2.1.

Study Area and Data

The study area, $5\mathrm{km}\times 10\mathrm{km}$ in size, was in Siak River Transect, Riau Province, Indonesia (Fig. 1), a rapidly developing region where the peatland has been intensively converted mostly into oil palm plantations over the last two decades.¹⁴ Generally, this area has a flat topography with a surface elevation ranging from 2 to 10 m above sea level. The temperature of this area is 26.2°C per year on average. The annual rainfall of this area varies from 2200 to 2600 mm per year, with the lowest and highest monthly rainfall in July (around 110 mm per month) and November (around 290 mm per month), respectively.

Fig. 1

Map of Indonesia showing the location of the study area (hatched rectangle).

In this study, two scenes of PALSAR dual-polarization data and a single scene of PALSAR fully polarimetric data were used as the primary data. The dual-polarization data were taken during the driest month of the year (ALPSRP236980000, coded as dry scene) and the wettest month of the year (ALPSRP257110000, coded as wet scene), allowing the seasonal variation analysis of tropical peatlands from the viewpoint of L-band dual-polarization SAR. The single scene of fully polarimetric data (ALPSRP175860000) was taken during the average monthly rainfall.

Moreover, Landsat 5 Thematic Mapper (TM) data were used to update the existing land use/cover map by means of visual interpretation and to select the training samples for classification purposes. A total of 2400 points were derived for training (1680 points) the algorithm of decision tree classification and for testing (720 points) the accuracy of the classification results (Fig. 2). An existing peatland distribution map and a topographic map provided by the Indonesian Agency for Agricultural Research and Development (IAARD) were used as reference maps. In addition, a joint ground truth survey with the IAARD was conducted on May 19–22, 2015 to provide basic information about the study area. The list of data used for analysis is shown in Table 1.

Fig. 2

Landsat 5 image of the study area showing the spatial distribution of (a) training and (b) testing points.

Table 1

List of data used in this study.

2.2.

Data Preprocessing

PALSAR data were imported to the PolSARPro Software for image processing. Thus, these data were radiometrically calibrated and multilooked one time in range and five times in azimuth direction for the dual-polarization data, and one time in range and seven times in azimuth direction for the fully polarimetric data. Speckle noise was reduced by applying a $7\times 7$ Lee-refined filter, followed by conversions to the scattering matrix (S2) and a $2\times 2$ covariance matrix (C2) for the dual-polarization data, and to the S2, coherency matrix (T3) and a $3\times 3$ covariance matrix (C3) for the fully polarimetric data. These matrices were used as inputs for deriving polarimetric features. PALSAR data were then geocoded using the advanced spaceborne thermal emission and reflection radiometer (ASTER) global digital elevation model (GDEM) to the UTM Zone 48N map projection.

2.3.

Polarimetric Features for Dual-Polarization Data

Polarimetric features for dual-polarization data were derived after PDs and backscatter coefficients were obtained. The Cloude–Pottier decomposition,¹⁵ an eigenvalue/eigenvector-based decomposition, was performed for the dual-polarization data to derive three features, viz. entropy (coded as dFBD_H for the dry-scene data and wFBD_H for the wet-scene data), anisotropy (coded as dFBD_A for the dry-scene data and wFBD_A for the wet-scene data), and alpha angle (coded as $\mathrm{dFBD}\_\alpha$ for the dry-scene data and $\mathrm{wFBD}\_\alpha$ for the wet-scene data). Two features represent the channels of polarization derived from the dual-polarization data, i.e., HH (coded as dFBD_HH for the dry-scene data and wFBD_HH for the wet-scene data) and HV (coded as dFBD_HV for the dry-scene data and wFBD_HV for the wet-scene data). Each scene of the dual-polarization data (wet and dry scenes) derived five polarimetric features. Thus, a total of 10 polarimetric features were computed using the dual-polarization data for the analysis carried out in this study.

2.4.

