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

Learn More →

Potential use of hyperspectral data to classify forest tree species

Potential use of hyperspectral data to classify forest tree species Background: Remote sensing techniques and data are becoming increasingly popular in forest management, e.g. for change detection and health condition analysis. Tree species recognition is a fundamental issue in taking forest inventories, especially in carbon budget modelling. Hyperspectral imagery provides an accurate classification results for large areas based on a relatively small amount of training data. Results: A hyperspectral image of a forest stand in north-eastern Poland taken using an AISA (Airborne Imaging Spectrometer for Application) Eagle camera was transformed to extract the most valuable spectral differences and was classified into seven tree types (birch, European beech, oak, hornbeam, European larch, Scots pine, and Norway spruce) using nine classification algorithms. The highest overall accuracy and kappa coefficient were 90.3% and 0.9 respectively using three minimum noise fraction bands and maximum likelihood classifier. Conclusions: Hyperspectral imaging of forests can be used to classify major forest tree species with a good degree of accuracy. It is time-efficient and user-friendly; however, the data and software required means that this approach is still expensive at present. Keywords: AISA, Hyperspectral classification, Minimum noise fraction, Trees Background Einzmann et al. 2014). The biotic characteristics of the Remote sensing is the science of acquiring information whole stands include leaf and branch density, angular about objects or areas from a distance, typically from distribution, clumping, tree size compared to neighbours aircraft or satellites. (Leckie et al. 2005, Korpela et al. 2011), and lichens, In tree canopies, the amount of radiation reflected in mosses, herbaceous vegetation, lianas, or other epiphytes regions of different wavelengths is related to the chem- (Clark and Roberts 2012). The abiotic characteristics of ical and physical properties of single trees as well as bi- whole stands include topography, soil type (and its influ- otic and abiotic characteristics of an entire stand. ence), moisture, and microclimate (Portigal et al. 1997). Among the chemical properties of single trees are the It is hardly feasible to identify species-specific absorp- levels of lignin, cellulose, nitrogen, chlorophyll, caroten- tion features in the visible and near-infrared (VIS-NIR) oids, anthocyanins (Asner 1998; Clark et al. 2005; Grant spectral region; however, this is much easier in the 1987; Clark and Roberts 2012; Ustin et al. 2009), and shortwave infrared (SWIR) spectral region (Asner 1998). water (Asner 1998, Gao and Hoetz 1990, Zarco-Tejada Salisbury (1986) presented leaf-level thermal infrared et al. 2003, Lee et al. 2010); these affect the health status (TIR) spectra of four species and identified well-defined of the trees (Waser et al. 2014). Among the physical and notably different spectral features. Salisbury and properties of single trees are leaf and wood morphology, Milton (1998) obtained close-range thermal reflectance transmission characteristics (Asner 1998; Clark et al. measurements for several other species and reported dif- 2005; Grant 1987; Clark and Roberts 2012; Ustin et al. ferences in the spectra in most of them. Ribeiro da Luz 2009), vertical leaf area density (Treuhaft et al. 2002), and Crowley (2007) found that TIR spectra were associ- and age (Ghiyamat et al. 2013; Roberts et al. 1997; ated with several chemical and structural compounds of plants such as cellulose, silica, xylan, and oleanolic acid * Correspondence: t.hycza@ibles.waw.pl levels, and reported that TIR signals were much more Department of Geomatics, Forest Research Institute, Braci Leśnej 3, 05-090 species-specific than the reflectance signals observed in Raszyn, Sękocin Stary, Poland © The Author(s). 2018 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. Hycza et al. New Zealand Journal of Forestry Science (2018) 48:18 Page 2 of 13 the visible, shortwave, and infrared regions. Many plants provided later in this paper. This method (MNF) was used develop chemical and aromatic compounds that might by Zagajewski (2010) to classify vegetation in the Tatra help define species-specific middle infrared and TIR sig- Mountains, by Olesiuk and Zagajewski (2008) to classify natures (Ribeiro da Luz and Crowley 2007; Ullah et al. the land cover of the Bystrzanka river drainage basin, and 2012). by Bartold (2008), Dian et al. (2014), and Han et al. (2004) Identifying tree species using remote sensing data is use- to classify forest tree species. Han et al. (2004) compared ful in the context of detecting changes (Adams et al. 1995) the results with those obtained by canonical transform- and managing water stress (Cho et al. 2010; Fassnacht et al. ation, while Harsanyi (1994)used the ‘orthogonal sub- 2016). It helps in the development of sustainable manage- space projection’ method. This method eliminates the ment policies (Dalponte et al. 2012, Jones et al. 2010, response from non-targets while applying a filter to match Plourde et al. 2007, Heinzel and Koch 2012, Kennedy and desired targets in the data, and is most efficient and effect- Southwood 1984) and performance of forest resource (Van ive when the target signatures are distinct. Aardt and Wynne 2007) and single tree inventories (Kor- Algorithms such as the Pixel Purity Index (PPI) pela and Tokola 2006; Immitzer et al. 2015;Tompalski et (Zagajewski 2010, Olesiuk and Zagajewski 2008, Bartold al. 2014). It enables the assessment and monitoring of bio- 2008) or linear spectral unmixing (LSU), which produces diversity, species compositions (Shang and Chisholm 2014; ‘maps of abundance’ in which each pixel is assigned to Wulder et al. 2006), wildlife habitats (Jansson and Angel- more than one class with a specified probability level (Luo stam 1999;Pausas etal. 1997), invasive species migrations and Chanussot 2009; Villa et al. 2013;Li et al. 2014), can (Chambers et al. 2013; Van Ewijk et al. 2014), and in the be used to extract the pixels most useful for the classifica- understanding of tree ecology (Chambers et al. 2013,Van tion (endmembers). Schull et al. (2010)alsoused pure Ewijk et al. 2014). It can also be applied to the estimation of spectral pixels to classify forests in the north-eastern USA insect abundance in forests (Kennedy and Southwood and achieved an overall accuracy of 92%. 1984) and the development of species-specific growth and The ability to successfully classify forest tree species yield models as well as allometric equations (Ørka et al. using hyperspectral data was proven for forests in the 2013; Vauhkonen et al. 2014). equatorial zone (Clark et al. 2005; Mickelson et al. 1998; Proper forest management and planning based on accur- Peerbhay et al. 2013; Goodwin et al. 2005), when seven ate distinction of tree species requires highly accurate clas- tree species were classified using linear discriminant ana- sification maps that cannot yet be produced using the lysis (LDA), maximum likelihood (ML), and spectral angle multispectral images typically acquired in four to eight wide mapping (SAM) methods, with accuracies of 80 to 100%. spectral bands. Hyperspectral data are more useful for clas- The hyperspectral data were also used in the tropical and sifying tree species: the only condition is that the species sub-tropical zones (Dalponte et al. 2008,Dian et al. 2014, must appear significantly different in the spectral reflect- Dennison and Roberts 2003, Lucas et al. 2008,Yang et al. ance measured in dozens of narrow spectral intervals (Clark 2009,Gong et al. 1997, Van Aarst and Norris-Rogers et al. 2005, Heinzel and Koch 2012, Carlson et al. 2010, 2008, Bellanti et al. 2016) with accuracies of over 90% and Dalponte et al. 2010, Dalponte et al. 2011,Stavrakoudiset in the temperate zone (Zagajewski 2010; Olesiuk and al. 2014,Farreiraetal. 2016). The reflectance of individual Zagajewski 2008; Bartold 2008;Dianet al. 2014;Martinet tree species is dependent on numerous factors, and the dif- al. 1998; Dalponte et al. 2013; Dmitriev 2014; Tarabalka ferences are sometimes too subtle to be observed using 2010; Richter et al. 2016)with accuracies of74 to 93%. wide, multispectral bands (Dalponte et al. 2009;Mickelson Classification results may be improved using hyper- et al. 1998). Since the technology was released, the cost of spectral data with light detection and ranging (LIDAR) hyperspectral images has decreased gradually. It is expected data (Alonzo et al. 2014). For the temperate and that it will be soon possible to use hyperspectral imagery to sub-tropical (Hainzel and Koch 2012; Dalponte et al. study forest ecology and develop management and planning 2008; Caiyun and Fang 2012) zones, the accuracies were techniques (Innes and Koch 1998;Dalponteetal. 2008; over 80%. Passive optical systems, particularly hyper- Voss and Sugumaran 2008). spectral ones, generally showed higher potential for tree However, hyperspectral images contain a huge amount species classification than active synthetic aperture radar of auto-correlated data. Principal component analysis (SAR) or LIDAR sensor systems. However, LIDAR data (PCA) is often used to solve this problem (Zagajewski have proven suitable for regions with a low number of 2010; Olesiuk and Zagajewski 2008; Bartold 2008). This species (Fassnacht et al. 2016). Forest stands classified widely known technique creates a set of artificial bands with the highest accuracy in the European temperate in which each band is less informative than the previous zone include mostly homogenous ones, dominated by one. The minimum noise fraction (MNF) transformation Scots pine (Pinus sylvestris L.) and Norway spruce (Picea works in a similar manner but reduces the noise first. abies L.). Of the broadleaved species, European beech More detailed information on these transformations is (Fagus sylvatica L.) and oak (Quercus spp. L.) are Hycza et al. New Zealand Journal of Forestry Science (2018) 48:18 Page 3 of 13 classified with the highest accuracy, but these classifica- information about this area was obtained from www.mi- tions have lower accuracies than those of coniferous spe- lomlyn.olsztyn.lasy.gov.pl/zasoby-lesne. The size of the cies (Wietecha et al. 2017). area and relative tree composition is given in Table 1 and The aim of this study was to evaluate the accuracy detailed information is listed in Appendix 1. The individ- of tree species classification methods using a hyper- ual compositions of Scots pine (Pinus sylvestris L.) and spectral Airborne Imaging Spectrometer for Applica- European larch (Larix decidua L.) were not provided. The tion (AISA) Eagle image for a forested area in study area was a 15 km (10 km long and 1.5 km wide) northern Poland. The following algorithms were eval- rectangle including three lakes: SzelągWielki, Tabórz uated in the study: PCA and MNF transformation (to (southern part), and Długie (northern part) (Fig. 2.). reduce noise and auto-correlated data), parallelepiped A local survey was performed on 9.85 ha of the (P), minimum distance (MD1), Mahalanobis distance Miłomłyn Forest District using a series of circular test (MD2), ML, SAM, spectral information divergence plots (radius: 12.62 m; area: 500.34 m ) in March 2014. (SID), neural net (NN), and support vector machine The sample plots were surveyed individually to achieve (SVM) to perform the supervised classification). The the highest level of diversity for various forest character- results were evaluated using a set of 300 test pixels, istics (e.g. age, species, forest type), where the influence deployed randomly across the study/sample plots area, of slope was minimal (Fig. 2). We corrected for the in- to achieve the most reliable assessment of accuracy. fluence of slope on the tree-position measurements. Each tree with a diameter at breast height (dbh) over 5 Materials and methods cm was inventoried and had the following information Study area recorded: distance from centre of the plot, azimuth The survey was performed in the Miłomłyn Forest District (measured from the centre of the plot to each tree), de- in the north-eastern Poland (Fig. 1). Background foliation (assessed using an expert method), and height. Fig. 1 The study area Hycza et al. New Zealand Journal of Forestry Science (2018) 48:18 Page 4 of 13 Table 1 Basic information about the Miłomłyn Forest District Pre-processing (source: www.milomlyn.olsztyn.lasy.gov.pl/zasoby-lesne) The image was geometrically corrected by MGGP AERO Parameter Amount (UTM, Zone 34N, WGS-84). It was also subject to radiometric calibration (using the ENVI ‘Radiometric Forest area (ha) 19,000 calibration’ tool–calibration type: reflectance, output Total area (ha) 48,000 interleave: BSQ (band sequential), output data type: Most common tree species (%) float, scale factor: 1.00) and ‘Quick Atmospheric Correc- Scots pine, European larch 71 tion–Quack’ atmospheric correction (Sensor Type: European beech (Fagus sylvatica L.) 12 AISA) (Dalponte et al. 2012; Bernstein et al. 2005)(Ap- Grey alder (Alnus glutinosa L.) 6 pendix 2).These procedures were performed using ENVI 5.0. Birch (Betula spp. L.) 5 Oak (Quercus spp. L.) 4 Norway spruce (Picea abies L.) 1 Data reduction After the atmospheric correction, the amount of data Other (e.g. hornbeam Carpinus betulus L.) 1 was reduced. The image containing 129 bands was not an ideal data set with which to perform supervised clas- The centre of the test plot was determined using the sification, because it contained too much auto-correlated Pathfinder ProXT (Trimble, Sunnyvale, California), Glo- data. The reduction of the data may be performed using bal Navigation Satellite System (GNSS) which functions one of two types of methods: data transformation (PCA in the DGPS (Differential Global Positioning System) or MNF transformation) (Clark et al. 2005) or band se- mode. Its vertical and horizontal accuracy was estimated lection. Data transformation is fully automatic but is to be 1.4 m and 0.97 m respectively. Tree heights were based on artificial bands. Band selection is based on ori- measured using a Vertex IV device (Haglof Sweden AB, ginal bands but is also very subjective. Both methods Langsele, Sweden) and dbh was measured using a