Estimation of design floods using distinct characteristics of the foliage of the considered tree univariate and multivariate flood frequency approach with species, i.e. leaves of the birch trees and needles of the pine regard to one wet year. Acta hydrotechnica, 27, 103–117. trees. The different interaction of needles and leaves to the rain Brasil, J.B., de Andrade, E.M., de Queiroz Palácio, H.A., dos drops and their characteristics has been already reported by Santos, J.C.N., Medeiros, P.H.A., 2020. Temporal variabil- other researchers (e.g., Holder, 2013; Nanko et al., 2016; Zabret ity of throughfall as a function of the canopy development et al., 2017; Zabret et al., 2018). stage: from seasonal to intra-event scale. Hydrol. Sci. J., 65, 1640–1651. CONCLUSIONS Breiman, L., 2001. Random Forests. Mach. Learn., 45, 5–32. Breiman, L., Cutler, A., Liaw, A., Wiener, M., 2018. Package The rainfall partitioning process is part of the hydrological ‘RandomForest’. https://cran.r-project.org/web/packages/ cycle, for which changes are expected due to more pronounced randomForest/randomForest.pdf (Accessed 25 March 2021) precipitation patterns, resulting in more intense wet and dry Dohnal, M., Černý, T., Votrubová, J., Tesař, M., 2014. Rainfall periods. Accordingly, the influence of meteorological factors interception and spatial variability of throughfall in spruce on throughfall, stemflow, and rainfall interception during a wet, stand. J. Hydrol. Hydromech., 62, 277–284. a dry, and an average year was analysed. Two similar statistical Elith, J., Leathwick, J.R., Hastie, T., 2008. A working guide to methods based on the regression tree approach were applied, boosted regression trees. J. Anim. Ecol., 77, 802–813. namely boosted regression trees and random forest. The com- Friedman, J.H., 2001. Greedy function approximation: a gradi- parison of the results showed that the methods are complemen- ent boosting machine. Ann. Stat., 29, 1189–1232. tary, since the BRT model indicates numerous variables with Friedman, J.H., Meulman, J.J., 2003. Multiple additive regres- relevant influence and the RF model highlights the variables sion trees with application in epidemiology. Stat. Med., 22, with the highest influence. 1365–1381. 465 Katarina Zabret, Mojca Šraj Gao, S., Zhou, T., Yi, C., Shi, P., Fang, W., Liu, R., Liang, E., Partitioning by Vegetation: A Global Synthesis. Springer Camarero, J.J., 2020. Asymmetric impacts of dryness and Nature, Berlin, Germany, pp. 49–70. wetness on tree growth and forest coverage. Agr. Forest. Schooling, J.T., Carlyle-Moses, D.E., 2015. The influence of Meteorol., 288–289, 107980. rainfall depth class and deciduous tree traits on stemflow Hao, Y., Wang, Y., Mei, X., Huang, X., Cui, X., Zhou, X., Niu, production in an urban park. Urban Ecosyst., 18, 1261–1284. H., 2008. CO2, H2O and energy exchange of an Inner Mon- Siegert, C.M., Levia, D.F., 2014. Seasonal and meteorological golia steppe ecosystem during a dry and wet year. Acta Oec- effects on differential stemflow funneling ratios for two de- ologica, 33, 133–143. ciduous tree species. J. Hydrol., 519, 446–454. Holder, C.D., 2013. Effects of leaf hydrophobicity and water Staelens, J., De Schrijver, A., Verheyen, K., Verhoest, N.E.C., droplet retention on canopy storage capacity. Ecohydrology, 2008. Rainfall partitioning into throughfall, stemflow, and 6, 483–490. interception within a single beech (Fagus sylvatica L.) cano- Honda, E.A., Mendonça, A.H., Durigan, G., 2014. Factors py: influence of foliation, rain event characteristics, and me- affecting the stemflow of trees in the Brazilian Cerrado. teorology. Hydrol. Process., 22, 33–45. Ecohydrology, 8, 1351–1362. Su, L., Xie, Z., Xu, W., Zhao, C., 2019. Variability of through- Hungate, B., Hampton, H., 2012. Valuing ecosystems for cli- fall quantity in a mixed evergreen-deciduous broadleaved mate. Nat. Clim. Change, 2, 151–152. forest in central China. J. Hydrol. Hydromech., 67, 225–231. Inglezakis, V.J., Poulopoulos, S.G., Arkhangelsky, E., Zorpas, Šraj, M., Brilly, M., Mikoš, M., 2008. Rainfall interception by A.A., Menegaki, A.N., 2016. Aquatic environment. In: Pou- two deciduous Mediterranean forests of contrasting stature lopoulos, S., Inglezakis, V. (Eds.): Environment and Devel- in Slovenia. Agr. Forest. Meteorol., 148, 121–134. opment: Basic Principles, Human