Discriminating Forest Successional Stages, Forest Degradation, and Land Use in Central Amazon Using ALOS / PALSAR-2 Full-Polarimetric Data

: We discriminated di ﬀ erent successional forest stages, forest degradation, and land use classes in the Tapaj ó s National Forest (TNF), located in the Central Brazilian Amazon. We used full polarimetric images from ALOS / PALSAR-2 that have not yet been tested for land use and land cover (LULC) classiﬁcation, neither for forest degradation classiﬁcation in the TNF. Our speciﬁc objectives were: (1) to test the potential of ALOS / PALSAR-2 full polarimetric images to discriminate LULC classes and forest degradation; (2) to determine the optimum subset of attributes to be used in LULC classiﬁcation and forest degradation studies; and (3) to evaluate the performance of Random Forest (RF) and Support Vector Machine (SVM) supervised classiﬁcations to discriminate LULC classes and forest degradation. PALSAR-2 images from 2015 and 2016 were processed to generate Radar Vegetation Index, Canopy Structure Index, Volume Scattering Index, Biomass Index, and Cloude–Pottier, van Zyl, Freeman–Durden, and Yamaguchi polarimetric decompositions. To determine the optimum subset, we used principal component analysis in order to select the best attributes to discriminate the LULC classes and forest degradation, which were classiﬁed by RF. Based on the variable importance score, we selected the four ﬁrst attributes for 2015, alpha, anisotropy, volumetric scattering, and double-bounce, and for 2016, entropy, anisotropy, surface scattering, and biomass index, subsequently classiﬁed by SVM. Individual backscattering indexes and polarimetric decompositions were also considered in both RF and SVM classiﬁers. Yamaguchi decomposition performed by RF presented the best results, with an overall accuracy (OA) of 76.9% and 83.3%, and Kappa index of 0.70 and 0.80 for 2015 and 2016, respectively. The optimum subset classiﬁed by RF showed an OA of 75.4% and 79.9%, and Kappa index of 0.68 and 0.76 for 2015 and 2016, respectively. RF exhibited superior performance in relation to SVM in both years. Polarimetric attributes exhibited an adequate capability to discriminate forest degradation and classes of di ﬀ erent ecological succession from the ones with less vegetation cover. canopy biomass variations due to ﬁre events, but they were not capable of discriminating among intermediate classes. Forests mildly a ﬀ ected by ﬁre had a higher contribution to the scattering in the VV polarization, while forests strongly a ﬀ ected by ﬁre and those with multiple ﬁre events had a larger sensitivity to the HH polarization. The results suggest a greater contribution of horizontally distributed components, such as fallen stems and branches in areas severely a ﬀ ected by ﬁre.

• HH/HV ratio showed the best classification results in the time series, with a producer's accuracy (PA) of 84.7%, and a user's accuracy (UA) of 96.3%. Martins et al. [38] evaluated the sensitivity of the full polarimetric ALOS/PALSAR-1 to forest degradation caused by fires in the Brazilian Amazon. They used polarimetric and derivate attributes from backscattering coefficients (σ • ) to estimate above-ground biomass (AGB), from multiple regression models. The subset formed by anisotropy, double-bounce, orientation angle, volume index, phase difference, and σ • HV presented the best results. These results showed that the attributes were sensitive to canopy biomass variations due to fire events, but they were not capable of discriminating among intermediate classes. Forests mildly affected by fire had a higher contribution to the scattering in the VV polarization, while forests strongly affected by fire and those with multiple fire events had a larger sensitivity to the HH polarization. The results suggest a greater contribution of horizontally distributed components, such as fallen stems and branches in areas severely affected by fire. Recent studies focusing on multi-sensor analysis for LULC mapping and change detection in tropical regions [39,40] have used dual-polarimetric ALOS/PALSAR-2 images integrated with Landsat multispectral images. Pavanelli et al. [39] highlighted the SAR potential to discriminate savannah-like vegetation in the Brazilian Amazon, with an improvement in OA of 6% in relation to the classification results obtained only from optical data. De Alban et al. [40], using a similar dataset, assessed the tropical landscapes of southern Myanmar using Random Forest (RF) classifier. They found that SAR-derived texture achieved an OA ranging from 56% to 71%, depending on the dataset used. However, these reported studies rely on low cloud cover conditions during the satellite image acquisition. Hagensieker and Waske [41] highlighted the importance of using SAR images for overcoming such drawbacks in tropical regions. These authors used dual-polarized ALOS/PALSAR-2, RADARSAT-2, and TerraSAR-X images for LULC classification in the Brazilian Amazon, achieving OA of 62% for mono-temporal analysis with ALOS/PALSAR-2 images.
ALOS/PALSAR-2 was launched in 2014 as a successor of ALOS/PALSAR-1 that stopped operating in early 2011. Full-polarimetric images of the new ALOS mission, which includes improved PALSAR-2 sensor with better radiometric and spatial resolutions, have not been tested for LULC classification purposes in the Tapajós region yet. Similarly, this sensor's potential to discriminate forest in different successional stages and to detect burn scars in forested areas has not yet been evaluated either. These better radiometric quality and higher spatial resolution are expected to increase classification accuracy.
In this study, the specific goals are: (1) to test the potential of ALOS/PALSAR-2 acquired in 2015 and 2016, full-polarimetric images processed by the polarimetric target decomposition techniques to discriminate different LULC classes and to detect forest degradation; (2) to determine the optimum subset of attributes to be used in LULC classifications and degradation forest studies, and (3) to evaluate the performance of both RF and Support Vector Machine (SVM) classification algorithms to discriminate LULC classes and forest degradation over the TNF region.
It is important to highlight that although two dates are available, this study does not aim to quantitatively compare the results of both dates in terms of LULC changes. This is due to the fact that the ALOS/PALSAR-2 images used were acquired in different periods and seasons, with increased effects of seasonality. Additionally, land management practices also vary throughout the year, which can cause substantial errors in any change detection analysis [42][43][44][45].