Codi- were tested. The data reduction was performed using mex calliper (Codimex, Warsaw, Poland). The data col- ENVI 5.0 software. lected were used to calibrate and verify the hyperspectral image classification process. No grey alder trees were found in the plots so this species was not considered fur- Classification ther. Although hornbeam occurs only occasionally in the Finally, four sets of data were classified (using four forest, it was found in one plot so was included in the algorithms): analysis. – The result of the PCA transformation—first three bands Data and software – The result of the MNF transformation—first seven The hyperspectral image was provided by MGGP AERO bands and taken by the AISA Eagle camera (SPECIM) on 3 – All 129 bands August 2013 at an altitude of 2303–2328 m (single – 36 original bands with the largest differences in the flight). The spectral resolution of the image was 400– spectral profiles generated from training pixels for 970 nm (129 spectral bands, 4–5 nm wide); the radio- each tree species metric resolution was 12 bits, while the spatial reso- lution was 1.5 m. The lens size was 18.5 mm and the To perform the supervised classification, it was im- field of view (FOV) was 37.7°. portant to choose representative pixels with which to The hyperspectral image classification process (as de- train the algorithm. This was performed using two tailed below) was performed using ENVI 5.0 (developed MNF band compositions and the data from the test by Exelis Inc.), ArcGIS 10.3 (developed by ESRI), and plots. A total of 260 training pixels were selected: 15 Statistica 8.0 (developed by StatSoft). The atmospheric of which represented birch, 80—European beech, 30— correction was carried out using a Quick Atmospheric European larch, 30—Scots pine, 30—oak, 10—horn- Correction (Quack) method, radiometric calibration, beam, 15—Norway spruce, and 50—no forest. The data reduction (PCA and MNF transformations), band pixels of each class were randomly divided into train- selection, and classifications using nine different algo- ing and validation sets within each plot. There was rithms; the accuracy analysis was performed using ENVI no spatial distinction between individual plots of 5.0, and ArcGIS 10.3 was used to select training and test training and test pixels; however, in some cases, only pixels. Figures were created using the ETRS 1989 Poland training or only test pixels might have been chosen C92 Projected Coordinate System. for a single plot. Hycza et al. New Zealand Journal of Forestry Science (2018) 48:18 Page 5 of 13 Fig. 2 AISA EAGLE hyperspectral image (natural colour composite) constructed by MGGP AERO; the sample plots The spectral reflectance of more than one object (tree, directly above the section of trunks that had their loca- bare ground, or any other) could have been contained in tion mapped during field measurements. By the end of a single 1.5-m pixel. The GNSS device could also have the classification process, the entire area was classified introduced an error. Therefore, the normalised Digital since all pixels, not only the ‘clear’ ones, were used. The Surface Model (nDSM) was used to overcome this prob- classifications were verified on separate data sets and lem. All areas below 1 m were removed. The spatial evaluated at the sample plot level. resolution of the nDSM was 0.5 m, so it was possible to To perform the supervised classification, nine algo- choose training and test pixels containing a single tree rithms were used on three out of four datasets: P, BE, or at least a group of trees of a single species. The spe- SID, MD1, MD2, ML, SVM, SAM, and NN. The settings cies were identified using data points representing the for the algorithms are provided in Appendix 2. The clas- location of tree tops. We assumed that they were sification was performed using ENVI 5.0 software. Hycza et al. New Zealand Journal of Forestry Science (2018) 48:18 Page 6 of 13 Accuracy assessment fraction of correctly classified pixels with regard to all The accuracy analysis was performed using 300 test pixels pixels of that ground truth class. The user’s accuracy is and 2 MNF-band compositions (5-4-3 and 4-3-2) in which the fraction of correctly classified pixels with regard to the differences in colour among the species were most ob- all pixels classified as this class in the classified image. servable. Some of the pixels representing trees in the sam- For the classification based on all 129 spectral bands ple plots were used as test pixels, but only those that were performed with the SVM algorithm, the highest pro- most recognisable were used (European beech: 50, birch: ducer’s accuracy was observed for European larch, 20, oak: 50, hornbeam: 10, European larch: 50, no forest: no-forest, and Scots pine (92–98%) and the lowest was 50, Scots pine: 50, and Norway spruce: 20). for birch and hornbeam (10%). The highest user accur- The pixels of each class within each plot were ran- acy was observed for Scots pine and hornbeam (94– domly divided into training and validation data sets. 100%) and the lowest was for European beech (52.4%). There was no spatial distinction between the plots of in- For the classification based on 36 spectral bands per- dividual training and test pixels; however, in some cases, formed with the NN algorithm, the highest producer’s only training or only test pixels might have been chosen accuracy was observed for European larch, no-forest, for a single plot. Nevertheless, both data sets covered and Scots pine (94%) and the lowest was observed for the entire study area randomly. birch and hornbeam (0%). The highest user accuracy A normalised Digital Surface Model was used to was observed for Scots pine and hornbeam (71.2–77.4%) choose the test pixels representing only one particular and the lowest was for birch, hornbeam, and Norway species and to overcome inaccuracies caused by the spruce (0%). spatial resolution of the image and the GNSS device. For the classification based on three PCA spectral Only pixels in which a tree top was located close to the bands performed with the ML algorithm, the highest centre were selected as test pixels. The accuracy assess- producer’s accuracy was observed for European larch ment was performed using ENVI 5.0 software. and no-forest (100%) and the lowest was for birch (10%). The classification results of 98 individual sample plots The highest user’s accuracy was observed for Scots (values represented in %) were compared to the number pine and Norway spruce (100%) and the lowest was for of trees (values represented in %) belonging to individual birch (40%). species on each plot using the coefficient of determin- For the classification based on 7 MNF spectral bands ation (R ) calculated in Statistica 8.0. Only trees from performed with the ML algorithm, the highest pro- the upper canopy were taken into consideration. ducer’s accuracy was observed for beech, European larch, and no-forest (100%) and the lowest was for birch Results (10%). The highest user’s accuracy was observed for The highest accuracy was obtained by the ML algorithm hornbeam, European larch, Scots pine, and Norway and the data set of the seven MNF bands. The final map spruce (100%) and the lowest was for birch (33.33%) was subjected to Majority Filter analysis. The overall ac- (Table 3). Birch was spread across the study area with no curacy was 91.3% (Kappa—0.9) (Fig. 3.). observable concentration while hornbeam was very rare; The classification performed on all 129 bands ranged only one sample plot contained enough of the latter from 31 (BE) to 76.7% (SVM), excluding P and NN (85%) to be observable from the aerial ceiling. (below 10%). Spectral Information Divergence also per- The visual comparison of these four classification ap- formed relatively well (64.7%). There were not enough proaches on a single-plot scale is shown for two chosen training pixels to perform ML and MD2. The classifica- plots in Figs. 4 and 5. The best results were achieved tion performed on the 36 original bands ranged from using the ML algorithm. 33.6 (BE) to 66.3% (NN), excluding P (below 20%). The coefficient of determination between the number Support vector machine also performed relatively well of trees and the classification results of individual sample (58.7%). There were not enough training pixels to per- plots ranged from 0.68 (birch) to 0.99 (European larch), form ML and MD2. The classification performed on the while those of Norway spruce, hornbeam, and oak were first three PCA bands ranged from 30.3 (P) to 88.3% approximately 0.9 (Table 4). (ML). NN, SAM, and MD2 ranged from 68.3 to 72.7%. The classification performed on the first seven MNF Discussion bands ranged from 10.3 (P) to 90.7% (ML). Spectral in- Hyperspectral images are difficult to use for classifica- formation divergence and SAM also performed relatively tion purposes because they contain several narrow bands well (84.7%–85%) (Table 2). that are correlated with one another. It is important to The producer’s and user’s accuracy for the four best reduce both the amount of data and the noise before classification results (each based on a different data set) performing classifications. Clark et al. (2005) observed a is provided in Table 3. The producer’s accuracy is the general increase in accuracy of up to 30 input bands Hycza et al. New Zealand Journal of Forestry Science (2018) 48:18 Page 7 of 13 Fig. 3 Classification results (maximum likelihood on seven minimum noise fraction bands). Legend: red—birch, orange—European beech, yellow—oak, pink—hornbeam, pale blue—European larch, green—Scots pine, dark blue—Norway spruce when using a feature selection algorithm combined high-dimensional data well without the need for a with a linear discriminant analysis classifier; including large training sample size. Thus, it is not strongly af- more bands produced a lower or equal accuracy when fected by the Hughes phenomenon (Dalponte et al. classifying tree species in a tropical environment. Dal- 2009, Hughes 1968), which states that as the number ponte et al. (2009) reported a slight decrease in ac- of hyperspectral narrow bands increases, the number curacy when dropping several bands from the initial of samples (i.e. training pixels) required to maintain a 126 in a tree-species classification that combined an minimum statistical confidence and functionality in SVM classifier with a feature-selection procedure. hyperspectral data for classification also increases ex- These findings were most likely also connected to the ponentially, making it very difficult to address this classifiers applied, given that SVM is known to handle issue adequately. Hycza et al. New Zealand Journal of Forestry Science (2018) 48:18 Page 8 of 13 Table 2 Classification results Parameter Classification All 129 bands 36 chosen bands First three Principal component First seven minimum noise analysis bands fraction bands Algorithm Accuracy (%) Kappa Accuracy (%) Kappa Accuracy (%) Kappa Accuracy (%) Kappa Parallelepiped 5.3 0.005 15.3 0.05 30.3 0.23 10.3 0.06 Minimum Distance 62.7 0.56 40.3 0.3 61.7 0.55 84.7 0.82 Mahalanobis distance NETP NETP NETP NETP 72.7 0.68 81.3 0.78 Maximum likelihood NETP NETP NETP NETP 88.3 0.86 90.7 0.89 Spectral angle mapping 75 0.7 39.3 0.3 69.3 0.64 85 0.82 Spectral information divergence 64.7 0.58 38.3 0.28 37 0.27 81.3 0.78 Binary encoding 31 0.22 33.6 0.23 11.7 0.05 44 0.37 Neural networks 6.7 0.0004 66.3 0.6 68.3 0.62 63.7 0.56 Support vector machine 76.7 0.72 58.7 0.5 61 0.53 72.7 0.68 NETP not enough training pixels The compositions made from the PCA or MNF 2013; Goodwin et al. 2005)at80–100%, in a tropical and bands may be useful to distinguish tree species and sub-tropical zone (Carlson et al. 2010; Dian et al. 2014; create a layer of training pixels or polygons used to Goodwin et al. 2005; Dennison and Roberts 2003; Lucas perform the supervised classification. However, PCA is et al. 2008; Yang et al. 2009; Gong et al. 1997; van Aardt not the most suitable method to reduce multidimen- and Norris-Rogers 2008) at over 90%, and in a temperate sionality when the objective is to classify remotely zone (Zagajewski 2010; Olesiuk and Zagajewski 2008; sensed data (Cheriyadat and Bruce 2003). Principal Bartold 2008; Dian et al. 2014; Martin et al. 1998; component analysis identifies variabilities that may not Dalponte et al. 2013; Dmitriev 2014; Tarabalka 2010; perform well in multi-class discrimination and does not Richter et al. 2016)at74–93%. differentiate between within-group and between-group Our results have a high correspondence with tree variations (Hobro et al. 2010). species frequencies at the sample-plot level. Differences The classification of the first seven MNF bands using between the classification results and data from the local the ML algorithm resulted in the best overall accuracy survey may be explained by the leaves and branches of (91.3%) and kappa (0.9). The results are comparable to the trees growing near, but outside the borders of, the those obtained for a forest species in an equatorial zone testing areas. Stumps were observed in the field, so it is (Clark et al. 2005; Mickelson et al. 1998; Peerbhay et al. possible that some parts of unmapped trees were Table 3 The producer’s and user’s accuracy for each class using different datasets and algorithms Classification Support vector machine over Neural net over 36 bands Maximum likelihood over first Maximum likelihood over 129 bands three principal component first seven minimum noise analysis bands fraction bands Species Producer’s User’s Producer’s User’s Producer’s User’s Producer’s User’s accuracy (%) accuracy (%) accuracy (%) accuracy (%) accuracy (%) accuracy (%) accuracy (%) accuracy (%) Birch 10.00 66.67 0.00 0.00 10.00 40.00 10.00 33.33 European 86.00 52.44 76.00 59.38 84.00 76.36 100.00 70.42 beech Oak species 68.00 82.93 80.00 55.56 92.00 85.19 94.00 95.92 Hornbeam 10.00 100.00 0.00 0.00 70.00 63.64 60.00 100.00 European 92.00 83.64 66.00 73.33 100.00 90.91 100.00 100.00 larch No-forest 98.00 83.64 82.00 77.36 100.00 96.15 100.00 98.04 Scots pine 94.00 94.00 94.00 71.21 98.00 100.00 98.00 100.00 Norway 40.00 88.89 0.00 0.00 95.00 100.00 90.00 100.00 spruce Hycza et al. New Zealand Journal of Forestry Science (2018) 48:18 Page 9 of 13 Table 4 Coefficient of determination between the number of trees (%) and the classification results (%) of individual