Polarimetric Features for Fully Polarimetric Data

Polarimetric features for fully polarimetric data were derived after PDs, backscatter coefficients measurements, and the RVI were obtained. Three PDs—Cloude–Pottier decomposition,¹⁵ Freeman–Durden three-component decomposition,¹⁶ and Yamaguchi four-component decomposition¹⁷—were applied to generate polarimetric features for the fully polarimetric data. Cloude–Pottier decomposition was also performed for the fully polarimetric data and generated three features, namely, entropy (PLR_H), anisotropy (PLR_A), and the alpha angle ($\mathrm{PLR}\_\alpha$). The Freeman–Durden decomposition, a physical scattering model-based decomposition, models the covariance matrix (C3) as the contribution of three component scattering mechanisms—surface (f_odd), double-bounce (f_double), and volume (f_volume) scattering mechanisms. The Yamaguchi decomposition scheme incorporates the nonreflection symmetry condition of copolarization and cross-polarization radar channels thus the covariance matrix (C3) is modeled as four component scattering mechanisms’ contributions—surface (y_odd), double-bounce (y_double), volume (y_volume), and helix (y_helix) scattering mechanisms. Four features representing the channels of polarization were derived from the fully polarimetric data, i.e., HH (coded as PLR_HH), HV (coded as PLR_HV), VH (coded as PLR_VH), and VV (coded as PLR_VV).

Three features representing a decomposition approach based on the Pauli matrix were generated.^12,15 These features are HH+VV, HH-VV, and 2HV, associated with surface scattering, double-bounce scattering, and volume scattering, respectively. A depolarization ratio (HV/HH) was calculated that indicated the level of volume scattering.¹² In addition, the total scattering power (SPAN) in four polarizations of the fully polarimetric data was generated. The SPAN is an average of HH, HV+VH, and VV, and thus has a lower speckle noise than those polarization channels individually.¹⁸ Concurrently, the RVI was derived for quantifying the amount of biomass presented in each pixel.¹⁹ The equation for the RVI is

Table 2

List of polarimetric features derived for analysis.

2.5.

Image Classification

To evaluate the performance of PALSAR dual-polarization and fully polarimetric data in classifying tropical peatlands, polarimetric features derived by four combinations of data were used in this study to apply to the algorithm of decision tree classification: (i) the combination of the two scenes of dual-polarization data, (ii) the single scene of the fully polarimetric data, (iii) the combination of the two scenes of dual-polarization data and the single scene of the fully polarimetric data, and (iv) the combination of the two scenes of dual-polarization data, the single scene of the fully polarimetric data, and the additional feature “distance to river.” These combinations of data were applied in this study to assess the capabilities of each type of data and their combinations in classifying tropical peatlands. Moreover, the integration of the additional feature “distance to river” was meant to evaluate the effectiveness of geographic information system (GIS) data integration in the SAR-based tropical peatlands classification. The feature “distance to river” was derived from a GIS data processing, which buffered the center line of the main river stream into three categories, namely, “less than 2 km from river,” “2 to 5 km from river,” and “more than 5 km from river.” Previous studies^4,20,21 have found “distance to river” to be an important topographic feature in estimating the peat depth as well as in discriminating spatial distributions of tropical peatlands.

In this study, the decision tree classifier was used for classifying tropical peatlands. According to the concept of decision tree classification,²² the decision tree is a procedure of classification that recursively partitions a set of data into smaller subdivisions based on a set of criteria determined at each branch in the tree. The decision tree classifier was characterized as strictly nonparametric and required no-assumptions regarding the distributions of the input data. Moreover, the design of classification is explicit and easy to understand. The main advantages of the decision tree classifier are its ability to handle complex relations among class properties, its computational speed, and its capability to handle data represented on different measurement scales.²³ Thus, the decision tree is an efficient tool for land cover classification purposes.²⁴ Furthermore, the distance factor (DF) was used in this study to evaluate the effectiveness of a feature for separating the class pairs by calculating the distance between the different class mean values compared with the standard deviations. Hence, if the distance between the different class mean values is large compared with the standard deviations, classes are said to be well separated following the concept of feature separation.²⁵ The expression for the DF is

In this study, six classes were selected to represent the common land use/cover on the study area from the viewpoint of L-band SAR data. These classes are water, artificial, bareland with peat, bareland without peat, vegetation with peat, and vegetation without peat. Hence, to apply the concept of feature separation on the algorithm of the decision tree classification, five combinations of class pairs were specified: (1) (“water” plus “artificial”) and others (“bareland with peat,” “bareland without peat,” “vegetation with peat,” plus “vegetation without peat”), (2) “water” and “artificial,” (3) (“bareland with peat” plus “bareland without peat”) and (“vegetation with peat” plus “vegetation without peat”), (4) “bareland with peat” and “bareland without peat,” and (5) “vegetation with peat” and “vegetation without peat.” These class pairs were then applied to the algorithm of the decision tree classification to classify each land use/cover class, as shown in Fig. 3.

Fig. 3

The algorithm of the decision tree classification.