Study Area
The study area is located near the city of Santarém in Pará State, Central Brazilian Amazon, along the BR-163 highway (Cuiabá-Santarém highway) and covers a portion of the TNF and surroundings (eastern side of the national park) (Figure 1). Some settlements, mainly the São Jorge community, are not inside the national park anymore, after the 2012 changes in the TNF boundary. Small-farm settlements are dominant in the São Jorge community with a landscape constituted by well and poorly managed fragments of pasturelands and forest fragments at different stages of succession.
The climate of the region is classified by Köppen as AmW (Monsoon type with short dry season), with an average annual temperature of 25 • C and average relative humidity of 85%. The average annual rainfall is 1909 mm, with a dry season between July and November [46]. The site is situated in a relatively flat area with an elevation between 80 m and 180 m [47,48]. Vegetation in the TNF is formed mainly by dense and open ombrophilous rainforests. These forests are composed of a continuous canopy of perennial trees with heights between 25 m and 30 m and occasional emergent trees that reach up to 50-60 m height [48]. This type of forest has several layers and sublayers under the canopy, including small trees, shrubby, and herbaceous plants. Tall trees are approximately evenly spaced due to the low light penetration into the canopy. There are high amounts of climbing plants, especially epiphytes and lianas, that often hide the outline of the trees [49]. The climate of the region is classified by Köppen as AmW (Monsoon type with short dry season), with an average annual temperature of 25 °C and average relative humidity of 85%. The average annual rainfall is 1909 mm, with a dry season between July and November [46]. The site is situated in a relatively flat area with an elevation between 80 m and 180 m [47,48]. Vegetation in the TNF is formed mainly by dense and open ombrophilous rainforests. These forests are composed of a continuous canopy of perennial trees with heights between 25 m and 30 m and occasional emergent trees that reach up to 50-60 m height [48]. This type of forest has several layers and sublayers under the canopy, including small trees, shrubby, and herbaceous plants. Tall trees are approximately evenly spaced due to the low light penetration into the canopy. There are high amounts of climbing plants, especially epiphytes and lianas, that often hide the outline of the trees [49].
Some areas inside the TNF are allowed to have sustainable selective logging while other areas, under full protection prerogatives, are in the advanced ecological successional stage. Part of the area has been degraded by fire, mainly by fire events that occurred at the end of 2015 and early 2016 [50,51]. Along the BR-163 highway (outside of the park), there are large areas with agricultural activities, cattle ranching, and forest fragments in different succession stages. Growing farming activities near the TNF and reduction of the original TNF area by the Federal Law no. 12.678/2012, have also imposed additional pressures on this conservation unit.

Field Data
A set of 288 samples were characterized in a field campaign conducted in September 2014 during the dry season. For each sample, representative field photos and geographic coordinates (latitude and longitude), using a Global Positioning System (GPS) receiver (Garmin eTrex10 model), were gathered. The sites were selected based on specific research interests and on previous field campaigns Some areas inside the TNF are allowed to have sustainable selective logging while other areas, under full protection prerogatives, are in the advanced ecological successional stage. Part of the area has been degraded by fire, mainly by fire events that occurred at the end of 2015 and early 2016 [50,51]. Along the BR-163 highway (outside of the park), there are large areas with agricultural activities, cattle ranching, and forest fragments in different succession stages. Growing farming activities near the TNF and reduction of the original TNF area by the Federal Law no. 12.678/2012, have also imposed additional pressures on this conservation unit.

Field Data
A set of 288 samples were characterized in a field campaign conducted in September 2014 during the dry season. For each sample, representative field photos and geographic coordinates (latitude and longitude), using a Global Positioning System (GPS) receiver (Garmin eTrex10 model), were gathered. The sites were selected based on specific research interests and on previous field campaigns carried out in 2012, 2014, and 2016 [35,48,52,53]. These previous authors have also defined site-specific land cover classes of interest, based on widely used criterion to differentiate succession processes in the Brazilian Amazon as presented by Vieira et al. [54] and Uhl et al. [55], as well as land use adequate to local landscape, presented by Pôssa et al. [35]. The dominant LULC typology for each site was defined as: Primary Forest (PF); Advanced Secondary Succession (SS3); Intermediate Secondary Succession (SS2); Initial Secondary Succession (SS1); Degraded Forest (DG); Poorly Managed Pasture (PP); Well Managed Pasture (WP); Cropland (CR); and Bare Soil and Fallow (BS) ( Table 1). Another field campaign was conducted in September 2016, where 267 samples were characterized. Most of these sites were the Remote Sens. 2020, 12, 3512 6 of 30 same as visited in 2014. It is important to highlight that although all the field surveys were conducted in a different time to SAR data acquisition, we assume that up to six months' time difference does not provide significant changes in LULC classes. A better delineation of the DG class produced by fires was done using the thermal anomalies and fire products from the Moderate Resolution Imaging Spectroradiometer (MODIS) sensor onboard TERRA and AQUA platforms (MOD14 and MYD14, respectively), obtained from the end of 2015 to early 2016. In addition, two Landsat-8 Operational Land Imager (OLI), multispectral images (path/row: 227/62) were used to visually analyze areas of degraded forests mainly by fire, comparing with those classified from ALOS/PALSAR-2 image. The images were obtained from the United States Geological Survey (USGS) website (earth explorer user interface). These orthorectified images have a spatial resolution of 30 m, radiometric resolution of 12 bits, and processing level of 1T. The selected images were obtained on 14 August 2015, which corresponds to the period before the fire events in the Tapajós region, and, from 29 June 2016, which corresponds to the period after the fire. Figure 2 shows the flowchart of the methodology used to discriminate the land cover, land use, and forest degradation in the TNF region. Our reference dataset consisted of samples obtained in the field campaigns. We used a combination of the reference samples and ALOS/PALSAR-2 full polarimetric images as predictor variables to discriminate the classes in the study area. Description of each step of the methodology is provided in the following subsections. To cover the entire study area, a mosaic of three ALOS/PALSAR-2 images in polarimetric StripMap-2 (SM2) mode was used for each year ( Figure 3). The images were obtained with 1.1 processing level (single look complex data) ( Table 2). For each ALOS/PALSAR-2 image, polarimetric scattering matrices and backscattering coefficients (σ°) were generated using the SNAP 7.0 software. A multilook processing was carried out with a window size of 3 pixels × 3 pixels (pixel sizes of 8.5 m by 9.5 m in azimuth and range directions, respectively). To reduce the speckle effect, the matrices were smoothed with the Refined Lee filter [56] (Figure 2). The filter was selected based on the results from previous studies developed in the same study area using ALOS/PALSAR-1 images [35,38,47]. In this study, the following window sizes were tested: 3 × 3, 5 × 5, 7 × 7, and 9 × 9 pixels. The window size of 5 × 5 was selected because it presented a best performance in the reduction of speckle effect and the best preservation of the integrity of spatial information in the SAR image (especially in areas with high landscape fragmentation). This filtering process reduced the speckle effects by 25% according to the Coefficient of Variation (CV) tests. To cover the entire study area, a mosaic of three ALOS/PALSAR-2 images in polarimetric StripMap-2 (SM2) mode was used for each year ( Figure 3). The images were obtained with 1.1 processing level (single look complex data) ( Table 2). For each ALOS/PALSAR-2 image, polarimetric scattering matrices and backscattering coefficients (σ • ) were generated using the SNAP 7.0 software. A multilook processing was carried out with a window size of 3 pixels × 3 pixels (pixel sizes of 8.5 m by 9.5 m in azimuth and range directions, respectively). To reduce the speckle effect, the matrices were smoothed with the Refined Lee filter [56] (Figure 2). The filter was selected based on the results from previous studies developed in the same study area using ALOS/PALSAR-1 images [35,38,47]. In this study, the following window sizes were tested: 3 × 3, 5 × 5, 7 × 7, and 9 × 9 pixels. The window size of 5 × 5 was selected because it presented a best performance in the reduction of speckle effect and the best preservation of the integrity of spatial information in the SAR image (especially in areas with high landscape fragmentation). This filtering process reduced the speckle effects by 25% according to the Coefficient of Variation (CV) tests.   After filtering, the following polarimetric decompositions were considered: (a) the Cloude-Pottier (CP) technique [57], generating the entropy (H), alpha angle (α), and anisotropy (A) attributes; (b) the van Zyl (VZ) technique [58], generating the surface (VZs), double-bounce (VZd), and volumetric (VZv) scatterings; (c) the Freeman-Durden (FD) technique [59], generating the surface (Ps), double-bounce (Pd), and volumetric (Pv) scattering attributes; and d) the Yamaguchi (YH) technique [60], generating the surface (Ys), double-bounce (Yd), volumetric scattering (Yv), and helix scattering (YH) attributes (see more details in Appendix A). These polarimetric decompositions were selected in order to test how effective they are for classification purposes. Although the Freeman-Durden and Yamaguchi technique similarly decomposes the polarimetric signal, the fourth additional component of Yamaguchi decomposition (YH) could perform well in the discrimination of specific successional forest stages. This component is an improved version of the tree-component Freeman technique [61]. After filtering, the following polarimetric decompositions were considered: (a) the Cloude-Pottier (CP) technique [57], generating the entropy (H), alpha angle (α), and anisotropy (A) attributes; (b) the van Zyl (VZ) technique [58], generating the surface (VZ s ), double-bounce (VZ d ), and volumetric (VZ v ) scatterings; (c) the Freeman-Durden (FD) technique [59], generating the surface (P s ), double-bounce (P d ), and volumetric (P v ) scattering attributes; and d) the Yamaguchi (YH) technique [60], generating the surface (Y s ), double-bounce (Y d ), volumetric scattering (Yv), and helix scattering (Y H ) attributes (see more details in Appendix A). These polarimetric decompositions were selected in order to test how effective they are for classification purposes. Although the Freeman-Durden and Yamaguchi technique similarly decomposes the polarimetric signal, the fourth additional component of Yamaguchi decomposition (Y H ) could perform well in the discrimination of specific successional forest stages. This component is an improved version of the tree-component Freeman technique [61].