Activities, and Environ- Xiao, Q., McPherson, E.G., Ustin, S.L., Grismer, M.E., Simp- mental Implications. Elsevier, pp. 137–212. son, J.R., 2000. Winter rainfall interception by two mature Klamerus-Iwan A., Link T.E., Keim R.F., Van Stan, J.T., 2020. open-grown trees in Davis, California. Hydrol. Process., 14, Storage and routing of precipitation through canopies. In: 763–784. Van Stan, J T., Gutmann, E., Friesen, J. (Eds.): Precipitation Xu, Z., Feng, Z., Zhao, C., Zheng, J., Yang, J., Tian, F., Peng, Partitioning by Vegetation: A Global Synthesis. Springer H., Wang, C., Peng, S., Sher, H., 2013. The canopy rainfall Nature, Berlin, Germany, pp. 17–34. interception in actual and potential distribution of Qinghai Levia, D.F., Germer, S., 2015. A review of stemflow generation spruce (Picea crassifolia) forest. J. Hydrol. Hydromech., 61, dynamics and stemflow-environment interactions in forests 64–72. and shrublands. Rev. Geophys., 53, 673–714. Xu, L., Cao, G., Wang, Y., Hao, J., Wang, Y., Yu, P., Liu, Z., Leyton, L., Reynolds, E.R.C., Thompson, F.B., 1967. Rainfall Xiong, W., Wang, X., 2020. Components of stand water bal- interception in forest and moorland. In: Sopper, W.E., Lull, ance of a larch plantation after thinning during the extremely H.W. (Eds.): Forest Hydrology. Pergamon, Oxford, pp. 163– wet and dry years in the Loess Plateau, China. Global Eco. 178. Conser., 24, e01307. Loh, W., 2011. Classification and regression trees. Data Min. Yue, K., De Frenne, P., Fornara, D.A., Van Meerbeek, K., Li, Knowl. Disc., 1, 14–23. W., Peng, X., Ni, X., Peng, Y., Wu, F., Yang, Y., Peñuelas, Mużyło, A., Llorens, P., Domingo, F., 2012. Rainfall partition- J., 2021. Global patterns and drivers of rainfall partitioning ing in a deciduous forest plot in leafed and leafless periods. by trees and shrubs. Glob. Change. Biol., 27, 3350–3357. Ecohydrology, 5, 759–767. Zabret, K., 2013. The influence of tree characteristics on rainfall Nadbath, M., 2008. Meteorological station Ljubljana Bežigrad. interception. Acta Hydrotech., 26, 99–116. (In Slovenian.) Naše okolje 15, 1. (In Slovenian.) Zabret, K., Rakovec, J., Mikoš, M., Šraj, M., 2017. Influence of Nanko, K., Hudson, S.A., Levia, D.F., 2016. Differences in raindrop size distribution on throughfall dynamics under throughfall drop size distributions in the presence and ab- pine and birch trees at the rainfall event level. Atmosphere, sence of foliage. Hydrolog. Sci. J., 61, 620–627. 8, 240. Peng, Y., Chen, L., Tian, J., Sun, B., Jiang, C., Lu, Y., Shang, Zabret, K., Rakovec, J., Šraj, M., 2018. Influence of meteoro- J., 2021. Ecosystem services help alleviate the intensity of logical variables on rainfall partitioning for deciduous and dryness/wetness. Global Ecol. Conser., 27, e01581. coniferous tree species in urban area. J. Hydrol., 558, 29–41. Perez-Harguindeguy, N., Diaz, S., Garnier, E. et al. 2013. New Zabret, K., Šraj, M., 2019a. Evaluating the influence of rain event handbook for standardized measurement of plant functional characteristics on rainfall interception by urban trees using traits worldwide. Aust. J. Bot., 61, 167–234 multiple correspondence analysis. Water, 11, 2659. R core team, 2020. R: A language and environment for statisti- Zabret, K., Šraj, M., 2019b. Rainfall interception by urban trees cal computing. R Foundation for Statistical Computing, Vi- and their impact on potential surface runoff. Clean Soil, Air, enna, Austria. http://www.R-project.org/ (Accessed 20 Au- Water, 47, 8, 1800327. gust 2020) Zabret, K., Šraj, M., 2021. How characteristics of a rainfall Ridgeway, G., 2020. Generalized Boosted Regression Models. event and the meteorological conditions determine the de- https://cran.r-project.org/web/packages/gbm/gbm.pdf (Ac- velopment of stemflow: A case study of a birch tree. Front. cessed 10 August 2020) Front. For. Glob. Change, 4, 663100. Sadeghi, S.M.M., Gordon, D.A., Van Stan, J.T., 2020. A global synthesis of throughfall and stemflow hydrometeorology. In: Received 28 May 2021 Van Stan, J T., Gutmann, E., Friesen, J. (Eds.): Precipitation Accepted 5 August 2021

Journal

Journal of Hydrology and Hydromechanicsde Gruyter

Published: Dec 1, 2021

Keywords: Throughfall; Stemflow; Rainfall interception; Rainfall microstructure; Boosted regression trees; Random forest

There are no references for this article.