species on individual test plots Species R European beech 0.79 Birch 0.68 Oak 0.93 Hornbeam 0.9 European larch 0.99 Scots pine 0.81 Norway spruce 0.89 included in the sample. It is also possible that the re- flectance of the input image was disturbed by the plants growing in the lower canopy layers of the stand. Add- itionally, even the same tree species may have different values of spectral reflectance depending on their age, weather and soil conditions, moisture, vegetation period, and many other factors (Ghiyamat and Shafri 2010), Fig. 4 Comparison of the results of four classification techniques (SVM-129, NN-36, ML-PCA, ML-MNF) on a single sample plot (Adams which is the premise for using hyperspectral imagery to et al. 1995). Legend: red—birch, orange—European beech, detect disease and nutrient deficiencies in even-aged yellow—oak, pink—hornbeam, pale blue—European larch, single-species stands. green—Scots pine, dark blue—Norway spruce The set of training polygons used in this study was suitable for performing the classification on a neighbour- ing area using the same type of data (AISA Eagle hyper- spectral image), acquired at the same flight height, during the peak of the vegetation season (July and August), when the weather conditions were similar (although the atmospheric correction was performed). Otherwise, the set of training polygons used should be separate because the spectral signatures of the different tree species varied due to the study area, data type, ac- quisition date, weather conditions, altitude, and other factors (Ghiyamat and Shafri 2010). However, this is a common issue when dealing with remotely sensed data. Unfortunately, more issues can be expected with hyper- spectral data; for example, a comparison to satellite im- ages and reference data is needed. This is due the fact that flight strips are relatively narrow and a longer time is needed to cover large areas. As the result, there will be large differences between single strips or groups of strips. In these cases, a smaller part of the data set is re- quired for training, and verification can be undertaken immediately. It is also important to select the training and test pixels from the same (or at least neighbouring) areas, using the same methodology, and with a similar propor- tion of class samples to avoid differences between the accuracy assessment and the true classification results. Fig. 5 Comparison of the results of four classification techniques (SVM-129, NN-36, ML-PCA, ML-MNF) on a single sample plot (Alonzo et al. 2014). Legend: red—birch, orange—European beech, Conclusions yellow—oak, pink—hornbeam, pale blue—European larch, The classification based on 7 MNF spectral bands per- green—Scots pine, dark blue—Norway spruce formed with the ML algorithm was found to be the most Hycza et al. New Zealand Journal of Forestry Science (2018) 48:18 Page 10 of 13 Appendix 1 Table 5 Detailed information on the Miłomłyn Forest District Forest area (ha) 19,000 Total area (ha) 48,000 Forest area (FA)/total area (TA) % 39.6 Mean FA/TA in Poland % 29.1 Abundance of forest site types (%) Fresh* mixed forest 42 Fresh* deciduous forest 23 Fresh* mixed coniferous forest 22 Other forest types 13 Abundance of forest species group types (%) Coniferous trees 71 Deciduous trees 29 Most common tree species (%) Scots pine (Pinus sylvestris L.) and European larch (Larix decidua L.) 71 European beech (Fagus sylvatica L.) 12 Grey alder (Alnus glutinosa L.) 6 Birch species (Betula spp. L.) 5 Oak species (Quercus spp. L.) 4 Norway spruce (Picea abies L.) 1 Other (e.g. hornbeam Carpinus betulus L.) 1 Age classes (%) I(0–20) 7 II (Dennison and Roberts 2003; Dian et al. 2014; Dmitriev 2014; Einzmann et al. 2014; Farreira et al. 2016; Fassnacht et al. 2014; Fassnacht et 14 al. 2016; Gao and Hoetz 1990; Ghiyamat and Shafri 2010; Ghiyamat et al. 2013; Gong et al. 1997; Goodwin et al. 2005; Grant 1987; Han et al. 2004; Hainzel and Koch 2012; Harsanyi 1994; Hobro et al. 2010; Hughes 1968; Immitzer et al. 2015; Innes and Koch 1998; Jansson and Angelstam 1999) III (Jansson and Angelstam 1999; Jones et al. 2010; Kennedy and Southwood 1984; Korpela and Tokola 2006; Korpela et al. 2011; Leckie et 23 al. 2005; Lee et al. 2010; Li et al. 2014; Lucas et al. 2008; Luo and Chanussot 2009; Martin et al. 1998; Mickelson et al. 1998; Olesiuk and Zagajewski 2008; Ørka et al. 2013; Pausas et al. 1997; Peerbhay et al. 2013; Plourde et al. 2007; Portigal et al. 1997; Ribeiro da Luz and Crowley 2007; Richter et al. 2016; Roberts et al. 1997) IV (Roberts et al. 1997; Salisbury 1986; Salisbury and Milton 1998; Shang and Chisholm 2014; Schull et al. 2010; Stavrakoudis et al. 2014; 17 Tarabalka 2010; Tompalski et al. 2014; Treuhaft et al. 2002; Ullah et al. 2012; Ustin et al. 2009; Van Aardt and Wynne 2007; Van Aardt and Norris-Rogers 2008; Van Ewijk et al. 2014; Vauhkonen et al. 2014; Villa et al. 2013; Voss and Sugumaran 2008; Waser et al. 2014; Wietecha et al. 2017; Wulder et al. 2006; Yang et al. 2009) V (80–100) 17 VI (100–120) 8 VII (120–140) 7 VII + (> 140) 7 Mean volume for species (m /ha) Scots pine (Pinus sylvestris L.) 241 Norway spruce (Picea abies L.) 216 European beech (Fagus sylvatica L.) 260 Oak species (Quercus spp. L.) 310 *According to the soil moisture level, forests are divided into dry, fresh, wet, and swampy Hycza et al. New Zealand Journal of Forestry Science (2018) 48:18 Page 11 of 13 Appendix 2 Table 6 Algorithm settings (Continued) Radiometric calibration Table 6 Algorithm settings Type Reflectance Radiometric calibration Penalty parameter 100 Type Reflectance Output interleave BSQ Pyramid levels 0 Output data type Float Classification probability threshold 0 Scale factor 1 Quick atmospheric correction accurate method for classifying species (overall accuracy of 90.3%), with the highest kappa coefficient of 0.9. The Sensor type AISA results from the study reported here showed that this PCA transformation method is sufficiently reliable, accurate and user-friendly Stats X resize factor 1 to be used in practice. However, the data and software Stats Y resize factor 1 required are still expensive, which may limit its practical Calculated using Covariance matrix use by forest managers at present. Output data file Floating point Select subset from eigenvalues No Abbreviations Number of output PC bands 129 AISA: Airborne Imaging Spectrometer for Application; BE: Binary encoding; MNF transformation BSQ: Band sequential; DGPS: Differential Global Positioning System; FOV: Field of view; GNSS: Global Navigation Satellite System; LDA: Linear Shift difference subset Full scene discriminant analysis; LIDAR: Light detection and ranging; LSU: Linear spectral Select subset from eigenvalues No unmixing; MD1: Minimum distance; MD2: Mahalanobis distance; ML: Maximum likelihood; MNF: Minimum noise fraction; nDSM: Normalised Number of output MNF bands 129 Digital Surface Model; NN: Neural net; P: Parallelepiped; PCA: Principal component analysis; PPI: Pixel Purity Index; SAM: Spectral angle mapping; Parallelepiped SAR: Synthetic aperture radar; SID: Spectral information divergence; Max standard deviation from mean 3 SVM: Support vector machine Minimum distance Acknowledgements Max standard deviation from mean 3 We would like to thank Mariusz Ciesielski, Leopold Leśko, Aleksander Rybski, Marek Przywózki, Maciej Sarnowski, and Michał Brach, who conducted the Max distance error 0 local survey and established 98 sample plots in the Miłomłyn Forest District. Mahalanobis distance Funding Max standard deviation from mean 3 The study was performed in relation to the project entitled “Modelling Maximum likelihood carbon budget on the local and global scale in the State Forests Holding and developing scientific input parameters and management scenarios for Max standard deviation from mean 3 Poland” funded by the State Forests (grant number BLP-392; grant recipient: Mr. Radomir Bałazy). Data scale factor 255 Spectral angle mapping Availability of data and materials The aerial image from the AISA Eagle camera was provided by MGGP Aero. Maximum angle 0.1 The image is property of the Forest Research Institute and may be disclosed Spectral information divergence with permission from the Director of the Forest Research Institute. The analysis was performed using ENVI 5.0 and ArcGIS 10.3 provided by ESRI Maximum divergence threshold 0.05 Polska. Binary encoding Authors’ contributions Minimum encoding threshold 0 TH undertook the literature review, methodology, analysis, manuscript Neural networks writing, and field survey. KS was involved in conceiving and planning the study, methodology, analysis, manuscript writing, and reviewing. RB Activation Logistic compiled the data and literature review, and reviewed the manuscript. All authors read and approved the final manuscript. Training threshold contribution 0.9 Training rate 0.2 Authors’ information TH—Master of Science in Remote Sensing and Geoinformatics, Assistant in Training momentum 0.9 the Forest Research Institute, and doctoral student. Training RMS exit criteria 0.1 KS—Ph.D.in Forestry, Adjunct in the Forest Research Institute, and coordinator on two projects (“A complex forest dynamics monitoring of the Number of hidden layers (and nodes) 1 (129, 36, 3, 7). Białowieża Forest based on remote sensing data” and “An estimation of Number of training iterations 1000 biomass and carbon resources in forests based on remote sensing data”) with 75 publications and 223 citations (Research Gate, February 2018). Minimum output activation threshold 0 RB—Master of Science in Forestry, Assistant in the Forest Research Institute, Support vector machine and coordinator of two projects (“A forest information system of monitoring and forest condition assessment of Sudety and West Beskidy” and “Carbon Hycza et al. New Zealand Journal of Forestry Science (2018) 48:18 Page 12 of 13 budget modelling of the Polish State Forests Holding on the local and Dalponte, M., Bruzzone, L., & Gianelle, D. (2012). Tree species classification in the global scale and the development of input parameters and economic southern Alps based on the fusion of very high geometrical resolution scenarios for Poland”) with 53 publications and 120 citations (Research Gate, multispectral/hyperspectral images and LIDAR data. Remote Sensing of February 2018). Environment, 123, 258–270. Dalponte, M., Bruzzone, L., Vescovo, L., & Gianelle, D. (2009). The role of spectral resolution and classifier complexity in the analysis of hyperspectral images of Ethics approval and consent to participate forest areas. Remote Sensing of Environment, 133(11), 2345–2355. Not applicable. Dalponte, M., Ørka, H. O., Ene, L. T., Gobakken, T., & Næsset, E. (2013). Tree crown delineation and tree species classification in boreal forests using Consent for publication hyperspectral and ALS data. Remote Sensing of Environment, 140, 306–317. Not applicable. Dennison, P. E., & Roberts, D. A. (2003). The effects of vegetation phenology on endmember selection and species mapping in southern California chaparral. Competing interests Remote Sensing of Environment, 87, 295–309. The authors declare that they have no competing interests. Dian, Y., Li, Z., & Pang, Y. (2014). Spectral and texture features combined for forest tree species classification with airborne hyperspectral imagery. Indian Society of Remote Sensing, 43(1), 101–107. Publisher’sNote Dmitriev, E. (2014). Classification of the forest cover of Tver oblast using Springer Nature remains neutral with regard to jurisdictional claims in hyperspectral airborne images. Izvestiya - Atmospheric and Oceanic Physics, published maps and institutional affiliations. 50(9), 929–942. Einzmann, K., Immitzer, M., Bachmann, M., Pinnel, N., & Atzberger, C. (2014). Received: 23 February 2018 Accepted: 27 November 2018 Method analysis for collecting and processing in-situ hyperspectral needle reflectance data for monitoring Norway spruce. Photogrammetrie, Fernerkundung, Geoinformation, 5, 423–434. References Farreira, M. P., Zanotta, D. C., Zortea, M., & andde Souza Filho, C. R. (2016). Mapping Adams, J. B., Sabol, D. E., Kapos, V., Filho, R. A., Roberts, D. A., Smith, M. O., & tree species in tropical seasonal semi-deciduous forests with hyperspectral and Gillespie, A. R. (1995). Classification of multispectral images based on multispectral data. Remote Sensing of Environment, 179,66–78. fractions of endmembers: Application to land−cover change in the Brazilian Fassnacht, F. E., Latifi, H., Stereńczak, K., Modzelewska, A., Lefsky, M., Waser, L. T., Amazon. Remote Sensing of Environment, 52(2), 137–154. Straub, C., & Ghosh, A. (2016). Review of studies on tree species classification Alonzo, M., Bookhagen, B., & Roberts, D. A. (2014). Urban tree species mapping from remotely sensed data. Remote Sensing of Environment, 186,64–87. using hyperspectral and LIDAR data fusion. Remote Sensing of Environment, Fassnacht, F. E., Neumann, C., Förster, M., Buddenbaum, H., Ghosh, A., Clasen, A., 148,70–83. Joshi, P. K., & Koch, B. (2014). Comparison of feature reduction algorithms for Asner, G. P. (1998). Biophysical and biochemical sources of variability in canopy classifying tree species with hyperspectral data on three central European reflectance. Remote Sensing of Environment, 64(3), 234–253. test sites. IEEE Journal of Selected Topics in Applied Earth Observation and Bartold, M. (2008). Klasyfikacja drzewostanów na obrazie satelitarnym Hyperion Remote Sensing, 7(6), 2547–2561. (EO-1). Teledetekcja Środowiska, 39,5–29. Gao, B. C., & Hoetz, A. F. H. (1990). Column atmospheric water-vapor and Bellanti, L., Blesius, L., Hinnes, E., & Kruse, B. (2016). Tree species classification vegetation liquid water retrievals from airborne imaging spectrometer data. using hyperspectral imagery: A comparison of two classifiers. Remote Sensing, Journal of Geophysical Research-Atmospheres, 95, 3549–3564. 