ALOS/PALSAR-2 Data and Preprocessing
The SAR images were also radiometrically calibrated into σ • whose attributes were used to calculate Radar Vegetation Index (RVI) [62] and the Pope indexes (PI) [63] (Figure 2) of Canopy Structure Index (CSI), Volume Scattering Index (VSI), and Biomass Index (BMI) (see more details in Appendix A). A summary of the attributes is presented in Table 3. All images were individually processed. Images were geometrically corrected using a 30 m digital elevation model (DEM) derived from the Shuttle Radar Topography Mission (SRTM). The images were coregistered to generate the mosaics of each year.  For this study, we generated only the variables of Table 3 because they had shown the best performances between all the SAR attributes generated in other studies for the same region in Amazon considering some similar classes but with different SAR data and methodological approaches [35,47,51].

Water Body Masking
Bare soils and water bodies, when interacting with electromagnetic energy, tend to produce relatively low levels of surface scattering, mainly at the L-band wavelengths. Similar backscattering of these targets can produce poor separability classes, resulting in misclassifications of both targets [64], so that we used a water body mask. To create the mask, we used the vector-based Main Rivers file available in the National Water Agency (ANA) database [65]. This product was mapped on 1:1,000,000 scale, and the drainage area was extracted from SRTM data. In the mask building process, we clipped the water bodies found in the study area and assigned them a value equal to zero. The remaining areas assigned a value equal to one, producing a binary mask: water/nonwater. During this process the areas with a value equal to one were considered in the calculation, whereas the areas with a value equal to zero were discarded [66]. Then, the mask was applied to all polarimetric decompositions and to the full-pol images calibrated into σ • .

LULC Classes
In order to classify the LULC classes, we used reference samples obtained in the field campaigns, as previously described in Section 2.2. Each field sample was georeferenced based on the GPS coordinates. To reduce the speckle effect that is inherent to SAR data, each sampling point was converted to polygons through the buffer function available in the QGIS software. After some tests, the buffer zone of 24.75 m was selected, which is equivalent to an area of~1800 m 2 for each polygon. The criterion used here was the visual analysis of the sample spatial distribution in our study area. High spatial resolution images from Google Earth™ platform (12/30/2014 and 30/12/2015) were used to support the inspection of each polygon. Figure 4 shows the location of the polygon samples over nine LULC classes found in the study area. The representativeness of the samples in terms of total number of polygons, pixels, and areas (in hectares) for each year is shown in Table 4. The samples were used as ground truth in the RF and SVM classifications.
Remote Sens. 2020, 12, x FOR PEER REVIEW 10 of 30 remaining areas assigned a value equal to one, producing a binary mask: water/nonwater. During this process the areas with a value equal to one were considered in the calculation, whereas the areas with a value equal to zero were discarded [66]. Then, the mask was applied to all polarimetric decompositions and to the full-pol images calibrated into σ°.