8(6), 445–463. Ghiyamat, A., & Shafri, H. Z. M. (2010). A review on hyperspectral remote sensing Bernstein, L. S., Sundberg, R. L., Levine, R. Y., Perkins, T. C., & Berk, A. (2005). A new for homogeneous and heterogeneous forest biodiversity assessment. method for atmospheric correction and aerosol optical property retrieval for International Journal of Remote Sensing, 31, 1837–1856. VIS−SWIR multi− and hyperspectral imaging sensors: QUAC (quick Ghiyamat, A., Shafri, H. Z. M., Mahdiraji, G. A., Shariff, A. R. M., & Mansor, S. (2013). atmospheric correction). Proceedings of the IEEE International Geoscience and Hyperspectral discrimination of tree species with different classifications Remote Sensing Symposium, 5, 3549–3552. using single- and multiple-endmember. International Journal of Applied Earth Caiyun, Z., & Fang, Q. (2012). Mapping individual tree species in an urban forest Observation and Geoinformation, 23, 177–191. using airborne LIDAR data and hyperspectral imagery. Photogrammetric Gong, P., Pu, R., & Zu, B. (1997). Conifer species recognition: An exploratory analysis Engineering and Remote Sensing, 78(10), 1079–1087. of in situ hyperspectral data. Remote Sensing of Environment, 62(2), 189–200. Carlson, K. M., Asner, G. P., Hughes, R. F., Ostertag, R., & Martin, R. E. (2010). Goodwin, N., Turner, R., & Merton, R. (2005). Classifying eucalyptus forests with high Hyperspectral remote sensing of canopy biodiversity in Hawaiian lowland spatial and spectral resolution imagery. An investigation of individual species rainforests. Ecosystems, 10, 526–549. and vegetation communities. Australian Journal of Botany, 53,337–345. Chambers, D., Périé, C., Casajus, N., & de Blois, S. (2013). Challenges in modelling Grant, L. (1987). Diffuse and specular characteristics of leaf reflectance. Remote the abundance of 105 tree species in eastern North America using climate, Sensing of Environment, 22, 309–322. edaphic, and topographic variables. Forest Ecology and Management, 291, Hainzel, J., & Koch, B. (2012). Investigating multiple data sources for tree species 20–29. classification in temperate forest and use for single tree delineation. International Cheriyadat, A., & Bruce, L. M. (2003). Why principal component analysis is not an Journal of Applied Earth Observation and Geoinformation, 18,101–110. appropriate feature extraction method for hyperspectral data. Proceedings of Han, T., Goodenough, D. G., Dyk, A., & Chen, H. (2004). Hyperspectral feature the IEEE International. Geoscience Remote Sensing Symposium, 6, 3420–3422. selection for forest classification IEEE International Geoscience and Remote Cho, M. A., Debba, P., Mathieu, R., Naidoo, L., van Aardt, J., & Asner, G. P. (2010). Sensing Symposium, 2) (pp. 1471–1474). Anchorage. Improving discrimination of savanna tree species through a multiple- Harsanyi, J. C. (1994). Hyperspectral image classification and dimensionality endmember spectral angle mapper approach: Canopy-level analysis. IEEE reduction: An orthogonal subspace projection approach. Geoscience and Transactions on Geoscience and Remote Sensing, 48(11), 4133–4142. Remote Sensing, 32(4), 779–785. Clark, M., Roberts, D., & Clark, D. (2005). Hyperspectral discrimination of tropical Hobro, A. J., Kuligowski, J., Döll, M., & Lendl, B. (2010). Differentiation of walnut rain forest tree species at leaf to crown scales. Remote Sensing of wood species and steam treatment using ATR-FTIR and partial least squares Environment, 96(3–4), 375–398. discriminant analysis (PLS-DA). Analytical and Bioanalytical Chemistry, 398(6), Clark, M. L., & Roberts, D. A. (2012). Species-level differences in hyperspectral 2713–2722. metrics among tropical rainforest trees as determined by a tree-based Hughes, G. F. (1968). On the mean accuracy of statistical pattern recognizers. classifier. Remote Sensing, 4(6), 1820–1855. Transactions on Information Theory, 14,55–63. Dalponte, M., Bruzzone, L., & Gianelle, D. (2008). Fusion of hyperspectral and LIDAR remote sensing data for classification of complex forest areas. IEEE Immitzer, M., Atzberger, C., Einzmann, K., Böck, S., Mattiuzzi, M., Wallner, A., Seitz, R., Transactions on Geoscience and Remote Sensing, 46, 1416–1427. Pinnel, N., Müller, A., & Frost, M. (2015). Fichten- und Kiefernkarte für Bayern. LWF Dalponte, M., Bruzzone, L., & Gianelle, D. (2011). Tree species classification in the aktuell. Bayerische Landesanstalt für Wald und Forstwirtschaft, 106,30–37. southern Alps with very high geometrical resolution multispectral and Innes, J. L., & Koch, B. (1998). Forest biodiversity and its assessment by remote hyperspectral data. Remote Sensing of Environment, 123, 258–270. sensing. Global Ecology and Biogeography Letters, 7, 397–419. Hycza et al. New Zealand Journal of Forestry Science (2018) 48:18 Page 13 of 13 Jansson, G., & Angelstam, P. (1999). Threshold levels of habitat composition for Salisbury, J. W., & Milton, N. M. (1998). Thermal infrared (2.5- to 13.5-μm) the presence of long-tailed tit (Aegithalos audatus) in a boreal landscape. directional hemispherical reflectance of leaves. Photogrammetric Engineering Landscape Ecology, 14, 283–290. and Remote Sensing, 54(9), 1301–1304. Jones, T. G., Coops, N. C., & Sharma, T. (2010). Assessing the utility of airborne Schull, M. A., Knyazikhin, Y., Xu, L., Samanta, A., Ganguly, S., & Latorre Carmona, P. hyperspectral and LIDAR data for species distribution mapping in the coastal (2010). Canopy spectral invariants. Part 2: Application to classification of Pacific Northwest, Canada. Remote Sensing of Environment, 114(12), 2841–2852. forest types from hyperspectral data. Journal Quantitative Spectroscopy Radiative Transfer, 112, 736–750. Kennedy, C. E. J., & Southwood, T. R. E. (1984). The number of species of insects Shang, X., & Chisholm, L. A. (2014). Classification of Australian native Forest associated with British trees: A re-analysis. Journal of Animal Ecology, 53(2), 455–478. species using hyperspectral remote sensing and machine-learning Korpela, I., Heikkinen, V., Hokavaara, E., Rohrbach, F., & Tokola, T. (2011). Variation classification algorithms. IEEE Journal of Selected Topics in Applied Earth and directional anisotropy of reflectance at the crown scale – Implications Observation and Remote Sensing, 7(6), 2481–2489. for tree species classification in digital aerial images. Remote Sensing of Stavrakoudis, D. G., Dragozi, E., Gitas, I. Z., & Karydas, C. G. (2014). Decision fusion Environment, 115, 2062–2074. based on hyperspectral and multispectral satellite imagery for accurate forest Korpela, I. S., & Tokola, T. E. (2006). Potential of aerial image-based monoscopic species mapping. Remote Sensing, 6(8), 6897–6928. and multiview single-tree forest inventory: A simulation approach. Forest Tarabalka, Y. (2010). Segmentation and classification of hyperspectral images Science, 52(2), 136–147. using minimum spanning forest grown automatically selected markers. IEEE Leckie, D. G., Tinis, S., Nelson, T., Burnett, C., Gougeon, F. A., Cloney, E., & Paradine, Transactions on Systems, Man, and Cybernetics, 40(5), 1267–1279. D. (2005). Issues in species classification of trees in old growth conifer stands. Tompalski, P., Coops, N. C., White, J. C., & Wulder, M. A. (2014). Simulating the Canadian Journal of Remote Sensing, 31(2), 175–190. impacts of error in species and height upon tree volume derived from Lee, W., S., Alchanatis, V., Yang, C., Hirafuji, M., Moshou, D., Li, C. (2010). Sensing airborne laser scanning data. Forest Ecology and Management, 327, 167–177. technologies for precision specialty crop production. Computers and Treuhaft, R. N., Asner, G. P., Law, B. E., & Tuyl Van, S. (2002). Forest leaf area Electronics in Agriculture, 74, 2–33. density profiles from the quantitative fusion of radar and hyperspectral data. Li, J., Xin, H., Gamba, P., & Bioucas-Dias, J. M. (2014). Multiple feature learning for Journal of Geophysical Research, 107(D21), ACL 7–1 – ACL 7-13. hyperspectral image classification. Transactions on Geoscience and Remote Ullah, S., Schlerf, M., Skidmore, A. K., & Hecker, C. (2012). Identifying plant species Sensing, 53(3), 1592–1606. using mid wave infrared (2.5–6 μm) and thermal infrared (8–14 μm) Lucas, R. M., Lee, A. C., & Bunting, P. J. (2008). Retrieving forest biomass through emissivity spectra. Remote Sensing of Environment, 118,95–102. integration of CASI and LIDAR data. International Journal of Remote Sensing, 29, Ustin, S., Gitelson, A. A., Jacquemoud, S. M., Asner, G. P., Gamon, J., & Zarco- 1553–1577. Tejada, P. J. (2009). Retrieval of foliar information about plant pigment Luo, B., & Chanussot, J. (2009). Hyperspectral image classification based on spectral systems from high resolution spectroscopy. Remote Sensing of Environment, and geometrical features (IEEE International Workshop on Machine Learning 113,67–77. for Signal Processing) (pp. 1–6). Van Aardt, J., & Norris-Rogers, M. (2008). Spectral-age interactions in managed, Martin, M. E., Newman, S. D., Aber, J. D., & Congalton, R. G. (1998). Determining even-aged eucalyptus plantations: Application of discriminant analysis and forest species composition using high spectral resolution remote sensing classification and regression trees approaches to hyperspectral data. data. Remote Sensing of Environment, 65, 249–254. International Journal of Remote Sensing, 29(6), 1841–1845. Mickelson, J. G., Civco, D. L., & Silander, J. A. (1998). Delineating forest canopy Van Aardt, J., & Wynne, R. H. (2007). Examining pine spectral separability using species in the north-eastern United States using multi-temporal TM imagery. hyperspectral data from an airborne sensor: An extension of field-based Photogrammetric Engineering and Remote Sensing, 64(9), 891–904. results. International Journal of Remote Sensing, 28(2), 431–436. Olesiuk, D., & Zagajewski, B. (2008). Wykorzystanie obrazów hiperspektralnych Van Ewijk, K. Y., Randin, C. F., Treitz, P. M., & Scott, N. A. (2014). Predicting fine-scale do klasyfikacji pokrycia terenu zlewni Bystrzanki. Teledetekcja Środowiska, tree species abundance patterns using biotic variables derived from LIDAR and 40, 125–148. high spatial resolution imagery. Remote Sensing of Environment, 150,120–131. Ørka, H. O., Dalponte, M., Gobakken, T., Næsset, E., & Ene, L. T. (2013). Vauhkonen, J., Ørka, H. O., Holmgren, J., Dalponte, M., Hainzel, J., & Koch, B. Characterizing forest species composition using multiple remote sensing (2014). Tree species recognition based on airborne laser scanning and data sources and inventory approaches. Scandinavian Journal of Forest complementary data sources. Forestry applications of airborne laser Research, 28(7), 677–688. scanning. In M. Maltamo, E. Naesset, & J. Vauhkonen (Eds.), Forestry Pausas, J. G., Austin, M. P., & Noble, I. R. (1997). A forest simulation model for Applications of Airborne Laser Scanning (pp. 135–156). Dordrecht: Springer. predicting eucalypt dynamics and habitat quality for arboreal marsupials. Villa, A., Chanossut, J., Benediktsson, J. A., Jutten, C., & Dambreville, R. (2013). Ecological Applications, 7(3), 921–993. Unsupervised methods for the classification of hyperspectral images with Peerbhay, K. Y., Mutanga, O., & Ismail, R. (2013). Commercial tree species low spatial resolution. Pattern Recognition, 46(6), 1556–1568. discrimination using airborne AISA Eagle hyperspectral imagery and partial Voss, M., & Sugumaran, R. (2008). Seasonal effect on tree species classification in least squared discriminant analysis (PLS-DA) in KwaZulu-Natal - South Africa. an urban environment using hyperspectral data. LIDAR, and an object- Remote Sensing, 79,19–28. oriented approach. Sensors, 8, 3020–3036. Plourde, L. C., Ollinger, S. V., Smith, M. L., & Martin, M. E. (2007). Estimating species Waser, L. T., Küchler, M., Jütte, K., & Stampfer, T. (2014). Evaluating the potential of abundance in a northern temperate forest using spectral mixture analysis. world-View-2 data to classify tree species and different levels of ash Photogrammetric Engineering and Remote Sensing, 73(7), 829–840. mortality. Remote Sensing, 6(5), 4515–4545. Portigal, F., Holasek, R., Mooradian, G., Owensby, P., Dicksion, M., Fene, M., Elliot, Wietecha, M., Modzelewska, A., & Stereńczak, K. (2017). Airborne hyperspectral M., Hall, E., and Driggett, D. (1997). Vegetation classification using red-edge data for the classification of tree species a temperate forests (Wykorzystanie first derivative and green peak statistical moment indices with the Advanced lotniczej teledetekcji hiperspektralnej w klasyfikacji gatunkowej lasów strefy Airborne Hyperspectral Imaging System (AAHIS). Third International Airborne umiarkowanej). Sylwan, 161(1), 3–17. Remote Sensing Conference and Exhibition. Copenhagen, Denmark, 7–10 Wulder, M. A., Dymond, C. C., White, J. C., Leckie, D. G., & Caroll, A. L. (2006). July 1997. II (Ann Arbor, MI: ERIM): 789–797. Surveying mountain pine beetle damage of forests: A review of remote Ribeiro da Luz, B., & Crowley, J. K. (2007). Spectral reflectance and emissivity sensing opportunities. Forest Ecology and Management, 221,27–41. features of broad leaf plants: Prospects for remote sensing in the thermal Yang, C., Everitt, J. H., Fletcher, R. S., Jensen, R. R., & Mausel, P. W. (2009). infrared (8.0–14.0 μm). Remote Sensing of Environment, 109, 393–405. Evaluating AISA+ hyperspectral imagery for mapping black mangrove along Richter, R., Reu, B., Wirth, C., Doktor, D., & Vohland, M. (2016). The use of airborne the South Texas. Gulf Coast. Photogrammetric Engineering and Remote hyperspectral data for tree species classification in a species-rich central Sensing, 75, 425–435. European forest area. International Journal of Applied Earth Observation and Zagajewski, B. (2010). Ocena przydatności sieci neuronowych i danych Geoinformation, 52, 464–474. hiperspektralnych do klasyfikacji roślinności Tatr Wysokich. Teledetekcja Roberts, D. A., Green, R. O., & Adams, J. B. (1997). Temporal and spatial patterns in Środowiska, 43, 1-113. vegetation and atmospheric properties from AVIRIS. Remote Sensing of Zarco-Tejada, P. J., Pushnik, J. C., Dobrowski, S., & Ustin, S. L. (2003). Steady-state Environment, 62, 223–240. chlorophyll a fluorescence detection from canopy derivative reflectance and Salisbury, J. W. (1986). Preliminary measurements of leaf spectral reflectance in the 8- double-peak red-edge effects. Remote Sensing of Environment, 84, 283–294. 14 μm region. International Journal of Remote Sensing, 7(12), 1879–1886. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png New Zealand Journal of Forestry Science Springer Journals

Potential use of hyperspectral data to classify forest tree species

Loading next page...
 
/lp/springer-journals/potential-use-of-hyperspectral-data-to-classify-forest-tree-species-tqiIedp6qP
Publisher
Springer Journals
Copyright
Copyright © 2018 by The Author(s).