LULC Classes
In order to classify the LULC classes, we used reference samples obtained in the field campaigns, as previously described in Section 2.2. Each field sample was georeferenced based on the GPS coordinates. To reduce the speckle effect that is inherent to SAR data, each sampling point was converted to polygons through the buffer function available in the QGIS software. After some tests, the buffer zone of 24.75 m was selected, which is equivalent to an area of ~1800 m² for each polygon. The criterion used here was the visual analysis of the sample spatial distribution in our study area. High spatial resolution images from Google Earth™ platform (12/30/2014 and 30/12/2015) were used to support the inspection of each polygon. Figure 4 shows the location of the polygon samples over nine LULC classes found in the study area. The representativeness of the samples in terms of total number of polygons, pixels, and areas (in hectares) for each year is shown in Table 4. The samples were used as ground truth in the RF and SVM classifications.   Table 1.

Optimum Subgroup Definition
To select the variables that better discriminate the LULC classes, subgroups of different polarimetric attributes and backscattering coefficients were formed and were named as optimum subgroups. To obtain these subgroups, we carried out for 2015 and 2016 separately all steps presented below. The selection of these optimum subgroups, the Principal Component Analysis (PCA), was initially applied in order to reduce the dimensionality of the datasets (see Section 2.6). Sequentially, the most important variables derived from the PCA were selected and used as input data set in the RF classification (see Section 2.7). To compose the training and validation set of the model, we selected 75% of samples for training and 25% for validation. During the RF classification process, the algorithm calculated the features based on the importance of each variable. In this study, we selected the first four most important variables. The variable importance shows the interaction between the variables and hierarchizes them within a level of contribution and importance for the classification [67]. The same first four variables based on variable importance selected in RF were also used for SVM classifier. It is important to highlight that, as well as the PCA and the classification in the RF for each year that were performed separately; the SVM classification also was carried out for each year separately. More details on PCA and RF and SVM classification steps are described in the next subsections.

Principal Component Analysis
PCA is considered an important statistical technique that enables reducing the data dimensionality [68,69], extracting the components accounting for as much variance as possible through a linear transformation [70]. In this context, to form an optimum subgroup for the LULC classification, we applied the PCA technique to reduce the number of variables from the polarimetric decompositions and σ • data sets. We used the first principal component (PC1) and the second principal component (PC2) only to aggregate information that was more relevant to the classification process. For these two principal components, we selected only those variables whose contributions were higher than 5%. In addition, similar variables derived from different polarimetric decompositions were excluded, for example, double-bounce scatterings derived from van Zyl, Freeman-Durden, and Yamaguchi decompositions. A total of 17 input parameters (13 from polarimetric decompositions and 4 from σ • ) were selected, being 9 variables from 2015 and 8 variables from 2016 images. The PCA technique was performed in the R 1.7.3 software [71].

Classification and Evaluation
All seven groups of attributes (CP, VZ, FD, and YH polarimetric decompositions, RVI, PI backscattering indexes, and optimum subgroups) from both years were used as input data in RF Remote Sens. 2020, 12, 3512 12 of 30 and SVM supervised classification algorithms (Tables 5 and 6). These algorithms were implemented in R software and received increasing attention, due to their accurate classification results and their processing performance [72,73]. Table 5. Performance of the optimum subgroup of attributes in Random Forest (RF) and Support Vector Machine (SVM) classifiers. Attribute identification: H (entropy), α (alpha angle), and A (anisotropy) from Cloude-Pottier; P v (volumetric scattering), and P s (surface scattering) from Freeman-Durden; VZ d (double-bounce scattering) from van Zyl, and BMI (Biomass Index) from Pope indexes. Details of the SAR derived attributes are presented in Table 3. RF and SVM support data with different statistical distributions and can support several input parameters, providing consistent results [67,74]. In SAR data applications, they are less affected by noisy data, showing good classification results involving LULC dynamics in forests [75][76][77]. RF algorithm consists of an ensemble of decision trees, where each tree contributes with just one vote for the assignment of the most frequent class to the input data set [78]. The final classification result is determined by the highest number of votes of all forest trees. The algorithm uses bagging (bootstrap) and random samples from training sets for tree building with replacement in the original training set [79,80]. The samples that are excluded from the bootstrap are called out-of-bag (OOB) samples. The OOB samples help to evaluate misclassification error and to estimate variable importance [78]. To estimate the importance of each predictor variable in the RF model, we used the mean decrease in accuracy (MDA) method. This method is calculated by the normalized difference between OOB accuracy of the original observations and the randomly permuted variables [81]. Higher scores of MDA indicate that variables are important predictors in the classification. There are two essential parameters in the RF model: the number of input variables in the random subset of variables at each node and the number of trees. The grown trees are not pruned to ensure lower diversity among individual trees and obtain low biased trees. RF can still measure the relative importance of each variable within the model. Each variable is ranked according to its contribution to the classification accuracy.

2015
In this study, to build the RF model, 75% of samples were used for training, and 25% for validation. A default value of the number of variables was adopted, which corresponds to the square root of the total number of input variables [78]. A total of 500 trees was considered. Each polarimetric decomposition and backscattering index was classified individually, following the parameter value mentioned above. Considering the optimum subgroup, the variables selected in PC1 and PC2 for each year were initially used as input data in the RF algorithm to perform the 2015 and 2016 classifications separately. Subsequently, these same variables were classified, and the variable importance of each one was calculated during the RF classification process. For each year, the first four variables that showed the highest contribution and importance for classifications were selected. In the second step, only these first four selected variables for each year were used for classifications, and their results, respectively, attributed to the optimal subgroup, 2015 and 2016.
The SVM algorithm is based on decision planes that determine decision boundaries in input space or high dimensional feature space [82]. An optimal hyperplane is defined as the linear decision function with the best separation among different training classes in order to maximize the distance among them [83]. The distance between the decision boundary and the closest data point (training data) determines the margin of the classifier. These borderline samples are called support vectors and correspond to the critical elements of the training samples, as they delimit the decision surface [84]. When the data are not linearly separable, which is a common case in the real world, the SVM algorithm uses kernel functions. The original input space is mapped in a higher dimensional feature space, allowing the adjustment of the linear hyperplane and improving the linear separability [85]. The most common kernel functions are linear, polynomial, sigmoid, and radial basis function (RBF).
In this study, to build the SVM model, 75% of samples for training, and 25% for validation were also used. The selected kernel was the RBF, making it necessary to adjust cost (C) and gamma (γ) parameters. C parameter is known as a regularization term that controls the tradeoff between maximization of the margin width and the fitting of the training data set as well as the width γ of the RBF kernel [86]. In order to find the optimum C and γ parameters, we applied the cross-validation method. This method allows obtaining an error estimate in the data sets, randomizing them before building the splits of support vectors [87]. To find the average error, we used 10-fold cross-validation using the training set and the best parameter combinations.
The processing times (training time plus classification time, unit in s) for RF and SVM were also recorded for the purpose of analyzing whether the different polarimetric decompositions and backscatter coefficients presented different runtime performances in different classification algorithms.
The computer processor used in the classification step was equipped with Intel ® Core™ i5 2450 CPU, 2.50 GHz, and 6.0 GB usable RAM.
The classification results from both years were analyzed through the Kappa index, OA, and the confidence interval at the significance level of 95% (CI.95%), from RF and SVM classifications. For the best classification results, we also calculated the producer's accuracy and user's accuracy as well as F1-scores.