Subject
Life Sciences; Forestry
eISSN
1179-5395
DOI
10.1186/s40490-018-0123-9
Publisher site
See Article on Publisher Site

Abstract

Background: Remote sensing techniques and data are becoming increasingly popular in forest management, e.g. for change detection and health condition analysis. Tree species recognition is a fundamental issue in taking forest inventories, especially in carbon budget modelling. Hyperspectral imagery provides an accurate classification results for large areas based on a relatively small amount of training data. Results: A hyperspectral image of a forest stand in north-eastern Poland taken using an AISA (Airborne Imaging Spectrometer for Application) Eagle camera was transformed to extract the most valuable spectral differences and was classified into seven tree types (birch, European beech, oak, hornbeam, European larch, Scots pine, and Norway spruce) using nine classification algorithms. The highest overall accuracy and kappa coefficient were 90.3% and 0.9 respectively using three minimum noise fraction bands and maximum likelihood classifier. Conclusions: Hyperspectral imaging of forests can be used to classify major forest tree species with a good degree of accuracy. It is time-efficient and user-friendly; however, the data and software required means that this approach is still expensive at present. Keywords: AISA, Hyperspectral classification, Minimum noise fraction, Trees Background Einzmann et al. 2014). The biotic characteristics of the Remote sensing is the science of acquiring information whole stands include leaf and branch density, angular about objects or areas from a distance, typically from distribution, clumping, tree size compared to neighbours aircraft or satellites. (Leckie et al. 2005, Korpela et al. 2011), and lichens, In tree canopies, the amount of radiation reflected in mosses, herbaceous vegetation, lianas, or other epiphytes regions of different wavelengths is related to the chem- (Clark and Roberts 2012). The abiotic characteristics of ical and physical properties of single trees as well as bi- whole stands include topography, soil type (and its influ- otic and abiotic characteristics of an entire stand. ence), moisture, and microclimate (Portigal et al. 1997). Among the chemical properties of single trees are the It is hardly feasible to identify species-specific absorp- levels of lignin, cellulose, nitrogen, chlorophyll, caroten- tion features in the visible and near-infrared (VIS-NIR) oids, anthocyanins (Asner 1998; Clark et al. 2005; Grant spectral region; however, this is much easier in the 1987; Clark and Roberts 2012; Ustin et al. 2009), and shortwave infrared (SWIR) spectral region (Asner 1998). water (Asner 1998, Gao and Hoetz 1990, Zarco-Tejada Salisbury (1986) presented leaf-level thermal infrared et al. 2003, Lee et al. 2010); these affect the health status (TIR) spectra of four species and identified well-defined of the trees (Waser et al. 2014). Among the physical and notably different spectral features. Salisbury and properties of single trees are leaf and wood morphology, Milton (1998) obtained close-range thermal reflectance transmission characteristics (Asner 1998; Clark et al. measurements for several other species and reported dif- 2005; Grant 1987; Clark and Roberts 2012; Ustin et al. ferences in the spectra in most of them. Ribeiro da Luz 2009), vertical leaf area density (Treuhaft et al. 2002), and Crowley (2007) found that TIR spectra were associ- and age (Ghiyamat et al. 2013; Roberts et al. 1997; ated with several chemical and structural compounds of plants such as cellulose, silica, xylan, and oleanolic acid * Correspondence: t.hycza@ibles.waw.pl levels, and reported that TIR signals were much more Department of Geomatics, Forest Research Institute, Braci Leśnej 3, 05-090 species-specific than the reflectance signals observed in Raszyn, Sękocin Stary, Poland © The Author(s). 2018 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. Hycza et al. New Zealand Journal of Forestry Science (2018) 48:18 Page 2 of 13 the visible, shortwave, and infrared regions. Many plants provided later in this paper. This method (MNF) was used develop chemical and aromatic compounds that might by Zagajewski (2010) to classify vegetation in the Tatra help define species-specific middle infrared and TIR sig- Mountains, by Olesiuk and Zagajewski (2008) to classify natures (Ribeiro da Luz and Crowley 2007; Ullah et al. the land cover of the Bystrzanka river drainage basin, and 2012). by Bartold (2008), Dian et al. (2014), and Han et al. (2004) Identifying tree species using remote sensing data is use- to classify forest tree species. Han et al. (2004) compared ful in the context of detecting changes (Adams et al. 1995) the results with those obtained by canonical transform- and managing water stress (Cho et al. 2010; Fassnacht et al. ation, while Harsanyi (1994)used the ‘orthogonal sub- 2016). It helps in the development of sustainable manage- space projection’ method. This method eliminates the ment policies (Dalponte et al. 2012, Jones et al. 2010, response from non-targets while applying a filter to match Plourde et al. 2007, Heinzel and Koch 2012, Kennedy and desired targets in the data, and is most efficient and effect- Southwood 1984) and performance of forest resource (Van ive when the target signatures are distinct. Aardt and Wynne 2007) and single tree inventories (Kor- Algorithms such as the Pixel Purity Index (PPI) pela and Tokola 2006; Immitzer et al. 2015;Tompalski et (Zagajewski 2010, Olesiuk and Zagajewski 2008, Bartold al. 2014). It enables the assessment and monitoring of bio- 2008) or linear spectral unmixing (LSU), which produces diversity, species compositions (Shang and Chisholm 2014; ‘maps of abundance’ in which each pixel is assigned to Wulder et al. 2006), wildlife habitats (Jansson and Angel- more than one class with a specified probability level (Luo stam 1999;Pausas etal. 1997), invasive species migrations and Chanussot 2009; Villa et al. 2013;Li et al. 2014), can (Chambers et al. 2013; Van Ewijk et al. 2014), and in the be used to extract the pixels most useful for the classifica- understanding of tree ecology (Chambers et al. 2013,Van tion (endmembers). Schull et al. (2010)alsoused pure Ewijk et al. 2014). It can also be applied to the estimation of spectral pixels to classify forests in the north-eastern USA insect abundance in forests (Kennedy and Southwood and achieved an overall accuracy of 92%. 1984) and the development of species-specific growth and The ability to successfully classify forest tree species yield models as well as allometric equations (Ørka et al. using hyperspectral data was proven for forests in the 2013; Vauhkonen et al. 2014). equatorial zone (Clark et al. 2005; Mickelson et al. 1998; Proper forest management and planning based on accur- Peerbhay et al. 2013; Goodwin et al. 2005), when seven ate distinction of tree species requires highly accurate clas- tree species were classified using linear discriminant ana- sification maps that cannot yet be produced using the lysis (LDA), maximum likelihood (ML), and spectral angle multispectral images typically acquired in four to eight wide mapping (SAM) methods, with accuracies of 80 to 100%. spectral bands. Hyperspectral data are more useful for clas- The hyperspectral data were also used in the tropical and sifying tree species: the only condition is that the species sub-tropical zones (Dalponte et al. 2008,Dian et al. 2014, must appear significantly different in the spectral reflect- Dennison and Roberts 2003, Lucas et al. 2008,Yang et al. ance measured in dozens of narrow spectral intervals (Clark 2009,Gong et al. 1997, Van Aarst and Norris-Rogers et al. 2005, Heinzel and Koch 2012, Carlson et al. 2010, 2008, Bellanti et al. 2016) with accuracies of over 90% and Dalponte et al. 2010, Dalponte et al. 2011,Stavrakoudiset in the temperate zone (Zagajewski 2010; Olesiuk and al. 2014,Farreiraetal. 2016). The reflectance of individual Zagajewski 2008; Bartold 2008;Dianet al. 2014;Martinet tree species is dependent on numerous factors, and the dif- al. 1998; Dalponte et al. 2013; Dmitriev 2014; Tarabalka ferences are sometimes too subtle to be observed using 2010; Richter et al. 2016)with accuracies of74 to 93%. wide, multispectral bands (Dalponte et al. 2009;Mickelson Classification results may be improved using hyper- et al. 1998). Since the technology was released, the cost of spectral data with light detection and ranging (LIDAR) hyperspectral images has decreased gradually. It is expected data (Alonzo et al. 2014). For the temperate and that it will be soon possible to use hyperspectral imagery to sub-tropical (Hainzel and Koch 2012; Dalponte et al. study forest ecology and develop management and planning 2008; Caiyun and Fang 2012) zones, the accuracies were techniques (Innes and Koch 1998;Dalponteetal. 2008; over 80%. Passive optical systems, particularly hyper- Voss and Sugumaran 2008). spectral ones, generally showed higher potential for tree However, hyperspectral images contain a huge amount species classification than active synthetic aperture radar of auto-correlated data. Principal component analysis (SAR) or LIDAR sensor systems. However, LIDAR data (PCA) is often used to solve this problem (Zagajewski have proven suitable for regions with a low number of 2010; Olesiuk and Zagajewski 2008; Bartold 2008). This species (Fassnacht et al. 2016). Forest stands classified widely known technique creates a set of artificial bands with the highest accuracy in the European temperate in which each band is less informative than the previous zone include mostly homogenous ones, dominated by one. The minimum noise fraction (MNF) transformation Scots pine (Pinus sylvestris L.) and Norway spruce (Picea works in a similar manner but reduces the noise first. abies L.). Of the broadleaved species, European beech More detailed information on these transformations is (Fagus sylvatica L.) and oak (Quercus spp. L.) are Hycza et al. New Zealand Journal of Forestry Science (2018) 48:18 Page 3 of 13 classified with the highest accuracy, but these classifica- information about this area was obtained from www.mi- tions have lower accuracies than those of coniferous spe- lomlyn.olsztyn.lasy.gov.pl/zasoby-lesne. The size of the cies (Wietecha et al. 2017). area and relative tree composition is given in Table 1 and The aim of this study was to evaluate the accuracy detailed information is listed in Appendix 1. The individ- of tree species classification methods using a hyper- ual compositions of Scots pine (Pinus sylvestris L.) and spectral Airborne Imaging Spectrometer for Applica- European larch (Larix decidua L.) were not provided. The tion (AISA) Eagle image for a forested area in study area was a 15 km (10 km long and 1.5 km wide) northern Poland. The following algorithms were eval- rectangle including three lakes: SzelągWielki, Tabórz uated in the study: PCA and MNF transformation (to (southern part), and Długie (northern part) (Fig. 2.). reduce noise and auto-correlated data), parallelepiped A local survey was performed on 9.85 ha of the (P), minimum distance (MD1), Mahalanobis distance Miłomłyn Forest District using a series of circular test (MD2), ML, SAM, spectral information divergence plots (radius: 12.62 m; area: 500.34 m ) in March 2014. (SID), neural net (NN), and support vector machine The sample plots were surveyed individually to achieve (SVM) to perform the supervised classification). The the highest level of diversity for various forest character- results were evaluated using a set of 300 test pixels, istics (e.g. age, species, forest type), where the influence deployed randomly across the study/sample plots area, of slope was minimal (Fig. 2). We corrected for the in- to achieve the most reliable assessment of accuracy. fluence of slope on the tree-position measurements. Each tree with a diameter at breast height (dbh) over 5 Materials and methods cm was inventoried and had the following information Study area recorded: distance from centre of the plot, azimuth The survey was performed in the Miłomłyn Forest District (measured from the centre of the plot to each tree), de- in the north-eastern Poland (Fig. 1). Background foliation (assessed using an expert method), and height. Fig. 1 The study area Hycza et al. New Zealand Journal of Forestry Science (2018) 48:18 Page 4 of 13 Table 1 Basic information about the Miłomłyn Forest District Pre-processing (source: www.milomlyn.olsztyn.lasy.gov.pl/zasoby-lesne) The image was geometrically corrected by MGGP AERO Parameter Amount (UTM, Zone 34N, WGS-84). It was also subject to radiometric calibration (using the ENVI ‘Radiometric Forest area (ha) 19,000 calibration’ tool–calibration type: reflectance, output Total area (ha) 48,000 interleave: BSQ (band sequential), output data type: Most common tree species (%) float, scale factor: 1.00) and ‘Quick Atmospheric Correc- Scots pine, European larch 71 tion–Quack’ atmospheric correction (Sensor Type: European beech (Fagus sylvatica L.) 12 AISA) (Dalponte et al. 2012; Bernstein et al. 2005)(Ap- Grey alder (Alnus glutinosa L.) 6 pendix 2).These procedures were performed using ENVI 5.0. Birch (Betula spp. L.) 5 Oak (Quercus spp. L.) 4 Norway spruce (Picea abies L.) 1 Data reduction After the atmospheric correction, the amount of data Other (e.g. hornbeam Carpinus betulus L.) 1 was reduced. The image containing 129 bands was not an ideal data set with which to perform supervised clas- The centre of the test plot was determined using the sification, because it contained too much auto-correlated Pathfinder ProXT (Trimble, Sunnyvale, California), Glo- data. The reduction of the data may be performed using bal Navigation Satellite System (GNSS) which functions one of two types of methods: data transformation (PCA in the DGPS (Differential Global Positioning System) or MNF transformation) (Clark et al. 2005) or band se- mode. Its vertical and horizontal accuracy was estimated lection. Data transformation is fully automatic but is to be 1.4 m and 0.97 m respectively. Tree heights were based on artificial bands. Band selection is based on ori- measured using a Vertex IV device (Haglof Sweden AB, ginal bands but is also very subjective. Both methods Langsele, Sweden) and dbh was