Optimum Subgroup Classification
In order to compose the optimum subset, we initially selected the most important variables derived from PC1 and PC2. For 2015, the variations explained by PC1 and PC2 were 51.3% and 13.0%, respectively (total accumulated variance of 64.3%). In 2016, the variations in PC1 and PC2 were 50.5% and 17.1%, respectively (total accumulated variance of 67.6%). Figure 5 shows  algorithms. The computer processor used in the classification step was equipped with Intel ® Core™ i5 2450 CPU, 2.50 GHz, and 6.0 GB usable RAM. The classification results from both years were analyzed through the Kappa index, OA, and the confidence interval at the significance level of 95% (CI.95%), from RF and SVM classifications. For the best classification results, we also calculated the producer's accuracy and user's accuracy as well as F1-scores.

Optimum Subgroup Classification
In order to compose the optimum subset, we initially selected the most important variables derived from PC1 and PC2. For 2015, the variations explained by PC1 and PC2 were 51.3% and 13.0%, respectively (total accumulated variance of 64.3%). In 2016, the variations in PC1 and PC2 were 50.5% and 17.1%, respectively (total accumulated variance of 67.6%). Figure 5 shows the percentage of each variable contribution for PC1 and PC2. These variables are considered the most important ones among all analyzed for 2015 and 2016. We highlight being for VZd variable, which presented the highest contribution (23.3%) for 2015. For 2016, the Yd variable showed the highest contribution (25.3%).  Table 3.
It is important to mention that the principal components of each year are formed by different variables, as shown in Figure 5. The distinct performance between 2015 and 2016, with different percentages of each variable contribution, may occur because: (a) some classes selected in this study presented high temporal and spatial dynamics (mainly PP, WP, CR, and BS). Such dynamics were also clearly observed in the DG class, which in 2016 showed high dynamics when compared with 2015, where the land cover was degraded by fire; (b) the acquisition of ALOS/PALSAR-2 images at different times of the year can influence the backscattering of the targets due to seasonal characteristics; and (c) the precipitation in the study area during some of the SAR acquisitions ( Table  2) can directly influence the amount of moisture in the dielectric properties of materials and their SAR backscatter [88].
Afterward, the variables presented in Figure 5 for 2015 were used as input data in the RF algorithm to classify the LULC classes for 2015. The four first variables for 2015 were: Pv, α, VZd, and A ( Figure 6). In the same way, the variables obtained for 2016 ( Figure 5) were used in RF classification for 2016. The four first variables for 2016 were: H, BMI, Ps, and A. Those aforementioned variables proved to be the most important in predicting to discriminate different forest types, forest degradation, and land use classes ( Figure 6).  Table 3.
It is important to mention that the principal components of each year are formed by different variables, as shown in Figure 5. The distinct performance between 2015 and 2016, with different percentages of each variable contribution, may occur because: (a) some classes selected in this study presented high temporal and spatial dynamics (mainly PP, WP, CR, and BS). Such dynamics were also clearly observed in the DG class, which in 2016 showed high dynamics when compared with 2015, where the land cover was degraded by fire; (b) the acquisition of ALOS/PALSAR-2 images at different times of the year can influence the backscattering of the targets due to seasonal characteristics; and (c) the precipitation in the study area during some of the SAR acquisitions (Table 2) can directly influence the amount of moisture in the dielectric properties of materials and their SAR backscatter [88].
Afterward, the variables presented in Figure 5 for 2015 were used as input data in the RF algorithm to classify the LULC classes for 2015. The four first variables for 2015 were: P v , α, VZ d , and A ( Figure 6). In the same way, the variables obtained for 2016 ( Figure 5) were used in RF classification for 2016. The four first variables for 2016 were: H, BMI, P s , and A. Those aforementioned variables proved to be the most important in predicting to discriminate different forest types, forest degradation, and land use classes ( Figure 6).  Table 3.
The four most important variables of each year were again classified separately by the RF algorithm. These same variables were also classified separately by the SVM algorithm. Table 5 shows the OA, CI. 95%, Kappa index, and processing time for the optimum subgroups classified by the RF and SVM algorithms for each year. As can be observed in Table 5, RF and SVM demonstrated their capabilities to classify LULC classes. The best OA and Kappa indices were exhibited by RF, with an OA of 75.4% and 79.9%, and Kappa index of 0.68 and 0.76, for 2015 and 2016, respectively. The RF processing for 2015, was 393.6 s faster than that of the SVM, while for 2016, the SVM was 21.6 s faster than that of RF. Table 6 (Table 6).  Table 3.