measured using a Codi- were tested. The data reduction was performed using mex calliper (Codimex, Warsaw, Poland). The data col- ENVI 5.0 software. lected were used to calibrate and verify the hyperspectral image classification process. No grey alder trees were found in the plots so this species was not considered fur- Classification ther. Although hornbeam occurs only occasionally in the Finally, four sets of data were classified (using four forest, it was found in one plot so was included in the algorithms): analysis. – The result of the PCA transformation—first three bands Data and software – The result of the MNF transformation—first seven The hyperspectral image was provided by MGGP AERO bands and taken by the AISA Eagle camera (SPECIM) on 3 – All 129 bands August 2013 at an altitude of 2303–2328 m (single – 36 original bands with the largest differences in the flight). The spectral resolution of the image was 400– spectral profiles generated from training pixels for 970 nm (129 spectral bands, 4–5 nm wide); the radio- each tree species metric resolution was 12 bits, while the spatial reso- lution was 1.5 m. The lens size was 18.5 mm and the To perform the supervised classification, it was im- field of view (FOV) was 37.7°. portant to choose representative pixels with which to The hyperspectral image classification process (as de- train the algorithm. This was performed using two tailed below) was performed using ENVI 5.0 (developed MNF band compositions and the data from the test by Exelis Inc.), ArcGIS 10.3 (developed by ESRI), and plots. A total of 260 training pixels were selected: 15 Statistica 8.0 (developed by StatSoft). The atmospheric of which represented birch, 80—European beech, 30— correction was carried out using a Quick Atmospheric European larch, 30—Scots pine, 30—oak, 10—horn- Correction (Quack) method, radiometric calibration, beam, 15—Norway spruce, and 50—no forest. The data reduction (PCA and MNF transformations), band pixels of each class were randomly divided into train- selection, and classifications using nine different algo- ing and validation sets within each plot. There was rithms; the accuracy analysis was performed using ENVI no spatial distinction between individual plots of 5.0, and ArcGIS 10.3 was used to select training and test training and test pixels; however, in some cases, only pixels. Figures were created using the ETRS 1989 Poland training or only test pixels might have been chosen C92 Projected Coordinate System. for a single plot. Hycza et al. New Zealand Journal of Forestry Science (2018) 48:18 Page 5 of 13 Fig. 2 AISA EAGLE hyperspectral image (natural colour composite) constructed by MGGP AERO; the sample plots The spectral reflectance of more than one object (tree, directly above the section of trunks that had their loca- bare ground, or any other) could have been contained in tion mapped during field measurements. By the end of a single 1.5-m pixel. The GNSS device could also have the classification process, the entire area was classified introduced an error. Therefore, the normalised Digital since all pixels, not only the ‘clear’ ones, were used. The Surface Model (nDSM) was used to overcome this prob- classifications were verified on separate data sets and lem. All areas below 1 m were removed. The spatial evaluated at the sample plot level. resolution of the nDSM was 0.5 m, so it was possible to To perform the supervised classification, nine algo- choose training and test pixels containing a single tree rithms were used on three out of four datasets: P, BE, or at least a group of trees of a single species. The spe- SID, MD1, MD2, ML, SVM, SAM, and NN. The settings cies were identified using data points representing the for the algorithms are provided in Appendix 2. The clas- location of tree tops. We assumed that they were sification was performed using ENVI 5.0 software. Hycza et al. New Zealand Journal of Forestry Science (2018) 48:18 Page 6 of 13 Accuracy assessment fraction of correctly classified pixels with regard to all The accuracy analysis was performed using 300 test pixels pixels of that ground truth class. The user’s accuracy is and 2 MNF-band compositions (5-4-3 and 4-3-2) in which the fraction of correctly classified pixels with regard to the differences in colour among the species were most ob- all pixels classified as this class in the classified image. servable. Some of the pixels representing trees in the sam- For the classification based on all 129 spectral bands ple plots were used as test pixels, but only those that were performed with the SVM algorithm, the highest pro- most recognisable were used (European beech: 50, birch: ducer’s accuracy was observed for European larch, 20, oak: 50, hornbeam: 10, European larch: 50, no forest: no-forest, and Scots pine (92–98%) and the lowest was 50, Scots pine: 50, and Norway spruce: 20). for birch and hornbeam (10%). The highest user accur- The pixels of each class within each plot were ran- acy was observed for Scots pine and hornbeam (94– domly divided into training and validation data sets. 100%) and the lowest was for European beech (52.4%). There was no spatial distinction between the plots of in- For the classification based on 36 spectral bands per- dividual training and test pixels; however, in some cases, formed with the NN algorithm, the highest producer’s only training or only test pixels might have been chosen accuracy was observed for European larch, no-forest, for a single plot. Nevertheless, both data sets covered and Scots pine (94%) and the lowest was observed for the entire study area randomly. birch and hornbeam (0%). The highest user accuracy A normalised Digital Surface Model was used to was observed for Scots pine and hornbeam (71.2–77.4%) choose the test pixels representing only one particular and the lowest was for birch, hornbeam, and Norway species and to overcome inaccuracies caused by the spruce (0%). spatial resolution of the image and the GNSS device. For the classification based on three PCA spectral Only pixels in which a tree top was located close to the bands performed with the ML algorithm, the highest centre were selected as test pixels. The accuracy assess- producer’s accuracy was observed for European larch ment was performed using ENVI 5.0 software. and no-forest (100%) and the lowest was for birch (10%). The classification results of 98 individual sample plots The highest user’s accuracy was observed for Scots (values represented in %) were compared to the number pine and Norway spruce (100%) and the lowest was for of trees (values represented in %) belonging to individual birch (40%). species on each plot using the coefficient of determin- For the classification based on 7 MNF spectral bands ation (R ) calculated in Statistica 8.0. Only trees from performed with the ML algorithm, the highest pro- the upper canopy were taken into consideration. ducer’s accuracy was observed for beech, European larch, and no-forest (100%) and the lowest was for birch Results (10%). The highest user’s accuracy was observed for The highest accuracy was obtained by the ML algorithm hornbeam, European larch, Scots pine, and Norway and the data set of the seven MNF bands. The final map spruce (100%) and the lowest was for birch (33.33%) was subjected to Majority Filter analysis. The overall ac- (Table 3). Birch was spread across the study area with no curacy was 91.3% (Kappa—0.9) (Fig. 3.). observable concentration while hornbeam was very rare; The classification performed on all 129 bands ranged only one sample plot contained enough of the latter from 31 (BE) to 76.7% (SVM), excluding P and NN (85%) to be observable from the aerial ceiling. (below 10%). Spectral Information Divergence also per- The visual comparison of these four classification ap- formed relatively well (64.7%). There were not enough proaches on a single-plot scale is shown for two chosen training pixels to perform ML and MD2. The classifica- plots in Figs. 4 and 5. The best results were achieved tion performed on the 36 original bands ranged from using the ML algorithm. 33.6 (BE) to 66.3% (NN), excluding P (below 20%). The coefficient of determination between the number Support vector machine also performed relatively well of trees and the classification results of individual sample (58.7%). There were not enough training pixels to per- plots ranged from 0.68 (birch) to 0.99 (European larch), form ML and MD2. The classification performed on the while those of Norway spruce, hornbeam, and oak were first three PCA bands ranged from 30.3 (P) to 88.3% approximately 0.9 (Table 4). (ML). NN, SAM, and MD2 ranged from 68.3 to 72.7%. The classification performed on the first seven MNF Discussion bands ranged from 10.3 (P) to 90.7% (ML). Spectral in- Hyperspectral images are difficult to use for classifica- formation divergence and SAM also performed relatively tion purposes because they contain several narrow bands well (84.7%–85%) (Table 2). that are correlated with one another. It is important to The producer’s and user’s accuracy for the four best reduce both the amount of data and the noise before classification results (each based on a different data set) performing classifications. Clark et al. (2005) observed a is provided in Table 3. The producer’s accuracy is the general increase in accuracy of up to 30 input bands Hycza et al. New Zealand Journal of Forestry Science (2018) 48:18 Page 7 of 13 Fig. 3 Classification results (maximum likelihood on seven minimum noise fraction bands). Legend: red—birch, orange—European beech, yellow—oak, pink—hornbeam, pale blue—European larch, green—Scots pine, dark blue—Norway spruce when using a feature selection algorithm combined high-dimensional data well without the need for a with a linear discriminant analysis classifier; including large training sample size. Thus, it is not strongly af- more bands produced a lower or equal accuracy when fected by the Hughes phenomenon (Dalponte et al. classifying tree species in a tropical environment. Dal- 2009, Hughes 1968), which states that as the number ponte et al. (2009) reported a slight decrease in ac- of hyperspectral narrow bands increases, the number curacy when dropping several bands from the initial of samples (i.e. training pixels) required to maintain a 126 in a tree-species classification that combined an minimum statistical confidence and functionality in SVM classifier with a feature-selection procedure. hyperspectral data for classification also increases ex- These findings were most likely also connected to the ponentially, making it very difficult to address this classifiers applied, given that SVM is known to handle issue adequately. Hycza et al. New Zealand Journal of Forestry Science (2018) 48:18 Page 8 of 13 Table 2 Classification results Parameter Classification All 129 bands 36 chosen bands First three Principal component First seven minimum noise analysis bands fraction bands Algorithm Accuracy (%) Kappa Accuracy (%) Kappa Accuracy (%) Kappa Accuracy (%) Kappa Parallelepiped 5.3 0.005 15.3 0.05 30.3 0.23 10.3 0.06 Minimum Distance 62.7 0.56 40.3 0.3 61.7 0.55 84.7 0.82 Mahalanobis distance NETP NETP NETP NETP 72.7 0.68 81.3 0.78 Maximum likelihood NETP NETP NETP NETP 88.3 0.86 90.7 0.89 Spectral angle mapping 75 0.7 39.3 0.3 69.3 0.64 85 0.82 Spectral information divergence 64.7 0.58 38.3 0.28 37 0.27 81.3 0.78 Binary encoding 31 0.22 33.6 0.23 11.7 0.05 44 0.37 Neural networks 6.7 0.0004 66.3 0.6 68.3 0.62 63.7 0.56 Support vector machine 76.7 0.72 58.7 0.5 61 0.53 72.7 0.68 NETP not enough training pixels The compositions made from the PCA or MNF 2013; Goodwin et al. 2005)at80–100%, in a tropical and bands may be useful to distinguish tree species and sub-tropical zone (Carlson et al. 2010; Dian et al. 2014; create a layer of training pixels or polygons used to Goodwin et al. 2005; Dennison and Roberts 2003; Lucas perform the supervised classification. However, PCA is et al. 2008; Yang et al. 2009; Gong et al. 1997; van Aardt not the most suitable method to reduce multidimen- and Norris-Rogers 2008) at over 90%, and in a temperate sionality when the objective is to classify remotely zone (Zagajewski 2010; Olesiuk and Zagajewski 2008; sensed data (Cheriyadat and Bruce 2003). Principal Bartold 2008; Dian et al. 2014; Martin et al. 1998; component analysis identifies variabilities that may not Dalponte et al. 2013; Dmitriev 2014; Tarabalka 2010; perform well in multi-class discrimination and does not Richter et al. 2016)at74–93%. differentiate between within-group and between-group Our results have a high correspondence with tree variations (Hobro et al. 2010). species frequencies at the sample-plot level. Differences The classification of the first seven MNF bands using between the classification results and data from the local the ML algorithm resulted in the best overall accuracy survey may be explained by the leaves and branches of (91.3%) and kappa (0.9). The results are comparable to the trees growing near, but outside the borders of, the those obtained for a forest species in an equatorial zone testing areas. Stumps were observed in the field, so it is (Clark et al. 2005; Mickelson et al. 1998; Peerbhay et al. possible that some parts of unmapped trees were Table 3 The producer’s and user’s accuracy for each class using different datasets and algorithms Classification Support vector machine over Neural net over 36 bands Maximum likelihood over first Maximum likelihood over 129 bands three principal component first seven minimum noise analysis bands fraction bands Species Producer’s User’s Producer’s User’s Producer’s User’s Producer’s User’s accuracy (%) accuracy (%) accuracy (%) accuracy (%) accuracy (%) accuracy (%) accuracy (%) accuracy (%) Birch 10.00 66.67 0.00 0.00 10.00 40.00 10.00 33.33 European 86.00 52.44 76.00 59.38 84.00 76.36 100.00 70.42 beech Oak species 68.00 82.93 80.00 55.56 92.00 85.19 94.00 95.92 Hornbeam 10.00 100.00 0.00 0.00 70.00 63.64 60.00 100.00 European 92.00 83.64 66.00 73.33 100.00 90.91 100.00 100.00 larch No-forest 98.00 83.64 82.00 77.36 100.00 96.15 100.00 98.04 Scots pine 94.00 94.00 94.00 71.21 98.00 100.00 98.00 100.00 Norway 40.00 88.89 0.00 0.00 95.00 100.00 90.00 100.00 spruce Hycza et al. New Zealand Journal of Forestry Science (2018) 48:18 Page 9 of 13 Table 4 Coefficient of determination between the number of trees (%) and the classification results (%) of individual species on individual test plots Species R European beech 0.79 Birch 0.68 Oak 0.93 Hornbeam 0.9 European larch 0.99 Scots pine 0.81 Norway spruce 0.89 included in the sample. It is also possible that the re- flectance of the input image was disturbed by the plants growing in the lower canopy layers of the stand. Add- itionally, even the same tree species may have different values of spectral reflectance depending on their age, weather and soil conditions, moisture, vegetation period, and many other factors (Ghiyamat and Shafri 2010), Fig. 4 Comparison of the results of four classification techniques (SVM-129, NN-36, ML-PCA, ML-MNF) on a single sample plot (Adams which is the premise for using hyperspectral imagery to et al. 1995). Legend: red—birch, orange—European beech, detect disease and nutrient deficiencies in even-aged yellow—oak, pink—hornbeam, pale blue—European larch, single-species stands. green—Scots pine, dark blue—Norway spruce The set of training polygons used in this study was suitable for performing the classification on a neighbour- ing area using the same type of data (AISA Eagle hyper- spectral image), acquired at the same flight height, during the peak of the vegetation season (July and August), when the weather conditions were similar (although the atmospheric correction was performed). Otherwise, the set of training polygons used should be separate because the spectral signatures of the different tree species varied due to the study area, data type, ac- quisition date, weather conditions, altitude, and other factors (Ghiyamat and Shafri 2010). However, this is a common issue when dealing with remotely sensed data. Unfortunately, more issues can be expected with hyper- spectral data; for example, a comparison to satellite im- ages and reference data is needed. This is due the fact that flight strips are relatively narrow and a longer time is needed to cover large areas. As the result, there will be large differences between single strips or groups of strips. In these cases, a smaller part of the data set is re- quired for training, and verification can be undertaken immediately. It is also important to select the training and test pixels from the same (or at least neighbouring) areas, using the same methodology, and with a similar propor- tion of class samples to avoid differences between the accuracy assessment and the true classification results. Fig. 5 Comparison of the results of four classification techniques (SVM-129, NN-36, ML-PCA, ML-MNF) on a single sample plot (Alonzo et al. 2014). Legend: red—birch, orange—European beech, Conclusions yellow—oak, pink—hornbeam, pale blue—European larch, The classification based on 7 MNF spectral bands per- green—Scots pine, dark blue—Norway spruce formed with the ML algorithm was found to be the most Hycza et al. New Zealand Journal of Forestry Science (2018) 48:18 Page 10 of 13 Appendix 1 Table 5 Detailed information on the Miłomłyn Forest District Forest area (ha) 19,000 Total area (ha) 48,000 Forest area (FA)/total area (TA) % 39.6 Mean FA/TA in Poland % 29.1 Abundance of forest site types (%) Fresh* mixed forest 42 Fresh* deciduous forest 23 Fresh* mixed coniferous forest 22 Other forest types 13 Abundance of forest species group types (%) Coniferous trees 71 Deciduous trees 29 Most common tree species (%) Scots pine (Pinus sylvestris L.) and European larch (Larix decidua L.) 71 European beech (Fagus sylvatica L.) 12 Grey alder (Alnus glutinosa L.) 6 Birch species (Betula spp. L.) 5 Oak species (Quercus spp. L.) 4 Norway spruce (Picea abies L.) 1 Other (e.g. hornbeam Carpinus betulus L.) 1 Age classes (%) I(0–20) 7 II (Dennison and Roberts 2003; Dian et al. 2014; Dmitriev 2014; Einzmann et al. 2014; Farreira et al. 2016; Fassnacht et al. 2014; Fassnacht et 14 al. 2016; Gao and Hoetz 1990; Ghiyamat and Shafri 2010; Ghiyamat et al. 2013; Gong et al. 1997; Goodwin et al. 2005; Grant 1987; Han et al. 2004; Hainzel and Koch 2012; Harsanyi 1994; Hobro et al. 2010; Hughes 1968; Immitzer et al. 2015; Innes and Koch 1998; Jansson and Angelstam 1999) III (Jansson and Angelstam 1999; Jones et al. 2010; Kennedy and Southwood 1984; Korpela and Tokola 2006; Korpela et al. 2011; Leckie et 23 al. 2005; Lee et al. 2010; Li et al. 2014; Lucas et al. 2008; Luo and Chanussot 2009; Martin et al. 1998; Mickelson et al. 1998; Olesiuk and Zagajewski 2008; Ørka et al. 2013; Pausas et al. 1997; Peerbhay et al. 2013; Plourde et al. 2007; Portigal et al. 1997; Ribeiro da Luz and Crowley 2007; Richter et al. 2016; Roberts et al. 1997) IV (Roberts et al. 1997; Salisbury 1986; Salisbury and Milton 1998; Shang and Chisholm 2014; Schull et al. 2010; Stavrakoudis et al. 2014; 17 Tarabalka 2010; Tompalski et al. 2014; Treuhaft et al. 2002; Ullah et al. 2012; Ustin et al. 2009; Van Aardt and Wynne 2007; Van Aardt and Norris-Rogers 2008; Van Ewijk et al. 2014; Vauhkonen et al. 2014; Villa et al. 2013; Voss and Sugumaran 2008; Waser et al. 2014; Wietecha et al. 2017; Wulder et al. 2006; Yang et al. 2009) V (80–100) 17 VI (100–120) 8 VII (120–140) 7 VII + (> 140) 7 Mean volume for species (m /ha) Scots pine (Pinus sylvestris L.) 241 Norway spruce (Picea abies L.) 216 European beech (Fagus sylvatica L.) 260 Oak species (Quercus spp. L.) 310 *According to the soil moisture level, forests are divided into dry, fresh, wet, and swampy Hycza et al. New Zealand Journal of Forestry Science (2018) 48:18 Page 11 of 13 Appendix 2 Table 6 Algorithm settings (Continued) Radiometric calibration Table 6 Algorithm settings Type Reflectance Radiometric calibration Penalty parameter 100 Type Reflectance Output interleave BSQ Pyramid levels 0 Output data type Float Classification probability threshold 0 Scale factor 1 Quick atmospheric correction accurate method for classifying species (overall accuracy of 90.3%), with the highest kappa coefficient of 0.9. The Sensor type AISA results from the study reported here showed that this PCA transformation method is sufficiently reliable, accurate and user-friendly Stats X resize factor 1 to be used in practice. However, the data and software Stats Y resize factor 1 required are still expensive, which may limit its practical Calculated using Covariance matrix use by forest managers at present. Output data file Floating point Select subset from eigenvalues No Abbreviations Number of output PC bands 129 AISA: Airborne Imaging Spectrometer for Application; BE: Binary encoding; MNF transformation BSQ: Band sequential; DGPS: Differential Global Positioning System; FOV: Field of view; GNSS: Global Navigation Satellite System; LDA: Linear Shift difference subset Full scene discriminant analysis; LIDAR: Light detection and ranging; LSU: Linear spectral Select subset from eigenvalues No unmixing; MD1: Minimum distance; MD2: Mahalanobis distance; ML: Maximum likelihood; MNF: Minimum noise fraction; nDSM: Normalised Number of output MNF bands 129 Digital Surface Model; NN: Neural net; P: Parallelepiped; PCA: Principal component analysis; PPI: Pixel Purity Index; SAM: Spectral angle mapping; Parallelepiped SAR: Synthetic aperture radar; SID: Spectral information divergence; Max standard deviation from mean 3 SVM: Support vector machine Minimum distance Acknowledgements Max standard deviation from mean 3 We would like to thank Mariusz Ciesielski, Leopold Leśko, Aleksander Rybski, Marek Przywózki, Maciej Sarnowski, and Michał Brach, who conducted the Max distance error 0 local survey and established 98 sample plots in the Miłomłyn Forest District. Mahalanobis distance Funding Max standard deviation from mean 3 The study was performed in relation to the project entitled “Modelling Maximum likelihood carbon budget on the local and global scale in the State Forests Holding and developing scientific input parameters and management scenarios for Max standard deviation from mean 3 Poland” funded by the State Forests (grant number BLP-392; grant recipient: Mr. Radomir Bałazy). Data scale factor 255 Spectral angle mapping Availability of data and materials The aerial image from the AISA Eagle camera was provided by MGGP Aero. Maximum angle 0.1 The image is property of the Forest Research Institute and may be disclosed Spectral information divergence with permission from the Director of the Forest Research Institute. The analysis was performed using ENVI 5.0 and ArcGIS 10.3 provided by ESRI Maximum divergence threshold 0.05 Polska. Binary encoding Authors’ contributions Minimum encoding threshold 0 TH undertook the literature review, methodology, analysis, manuscript Neural networks writing, and field survey. KS was involved in conceiving and planning the study, methodology, analysis, manuscript writing, and reviewing. RB Activation Logistic compiled the data and literature review, and reviewed the manuscript. All authors read and approved the final manuscript. Training threshold contribution 0.9 Training rate 0.2 Authors’ information TH—Master of Science in Remote Sensing and Geoinformatics, Assistant in Training momentum 0.9 the Forest Research Institute, and doctoral student. Training RMS exit criteria 0.1 KS—Ph.D.in Forestry, Adjunct in the Forest Research Institute, and coordinator on two projects (“A complex forest dynamics monitoring of the Number of hidden layers (and nodes) 1 (129, 36, 3, 7). Białowieża Forest based on remote sensing data” and “An estimation of Number of training iterations 1000 biomass and carbon resources in forests based on remote sensing data”) with 75 publications and 223 citations (Research Gate, February 2018). Minimum output activation threshold 0 RB—Master of Science in Forestry, Assistant in the Forest Research Institute, Support vector machine and coordinator of two projects (“A forest information system of monitoring and forest condition assessment of Sudety and West Beskidy” and “Carbon Hycza et al. New Zealand Journal of Forestry Science (2018) 48:18 Page 12 of 13 budget modelling of the Polish State Forests Holding on the local and Dalponte, M., Bruzzone, L., & Gianelle, D. (2012). Tree species classification in the global scale and the development of input parameters and economic southern Alps based on the fusion of very high geometrical resolution scenarios for Poland”) with 53 publications and 120 citations (Research Gate, multispectral/hyperspectral images and LIDAR data. Remote Sensing of February 2018). Environment, 123, 258–270. Dalponte, M., Bruzzone, L., Vescovo, L., & Gianelle, D. (2009). The role of spectral resolution and classifier complexity in the analysis of hyperspectral images of Ethics approval and consent to participate forest areas. Remote Sensing of Environment, 133(11), 2345–2355. Not applicable. Dalponte, M., Ørka, H. O., Ene, L. T., Gobakken, T., & Næsset, E. (2013). Tree crown delineation and tree species classification in boreal forests using Consent for publication hyperspectral and ALS data. Remote Sensing of Environment, 140, 306–317. Not applicable. Dennison, P. E., & Roberts, D. A. (2003). The effects of vegetation phenology on endmember selection and species mapping in southern California chaparral. Competing interests Remote Sensing of Environment, 87, 295–309. The authors declare that they have no competing interests. Dian, Y., Li, Z., & Pang, Y. (2014). Spectral and texture features combined for forest tree species classification with airborne hyperspectral imagery. Indian Society of Remote Sensing, 43(1), 101–107. Publisher’sNote Dmitriev, E. (2014). Classification of the forest cover of Tver oblast using Springer Nature remains neutral with regard to jurisdictional claims in hyperspectral airborne images. Izvestiya - Atmospheric and Oceanic Physics, published maps and institutional affiliations. 50(9), 929–942. Einzmann, K., Immitzer, M., Bachmann, M., Pinnel, N., & Atzberger, C. (2014). Received: 23 February 2018 Accepted: 27 November 2018 Method analysis for collecting and processing in-situ hyperspectral needle reflectance data for monitoring Norway spruce. Photogrammetrie, Fernerkundung, Geoinformation, 5, 423–434. References Farreira, M. P., Zanotta, D. C., Zortea, M., & andde Souza Filho, C. R. (2016). Mapping Adams, J. B., Sabol, D. E., Kapos, V., Filho, R. A., Roberts, D. A., Smith, M. O., & tree species in tropical seasonal semi-deciduous forests with hyperspectral and Gillespie, A. R. (1995). Classification of multispectral images based on multispectral data. Remote Sensing of Environment, 179,66–78. fractions of endmembers: Application to land−cover change in the Brazilian Fassnacht, F. E., Latifi, H., Stereńczak, K., Modzelewska, A., Lefsky, M., Waser, L. T., Amazon. Remote Sensing of Environment, 52(2), 137–154. Straub, C., & Ghosh, A. (2016). Review of studies on tree species classification Alonzo, M., Bookhagen, B., & Roberts, D. A. (2014). Urban tree species mapping from remotely sensed data. Remote Sensing of Environment, 186,64–87. using hyperspectral and LIDAR data fusion. Remote Sensing of Environment, Fassnacht, F. E., Neumann, C., Förster, M., Buddenbaum, H., Ghosh, A., Clasen, A., 148,70–83. Joshi, P. K., & Koch, B. (2014). Comparison of feature reduction algorithms for Asner, G. P. (1998). Biophysical and biochemical sources of variability in canopy classifying tree species with hyperspectral data on three central European reflectance. Remote Sensing of Environment, 64(3), 234–253. test sites. IEEE Journal of Selected Topics in Applied Earth Observation and Bartold, M. (2008). Klasyfikacja drzewostanów na obrazie satelitarnym Hyperion Remote Sensing, 7(6), 2547–2561. (EO-1). Teledetekcja Środowiska, 39,5–29. Gao, B. C., & Hoetz, A. F. H. (1990). Column atmospheric water-vapor and Bellanti, L., Blesius, L., Hinnes, E., & Kruse, B. (2016). Tree species classification vegetation liquid water retrievals from airborne imaging spectrometer data. using hyperspectral imagery: A comparison of two classifiers. Remote Sensing, Journal of Geophysical Research-Atmospheres, 95, 3549–3564. 