LULC Classification
The four most important variables of each year were again classified separately by the RF algorithm. These same variables were also classified separately by the SVM algorithm. Table 5 shows the OA, CI. 95%, Kappa index, and processing time for the optimum subgroups classified by the RF and SVM algorithms for each year.
As can be observed in Table 5, RF and SVM demonstrated their capabilities to classify LULC classes. The best OA and Kappa indices were exhibited by RF, with an OA of 75.4% and 79.9%, and Kappa index of 0.68 and 0.76, for 2015 and 2016, respectively. The RF processing for 2015, was 393.6 s faster than that of the SVM, while for 2016, the SVM was 21.6 s faster than that of RF. Table 6 (Table 6). Figure 7 shows the performance of RF and SVM classifiers in relation to the OA and processing time of the polarimetric decompositions, σ • , and optimum subgroup. RF exhibited superior performances for the groups of attributes investigated in both years. On average, for 2015, SVM algorithm demanded more than 58% of the classification time in relation to RF and more than 46% for 2016. derived attributes are presented in Table 3. Figure 7 shows the performance of RF and SVM classifiers in relation to the OA and processing time of the polarimetric decompositions, σ°, and optimum subgroup. RF exhibited superior performances for the groups of attributes investigated in both years. On average, for 2015, SVM algorithm demanded more than 58% of the classification time in relation to RF and more than 46% for 2016. In summary, for 2015, YH decomposition presented an improvement of 16.7% for OA and 23.9% for Kappa index, with an additional processing time of more than 246 seconds, when compared to CP decomposition; an increase of 6.1% for OA and 8.7% for Kappa index, with an additional processing time of more than 301.7 s, when compared with VZ decomposition; and 4.1% for OA, 6.1% for Kappa index and additional processing time of 307.6 seconds, when compared with FD decomposition. Concerning PI index, YH decomposition showed an increase of 12.8% for OA, 18.4% for Kappa index, and additional processing time of 246.6 seconds. The gains in relation to RVI were 36.5% for OA and 46.8% for Kappa index. Here, we found the largest differences, indicating that RVI had the worst classification performance. On the other hand, RVI processing time was 494.6 seconds faster than that of YH decomposition. The smallest difference in performance was observed with the optimum subset (α, A, P v , and VZ d ), with increments of 1.4% for OA, 2.3% for Kappa index, and 125.1 seconds in processing time than that of the optimum subgroup (Tables 5 and 6 and Figure 7).

LULC Classification
In 2016, YH decomposition also showed an increase of 10.4% for OA and 12.9% for Kappa index, with a processing time shorter than 38.8 s, when compared with that of CP classification; 4.1% for OA, 5.3% for Kappa index, and 130.2 s slower than that of VZ decomposition; 27.0% for OA and 33.6% for Kappa index, with a processing time longer than 79.9 s, when compared to that of PI index; 7.7% for OA and 10.2% for Kappa index, with a processing time shorter than 231.1 s in relation to that of the optimum subset (H, A, P s , and BMI). RVI for 2016 also showed the worst classification performance when compared with that of YH decomposition, with an increase YH of 51.0% for OA and of 61.8% for Kappa index. However, YH was around 315.0 s slower in the processing time. In relation to FD decomposition, YH decomposition showed an increase of 3.1% for OA and of 3.6% for Kappa index, with a processing time higher than 68.5 s, indicating the smallest performance difference among the decompositions (Tables 5 and 6 and Figure 7).
Based on the values of the parameters shown in Tables 5 and 6, and aiming to obtain the attributes with the greatest potential to discriminate LULC classes, we selected YH decomposition classified by the RF algorithm considering both years (2015 and 2016).
According to the classification results from YH decomposition for 2015 (Figure 8), among the classes with more dense forest structures investigated, we highlighted the results for the PF class, with a PA of 93.4%, a UA of 81.5%, and a F1-score of 0.87, followed by SS2 class, with PA of 68.2%, UA of 64.4%, and F1-score of 0.66. The highest misclassification was observed for the SS3 class, being misclassified as PF and SS2. The SS3 presented PA of 17.8%, UA of 44.4%, and F1-score of 0.25. The DG class showed PA of 41.2%, UA of 53.8, and F1-score of 0.47 (Figure 8).
According to the classification results from YH decomposition for 2015 (Figure 8), among the classes with more dense forest structures investigated, we highlighted the results for the PF class, with a PA of 93.4%, a UA of 81.5%, and a F1-score of 0.87, followed by SS2 class, with PA of 68.2%, UA of 64.4%, and F1-score of 0.66. The highest misclassification was observed for the SS3 class, being misclassified as PF and SS2. The SS3 presented PA of 17.8%, UA of 44.4%, and F1-score of 0.25. The DG class showed PA of 41.2%, UA of 53.8, and F1-score of 0.47 (Figure 8).  Table 1.
Concerning the classes with less vegetation coverage, the best results (PA and UA > 74.0%, and F1-score ≥ 0.77) were found for PP, WP, and BS. On the other hand, CR showed the highest misclassification, with PA of 55.0%, UA of 71.0%, and F1-score of 0.62 (Figure 8). According to YH classification results for 2016 (Figure 9), among the investigated classes that presented a dense forest structure, PF exhibited one of the best accuracy results, with PA of 96.8%, UA of 78.3%, and F1-score of 0.87, followed by the DG class, with PA of 90.1%, UA of 91.7%, and F1-score of 0.91. SS1 showed adequate accuracy, with PA of 71.8%, UA of 63.6%, and F1-score of 0.67. As observed in 2015, SS3 class also showed the highest misclassification, especially in terms of the PA. Around 78.6% of the samples were misclassified and the main misclassification was associated with PF and SS2 classes.  Table 1.
Concerning the classes with less vegetation coverage, the best results (PA and UA > 74.0%, and F1-score ≥ 0.77) were found for PP, WP, and BS. On the other hand, CR showed the highest misclassification, with PA of 55.0%, UA of 71.0%, and F1-score of 0.62 ( Figure 8). According to YH classification results for 2016 (Figure 9), among the investigated classes that presented a dense forest structure, PF exhibited one of the best accuracy results, with PA of 96.8%, UA of 78.3%, and F1-score of 0.87, followed by the DG class, with PA of 90.1%, UA of 91.7%, and F1-score of 0.91. SS1 showed adequate accuracy, with PA of 71.8%, UA of 63.6%, and F1-score of 0.67. As observed in 2015, SS3 class also showed the highest misclassification, especially in terms of the PA. Around 78.6% of the samples were misclassified and the main misclassification was associated with PF and SS2 classes. The classes with low levels of vegetation (PP, WP, CR, and BS) showed adequate accuracy (PA and UA > 77.0%, F1-score ≥ 0.78) (Figure 9). The obtained results demonstrated high reliability for each discriminated class, as well as high sensitivity of the classifier to correctly distinguish these classes from the others.
The LULC and forest degradation classification for 2015 and 2016 were produced using the Ys, Yd, Yv, and YH attributes from the YH decomposition ( Figure 10). The subsets in Figure 11b,c emphasize the areas with the most predominant of land use types in the study area. PP, WP, CR, and BS exhibited high dynamics in terms of areas occupied between 2015 and 2016. Concerning land cover classes (Figure 12), between 2015 and 2016, the main changes observed are related to an increase of areas occupied by DG and a reduction of areas occupied mostly by primary forest and secondary succession classes, where part of the TNF and surroundings has been degraded by fire, because of The classes with low levels of vegetation (PP, WP, CR, and BS) showed adequate accuracy (PA and UA > 77.0%, F1-score ≥ 0.78) (Figure 9). The obtained results demonstrated high reliability for each discriminated class, as well as high sensitivity of the classifier to correctly distinguish these classes from the others.
The LULC and forest degradation classification for 2015 and 2016 were produced using the Y s , Y d , Y v , and Y H attributes from the YH decomposition ( Figure 10). The subsets in Figure 11b,c emphasize the areas with the most predominant of land use types in the study area. PP, WP, CR, and BS exhibited high dynamics in terms of areas occupied between 2015 and 2016. Concerning land cover classes (Figure 12), between 2015 and 2016, the main changes observed are related to an increase of areas occupied by DG and a reduction of areas occupied mostly by primary forest and secondary succession classes, where part of the TNF and surroundings has been degraded by fire, because of the occurrence of forest fires events at the end of 2015 and the beginning of 2016.
The changes that occurred in LULC and forest degradation between 2015 and 2016 are only based on qualitative analysis (i.e., a visual analysis) of the YH decomposition classification results, which are clearly noticeable in Figures 10-12. A detailed analysis to quantify the losses and gains requires additional change detection techniques that are suggested for future studies. In addition, some thematic classes in the study area such as the cropland, well managed pasture, and bare soil, show high dynamics over the year, which could introduce false changes.    Table 1.   Table 1.