8(6), 445–463. Ghiyamat, A., & Shafri, H. Z. M. (2010). A review on hyperspectral remote sensing Bernstein, L. S., Sundberg, R. L., Levine, R. Y., Perkins, T. C., & Berk, A. (2005). A new for homogeneous and heterogeneous forest biodiversity assessment. method for atmospheric correction and aerosol optical property retrieval for International Journal of Remote Sensing, 31, 1837–1856. VIS−SWIR multi− and hyperspectral imaging sensors: QUAC (quick Ghiyamat, A., Shafri, H. Z. M., Mahdiraji, G. A., Shariff, A. R. M., & Mansor, S. (2013). atmospheric correction). Proceedings of the IEEE International Geoscience and Hyperspectral discrimination of tree species with different classifications Remote Sensing Symposium, 5, 3549–3552. using single- and multiple-endmember. International Journal of Applied Earth Caiyun, Z., & Fang, Q. (2012). Mapping individual tree species in an urban forest Observation and Geoinformation, 23, 177–191. using airborne LIDAR data and hyperspectral imagery. Photogrammetric Gong, P., Pu, R., & Zu, B. (1997). Conifer species recognition: An exploratory analysis Engineering and Remote Sensing, 78(10), 1079–1087. of in situ hyperspectral data. Remote Sensing of Environment, 62(2), 189–200. Carlson, K. M., Asner, G. P., Hughes, R. F., Ostertag, R., & Martin, R. E. (2010). Goodwin, N., Turner, R., & Merton, R. (2005). Classifying eucalyptus forests with high Hyperspectral remote sensing of canopy biodiversity in Hawaiian lowland spatial and spectral resolution imagery. An investigation of individual species rainforests. Ecosystems, 10, 526–549. and vegetation communities. Australian Journal of Botany, 53,337–345. Chambers, D., Périé, C., Casajus, N., & de Blois, S. (2013). Challenges in modelling Grant, L. (1987). Diffuse and specular characteristics of leaf reflectance. Remote the abundance of 105 tree species in eastern North America using climate, Sensing of Environment, 22, 309–322. edaphic, and topographic variables. Forest Ecology and Management, 291, Hainzel, J., & Koch, B. (2012). Investigating multiple data sources for tree species 20–29. classification in temperate forest and use for single tree delineation. International Cheriyadat, A., & Bruce, L. M. (2003). Why principal component analysis is not an Journal of Applied Earth Observation and Geoinformation, 18,101–110. appropriate feature extraction method for hyperspectral data. Proceedings of Han, T., Goodenough, D. G., Dyk, A., & Chen, H. (2004). Hyperspectral feature the IEEE International. Geoscience Remote Sensing Symposium, 6, 3420–3422. selection for forest classification IEEE International Geoscience and Remote Cho, M. A., Debba, P., Mathieu, R., Naidoo, L., van Aardt, J., & Asner, G. P. (2010). Sensing Symposium, 2) (pp. 1471–1474). Anchorage. Improving discrimination of savanna tree species through a multiple- Harsanyi, J. C. (1994). Hyperspectral image classification and dimensionality endmember spectral angle mapper approach: Canopy-level analysis. IEEE reduction: An orthogonal subspace projection approach. Geoscience and Transactions on Geoscience and Remote Sensing, 48(11), 4133–4142. Remote Sensing, 32(4), 779–785. Clark, M., Roberts, D., & Clark, D. (2005). Hyperspectral discrimination of tropical Hobro, A. J., Kuligowski, J., Döll, M., & Lendl, B. (2010). Differentiation of walnut rain forest tree species at leaf to crown scales. Remote Sensing of wood species and steam treatment using ATR-FTIR and partial least squares Environment, 96(3–4), 375–398. discriminant analysis (PLS-DA). Analytical and Bioanalytical Chemistry, 398(6), Clark, M. L., & Roberts, D. A. (2012). Species-level differences in hyperspectral 2713–2722. metrics among tropical rainforest trees as determined by a tree-based Hughes, G. F. (1968). On the mean accuracy of statistical pattern recognizers. classifier. Remote Sensing, 4(6), 1820–1855. Transactions on Information Theory, 14,55–63. Dalponte, M., Bruzzone, L., & Gianelle, D. (2008). Fusion of hyperspectral and LIDAR remote sensing data for classification of complex forest areas. IEEE Immitzer, M., Atzberger, C., Einzmann, K., Böck, S., Mattiuzzi, M., Wallner, A., Seitz, R., Transactions on Geoscience and Remote Sensing, 46, 1416–1427. Pinnel, N., Müller, A., & Frost, M. (2015). Fichten- und Kiefernkarte für Bayern. LWF Dalponte, M., Bruzzone, L., & Gianelle, D. (2011). Tree species classification in the aktuell. Bayerische Landesanstalt für Wald und Forstwirtschaft, 106,30–37. southern Alps with very high geometrical resolution multispectral and Innes, J. L., & Koch, B. (1998). Forest biodiversity and its assessment by remote hyperspectral data. Remote Sensing of Environment, 123, 258–270. sensing. Global Ecology and Biogeography Letters, 7, 397–419. Hycza et al. New Zealand Journal of Forestry Science (2018) 48:18 Page 13 of 13 Jansson, G., & Angelstam, P. (1999). Threshold levels of habitat composition for Salisbury, J. W., & Milton, N. M. (1998). Thermal infrared (2.5- to 13.5-μm) the presence of long-tailed tit (Aegithalos audatus) in a boreal landscape. directional hemispherical reflectance of leaves. Photogrammetric Engineering Landscape Ecology, 14, 283–290. and Remote Sensing, 54(9), 1301–1304. Jones, T. G., Coops, N. C., & Sharma, T. (2010). Assessing the utility of airborne Schull, M. A., Knyazikhin, Y., Xu, L., Samanta, A., Ganguly, S., & Latorre Carmona, P. hyperspectral and LIDAR data for species distribution mapping in the coastal (2010). Canopy spectral invariants. Part 2: Application to classification of Pacific Northwest, Canada. Remote Sensing of Environment, 114(12), 2841–2852. forest types from hyperspectral data. Journal Quantitative Spectroscopy Radiative Transfer, 112, 736–750. Kennedy, C. E. J., & Southwood, T. R. E. (1984). The number of species of insects Shang, X., & Chisholm, L. A. (2014). Classification of Australian native Forest associated with British trees: A re-analysis. Journal of Animal Ecology, 53(2), 455–478. species using hyperspectral remote sensing and machine-learning Korpela, I., Heikkinen, V., Hokavaara, E., Rohrbach, F., & Tokola, T. (2011). Variation classification algorithms. IEEE Journal of Selected Topics in Applied Earth and directional anisotropy of reflectance at the crown scale – Implications Observation and Remote Sensing, 7(6), 2481–2489. for tree species classification in digital aerial images. Remote Sensing of Stavrakoudis, D. G., Dragozi, E., Gitas, I. Z., & Karydas, C. G. (2014). Decision fusion Environment, 115, 2062–2074. based on hyperspectral and multispectral satellite imagery for accurate forest Korpela, I. S., & Tokola, T. E. (2006). Potential of aerial image-based monoscopic species mapping. Remote Sensing, 6(8), 6897–6928. and multiview single-tree forest inventory: A simulation approach. Forest Tarabalka, Y. (2010). Segmentation and classification of hyperspectral images Science, 52(2), 136–147. using minimum spanning forest grown automatically selected markers. IEEE Leckie, D. G., Tinis, S., Nelson, T., Burnett, C., Gougeon, F. A., Cloney, E., & Paradine, Transactions on Systems, Man, and Cybernetics, 40(5), 1267–1279. D. (2005). Issues in species classification of trees in old growth conifer stands. Tompalski, P., Coops, N. C., White, J. C., & Wulder, M. A. (2014). Simulating the Canadian Journal of Remote Sensing, 31(2), 175–190. impacts of error in species and height upon tree volume derived from Lee, W., S., Alchanatis, V., Yang, C., Hirafuji, M., Moshou, D., Li, C. (2010). Sensing airborne laser scanning data. Forest Ecology and Management, 327, 167–177. technologies for precision specialty crop production. Computers and Treuhaft, R. N., Asner, G. P., Law, B. E., & Tuyl Van, S. (2002). Forest leaf area Electronics in Agriculture, 74, 2–33. density profiles from the quantitative fusion of radar and hyperspectral data. Li, J., Xin, H., Gamba, P., & Bioucas-Dias, J. M. (2014). Multiple feature learning for Journal of Geophysical Research, 107(D21), ACL 7–1 – ACL 7-13. hyperspectral image classification. Transactions on Geoscience and Remote Ullah, S., Schlerf, M., Skidmore, A. K., & Hecker, C. (2012). Identifying plant species Sensing, 53(3), 1592–1606. using mid wave infrared (2.5–6 μm) and thermal infrared (8–14 μm) Lucas, R. M., Lee, A. C., & Bunting, P. J. (2008). Retrieving forest biomass through emissivity spectra. Remote Sensing of Environment, 118,95–102. integration of CASI and LIDAR data. International Journal of Remote Sensing, 29, Ustin, S., Gitelson, A. A., Jacquemoud, S. M., Asner, G. P., Gamon, J., & Zarco- 1553–1577. Tejada, P. J. (2009). Retrieval of foliar information about plant pigment Luo, B., & Chanussot, J. (2009). Hyperspectral image classification based on spectral systems from high resolution spectroscopy. Remote Sensing of Environment, and geometrical features (IEEE International Workshop on Machine Learning 113,67–77. for Signal Processing) (pp. 1–6). Van Aardt, J., & Norris-Rogers, M. (2008). Spectral-age interactions in managed, Martin, M. E., Newman, S. D., Aber, J. D., & Congalton, R. G. (1998). Determining even-aged eucalyptus plantations: Application of discriminant analysis and forest species composition using high spectral resolution remote sensing classification and regression trees approaches to hyperspectral data. data. Remote Sensing of Environment, 65, 249–254. International Journal of Remote Sensing, 29(6), 1841–1845. Mickelson, J. G., Civco, D. L., & Silander, J. A. (1998). Delineating forest canopy Van Aardt, J., & Wynne, R. H. (2007). Examining pine spectral separability using species in the north-eastern United States using multi-temporal TM imagery. hyperspectral data from an airborne sensor: An extension of field-based Photogrammetric Engineering and Remote Sensing, 64(9), 891–904. results. International Journal of Remote Sensing, 28(2), 431–436. Olesiuk, D., & Zagajewski, B. (2008). Wykorzystanie obrazów hiperspektralnych Van Ewijk, K. Y., Randin, C. F., Treitz, P. M., & Scott, N. A. (2014). Predicting fine-scale do klasyfikacji pokrycia terenu zlewni Bystrzanki. Teledetekcja Środowiska, tree species abundance patterns using biotic variables derived from LIDAR and 40, 125–148. high spatial resolution imagery. Remote Sensing of Environment, 150,120–131. Ørka, H. O., Dalponte, M., Gobakken, T., Næsset, E., & Ene, L. T. (2013). Vauhkonen, J., Ørka, H. O., Holmgren, J., Dalponte, M., Hainzel, J., & Koch, B. Characterizing forest species composition using multiple remote sensing (2014). Tree species recognition based on airborne laser scanning and data sources and inventory approaches. Scandinavian Journal of Forest complementary data sources. Forestry applications of airborne laser Research, 28(7), 677–688. scanning. In M. Maltamo, E. Naesset, & J. Vauhkonen (Eds.), Forestry Pausas, J. G., Austin, M. P., & Noble, I. R. (1997). A forest simulation model for Applications of Airborne Laser Scanning (pp. 135–156). Dordrecht: Springer. predicting eucalypt dynamics and habitat quality for arboreal marsupials. Villa, A., Chanossut, J., Benediktsson, J. A., Jutten, C., & Dambreville, R. (2013). Ecological Applications, 7(3), 921–993. Unsupervised methods for the classification of hyperspectral images with Peerbhay, K. Y., Mutanga, O., & Ismail, R. (2013). Commercial tree species low spatial resolution. Pattern Recognition, 46(6), 1556–1568. discrimination using airborne AISA Eagle hyperspectral imagery and partial Voss, M., & Sugumaran, R. (2008). Seasonal effect on tree species classification in least squared discriminant analysis (PLS-DA) in KwaZulu-Natal - South Africa. an urban environment using hyperspectral data. LIDAR, and an object- Remote Sensing, 79,19–28. oriented approach. Sensors, 8, 3020–3036. Plourde, L. C., Ollinger, S. V., Smith, M. L., & Martin, M. E. (2007). Estimating species Waser, L. T., Küchler, M., Jütte, K., & Stampfer, T. (2014). Evaluating the potential of abundance in a northern temperate forest using spectral mixture analysis. world-View-2 data to classify tree species and different levels of ash Photogrammetric Engineering and Remote Sensing, 73(7), 829–840. mortality. Remote Sensing, 6(5), 4515–4545. Portigal, F., Holasek, R., Mooradian, G., Owensby, P., Dicksion, M., Fene, M., Elliot, Wietecha, M., Modzelewska, A., & Stereńczak, K. (2017). Airborne hyperspectral M., Hall, E., and Driggett, D. (1997). Vegetation classification using red-edge data for the classification of tree species a temperate forests (Wykorzystanie first derivative and green peak statistical moment indices with the Advanced lotniczej teledetekcji hiperspektralnej w klasyfikacji gatunkowej lasów strefy Airborne Hyperspectral Imaging System (AAHIS). Third International Airborne umiarkowanej). Sylwan, 161(1), 3–17. Remote Sensing Conference and Exhibition. Copenhagen, Denmark, 7–10 Wulder, M. A., Dymond, C. C., White, J. C., Leckie, D. G., & Caroll, A. L. (2006). July 1997. II (Ann Arbor, MI: ERIM): 789–797. Surveying mountain pine beetle damage of forests: A review of remote Ribeiro da Luz, B., & Crowley, J. K. (2007). Spectral reflectance and emissivity sensing opportunities. Forest Ecology and Management, 221,27–41. features of broad leaf plants: Prospects for remote sensing in the thermal Yang, C., Everitt, J. H., Fletcher, R. S., Jensen, R. R., & Mausel, P. W. (2009). infrared (8.0–14.0 μm). Remote Sensing of Environment, 109, 393–405. Evaluating AISA+ hyperspectral imagery for mapping black mangrove along Richter, R., Reu, B., Wirth, C., Doktor, D., & Vohland, M. (2016). The use of airborne the South Texas. Gulf Coast. Photogrammetric Engineering and Remote hyperspectral data for tree species classification in a species-rich central Sensing, 75, 425–435. European forest area. International Journal of Applied Earth Observation and Zagajewski, B. (2010). Ocena przydatności sieci neuronowych i danych Geoinformation, 52, 464–474. hiperspektralnych do klasyfikacji roślinności Tatr Wysokich. Teledetekcja Roberts, D. A., Green, R. O., & Adams, J. B. (1997). Temporal and spatial patterns in Środowiska, 43, 1-113. vegetation and atmospheric properties from AVIRIS. Remote Sensing of Zarco-Tejada, P. J., Pushnik, J. C., Dobrowski, S., & Ustin, S. L. (2003). Steady-state Environment, 62, 223–240. chlorophyll a fluorescence detection from canopy derivative reflectance and Salisbury, J. W. (1986). Preliminary measurements of leaf spectral reflectance in the 8- double-peak red-edge effects. Remote Sensing of Environment, 84, 283–294. 14 μm region. International Journal of Remote Sensing, 7(12), 1879–1886.

Journal

New Zealand Journal of Forestry ScienceSpringer Journals

Published: Dec 28, 2018

References