Discussion
The attributes extracted from the YH decomposition showed the best classification results (Table 5), demonstrating an adequate capability to discriminate different land use types, forest degradation, and land cover classes. The potential to discriminate classes with low levels of vegetation coverage (PP, WP, CR, and BS) was highlighted (Figures 8 and 9). The only exception occurs for the CR class, with PA of 55.0% in 2015 formed by crops with different phenological stages. This class showed the lowest classification performance and was therefore an exception. Classes with low vegetation coverage levels, like CR, can be misclassified due to the flat surfaces that show low levels of backscattering of transmitted signals in L-band, consequently, they cannot be easily distinguished [89]. In addition, some agricultural crops in more advanced stages of development can be easily distinguished from flat surfaces, as mature crops cause more volume scattering.
Some misclassifications that occurred in PP and WP classes can be explained by the presence of inajás (Attalea maripa) and babaçus (Attalea speciosa) [35]. They are pioneer trees species sparsely distributed in both poorly and well managed pastures. The structure of these tree species may cause certain confusion during the classification process, mainly when associated with other forested classes.
Concerning classes that present a dense forest structure, the PF was the class that exhibited the highest classification results in 2015 and 2016 (Figures 8 and 9), highlighting the great capability of YH decomposition to discriminate primary forest in rainforest environments. However, these attributes were not effective in discriminating SS3 class, especially the misclassification associated with the PA in both years. SS2 also showed intermediate values of PA and UA in both years. Misclassifications between SS2 and SS3 probably occurred due to similarities in their respective structural physiognomic characteristics and the high levels of regenerative development. SS1 also showed intermediate accuracies in both years. In this ecological successional stage, the forest class structure is composed predominantly of only one homogeneous superior stratum and one inferior stratum. The latter is composed of regeneration in its initial succession stages, which are arranged in an irregular form and dispersed in the environment, being influenced mainly by the distribution of pioneer species. Such condition probably was responsible for the misclassification to SS2 and PP classes.  (Table 4). In addition, the degraded forest samples collected during the field campaign in 2015 correspond to forests severely affected by fires and by unsustainable logging, while the samples collected in 2016 corresponded mostly to forests severely affected by fires. In this context, the most homogeneous samples in the DG class were in 2016, which consequently may have contributed to reach better results, demonstrating the capability of the YH attributes to classify forest degradation, mostly degraded by fire ( Figure 12).
In general, the intermediate values of accuracy for SS2 and SS1, and the lowest values of accuracy obtained for SS3 in both years, demonstrate that the YH decomposition has a greater capacity to discriminate ecological succession classes from those with less vegetation coverage, despite its poor ability in distinguishing among classes of different ecological succession and disturbance stages, with the exception of PF class. Ullmann et al. [90] examined the scattering characteristics from polarimetric data obtained at X-, C-, and L-band in a tundra-dominant ecosystem and found that L-band data were more appropriate to differentiate classes with low levels of vegetation cover. Pavanelli et al. [39] integrated data from Landsat-8/OLI and ALOS/PALSAR-2 satellites and classified them in the RF algorithm, to map the LULC classes in a fragment located in the northern part of the Brazilian Amazon. They found that ALOS/PALSAR-2 data's most important contribution was the possibility to discriminate classes with low levels of biomass (grasslands and wooded savanna).
According to the classification results from the optimum subset of attributes for 2015 (α, A, P v , and VZ d ) and 2016 (H, A, P s , and BMI) ( Table 5), these groups also showed a satisfactory capability to discriminate LULC classes. P v and H attributes are more sensitive to volumetric scattering and are mainly associated with dense forest classes. Thus, both attributes can be associated with a higher scattering volume of SAR signals in more dense forest structures. Trisasongko [91] used the CP decomposition to monitor disturbances in a dense tropical forest in Papua Tengah Province, Indonesia. This author associated high H values with intact forest classes, indicating a volume scattering mechanism's dominance. Silva et al. [92] modeled and estimated forest biomass from ALOS/PALSAR-1 polarimetric data in from the TNF and found a significant P v attribute contribution in primary and secondary forests. According to the results obtained by Kuplich et al. [93], P v component also presented the highest contribution in the discrimination of LULC classes of the TNF and its surroundings. In relation to classes with dense forest coverage, a greater contribution of BMI index in SAR signal response is expected since this parameter is related to the quantity of AGB. This index tends to be more sensitive in areas of dense forests, which is the case of the TNF.
Concerning A and VZ d attributes, Tanase et al. [94] found a better correlation with severe burning levels in forests affected by fire in a Mediterranean region for C-band data acquired at steep incidences. Martins et al. [38] and Plank et al. [95] used full polarimetric data from ALOS/PALSAR, obtaining high A values in forested areas severely affected by fire because of an increase of canopy openings and a decrease of AGB.
The P s attribute is more sensitive to surface scattering mechanism and can be associated with areas with less vegetation coverage or nonforest areas, such as bare soil and water bodies. Qi et al. [96] observed that the dominance of surface scattering extracted from the FD decomposition, which helped distinguishing between forest and barren/sparsely vegetated classes. In the context of the present study, the P s attribute can be associated mainly with bare soil and croplands, especially at the beginning of the crop growing cycle. The plants are starting to emerge above the surface; and consequently, there is a higher influence of soil surface in the backscattered SAR signals.
In this study, the YH polarimetric decomposition classified by the RF algorithm showed the best results in terms of OA and Kappa index for most of the thematic classes, including forest degradation class, in both years. In this context, Varghese et al. [32] analyzed different polarimetric decompositions to classify forest canopy density, based on the full polarimetric RADARSAT-2 data classified by the SVM algorithm. YH decomposition also showed the best results, followed by VZ, FD, and CP decompositions. Classes with low levels of vegetation coverage, such as cropland and fallows, showed good classification results. In this context, both classes that also were analyzed in our study exhibited similar classification results (Figures 8 and 9). Avtar et al. [89] also used YH, FD, and CP polarimetric decompositions to monitor land cover in a tropical region of Cambodia, based on ALOS/PALSAR-1 full polarimetric data. The authors used the Maximum Likelihood algorithm to classify the decomposition attributes, obtaining the highest value of accuracy, in terms of OA and Kappa index, for the YH decomposition, followed by FD and CP decompositions. Both studies reported above corroborate the results found in our research.
However, some studies presented different classification results. Middinti et al. [97] used ALOS/PALSAR full polarimetric data to classify different forest types from northeastern India. These authors extracted different attributes derived from backscatter coefficients and polarimetric decompositions. The results found for SVM (CP, FD, YH, and VZ decompositions) exhibited nearly similar classification accuracies, with intermediate values. Different to results presented by these authors, the classification results found in our study showed that YH and FD decompositions obtained high values of accuracy. Pôssa et al. [35] also analyzed the different polarimetric decompositions in order to map LULC classes from TNF. CP decomposition showed the best classification results (OA of 77.37% and Kappa index of 0.70), followed by FD and YH decompositions. Mirelva and Nagasawa [98] analyzed FD and YH decompositions, in addition to the backscatter intensities separately, and integrated them into ALOS/PALSAR-1 full polarimetric images for agriculture croplands classification in Indonesia. In their research, the FD and YH decompositions exhibited intermediate overall classification accuracy. Meanwhile, the best classification results were obtained in the integration of backscatter intensities (HH, HV, VH, VV, and HH+HV) and FD decomposition.
The classification results obtained by VZ, FD, and YH polarimetric decompositions and by the optimum subset demonstrated better classification performances in comparison with the classification based on backscatter coefficients, such as RVI and Pope index (Tables 5 and 6). A performance investigation of the two advanced machine learning methods (RF and SVM), based on classification accuracy, demonstrated that both algorithms showed sensitivity to discriminate different LULC classes and forest degradation. RF presented the best computational efficiency (Figure 7). The higher processing time in YH and optimal subset classifications for 2015 and 2016 may be related to the larger number of components (4), compared with the other groups (3 and 1).
Among the classes with different ecological successions (Figure 12), the most noteworthy class observed from 2015 to 2016 was the DG class. Unusual relative low air humidity, high temperature, and shortage of rainfall events were reported by MODIS products [50,51]. A set of 91 heat points were detected inside the TNF and adjacent areas. In this context, forest degradation, deforestation, and expansion of agricultural and livestock frontiers often cause negative environmental impacts.
In order to mitigate these negative impacts, we recommend an effective monitoring and surveillance program in the TNF region.
In this study, only full polarimetric data from ALOS/PALSAR-2 were explored. However, further studies are recommended adding full polarimetric SAR data obtained at different frequencies (currently available from the RADARSAT-2 and the experimental mode TERRASAR-X satellites). Deep learning methods are also recommended to extract more information from targets, thus improving the classification accuracy.

Conclusions
The polarimetric attributes extracted from ALOS/PALSAR-2 imagery and classified by RF and SVM algorithms exhibited a good ability to discriminate different types of tropical forests. In terms of OA, Kappa index, and processing time, the RF algorithm, in general, presented the best performance, when compared with that of SVM. YH decomposition, classified by the RF algorithm, presented the best performance concerning LULC and forest degradation classification. The thematic classes with less vegetation coverage were better discriminated than more dense forest structure classes, except for the PF. YH decomposition showed a strong capability of discrimination, with accuracies above 78.0%. However, in terms of processing time, the four components extracted from YH demanded the longest processing times.
The optimum subset of attributes for 2015 (α, A, P v , and VZ d ) and for 2016 (H, A, P s , and BMI) and FD and VZ decompositions also showed satisfactory classification results, indicating an adequate capability to discriminate different types of LULC present in rainforests. These polarimetric decompositions involving three attributes showed the shortest processing time performance compared with that of the optimum subgroup. The polarimetric decompositions showed a better performance than the backscatter coefficients represented by RVI and Pope index. During the period 2015-2016, there was a loss of forest classes due to forest degradation caused by fire events. In this context, the classification based on YH decomposition involving the ALOS/PALSAR-2 full polarimetric images demonstrated the capability to discriminate burned areas in rainforests. Furthermore, forest degradation by fire and the conversion of forest into pasture or cropland was also noticed, demanding initiatives to monitor LULC changes in such an important conservation unit. Future research employing deep learning methods, besides a synergistic approach involving other frequency SAR data, is recommended.
Further research for verifying the impact of the time difference between field data collection and SAR data acquisitions on the LULC changes are strongly recommended. Seasonality effects and temporal changes can induce some mislabeled samples in classification models, mainly when collected in land use classes with high temporal and spatial dynamics. BMI can be used as an indirect indicator of the relative quantity of wood, through leaf biomass (Equation (A13)). Lower BMI values are generally associated with environments dominated by the perennial forests, due to the absorption of microwave EM radiation by the green leaves [63].
The VSI is related to the density of the vegetation canopy, providing a measure of the backscatter depolarization, through the number of elements that generate multiple spreading, such as the cylindrical elements represented by the branches and trunks of the trees (Equation (A14)).
The CSI can be used to indicate the thickness or density of the treetops. This index indicates the importance of structures arranged in vertical orientation in relation to structures in horizontal. Forests that have a predominance of vertically arranged trunks and branches tend to have high CSI values.
Lower CSI values are expected in environments of forests dominated by horizontal branches, where there is a greater sensitivity to surface spreading [63] (Equation (A15)).