Thus, polarimetric features with a relatively high DF on each class pair were analyzed and applied to the algorithm for decision tree classification, then the classification results’ accuracy was assessed using the available testing points. Features that increase the accuracy of the classification results were then added to the algorithm and vice versa. The confusion matrix was used to perform the accuracy assessment on the classification results by means of deriving the accuracy indicators, namely, producer’s accuracy (PA), user’s accuracy (UA), overall accuracy (OA), and kappa coefficient (K). These indicators were used to assess the quality of the classification results.

3.1.

Class Pairs’ Separability

Table 3 shows the separability of five class pairs by each polarimetric feature, as measured by the DF. An italic value indicates the low performance of a feature in separating class pairs, whereas a bold value illustrates the average performance. A bold italic value indicates the excellent performance of a feature in separating class pairs that equates to a nearly complete separation of class pairs. The values of DF for all class pairs were quite varied, depending on the feature. In general, the DF derived by the fully polarimetric data tends to have a higher value than those derived by the dual-polarization data, particularly for class pairs 1 to 4. However, for the class pair 5, the DF derived by the fully polarimetric data was inferior to those derived by the dual-polarization data. In addition, the quality of DF for class pairs 4 and 5 generated by all polarimetric features was low. The low performance of DF occurred due to the highly overlapping mean and standard deviation values of land use/cover classes in the associated class pairs that corresponded to the stage of identifying the existence of peatland in a land use/cover class.

Table 3

Separability of five class pairs by each polarimetric feature.

3.2.

Selected Features for Classification

3.3.

Results of the Image Classification

As presented in Fig. 4, the results of tropical peatlands classification from four combinations of data were generated by means of the decision tree classifier. The classification results showed six land use/cover classes, i.e., water, artificial, bareland with peat, bareland without peat, vegetation with peat, and vegetation without peat. The “bareland with peat” and “vegetation with peat” classes represented land use/cover classes identified with the existence of peatland, whereby the “bareland without peat” and “vegetation without peat” classes represented land use/cover classes identified as having no peatland. The classification results of data combinations (i), (ii), (iii), and (iv) had 53.6%, 40.4%, 50.9%, and 32.3% of the total area identified as having peatland, respectively. The results of tropical peatlands classification derived by the dual-polarization data generated the biggest area identified as having peatland, whereas those derived by the combination of dual-polarization data, fully polarimetric data, and the additional feature “distance to river” produced the smallest area identified as having peatland. Thus, in this study, the dual-polarization data tended to generate more peatland area, whereas the full polarimetric data tended to generate less peatland area. In addition, the feature “distance to river” was added to the algorithm of decision tree classification to ignore the existence of peatland $<2\mathrm{km}$ from the river stream, so that it influenced the classification results of data combination (iv) producing the smallest area of peatland.

Fig. 4

Result of tropical peatlands classification for each combination of data: (a) data combination i, (b) data combination ii, (c) data combination iii, and (d) data combination iv.

3.4.

Accuracies of the Image Classification

Table 5 shows the accuracy indicators of tropical peatlands classification results from four combinations of data by means of the decision tree classifier. Accuracy indicators were low for the classification results derived by the dual-polarization data, which also yielded the lowest OA and K of 55.0% and 0.39, respectively. In contrast, the accuracy indicators for the classification results generated by the fully polarimetric data produced better accuracy indicators than those produced by the dual-polarization data; the fully polarimetric data gave accuracies of 65.0% and 0.51 for OA and K, respectively. Significant increases in OA and K were obtained by the classification results derived by the combination of dual-polarization and fully polarimetric data; this combination yielded an OA of 69.0% and a K of 0.57. Thus, accuracy indicators were increased by applying the combination of dual-polarization and fully polarimetric data to the algorithm of decision tree classification for the SAR-based tropical peatlands classification. This combination showed that the complementarity of dual-polarization and fully polarimetric data produced higher accuracy than the single usage of each type of data for the purpose of tropical peatlands classification by means of the decision tree classifier. In addition, the best accuracy was obtained by means of integrating the feature “distance to river” into the algorithm of decision tree classification. The accuracy indicators derived by the combination of dual-polarization data, fully polarimetric data, and the additional feature “distance to river” yielded the best OA and K of 76.0% and 0.66, respectively.

Table 5

Accuracy indicators for tropical peatlands classification for each combination of data.

The integration of GIS data in the SAR-based tropical peatlands classification was found to be effective in improving the classification accuracy, whereas the feature “distance to river” ignored the existence of peatland $<2\mathrm{km}$ calculated from the main river stream. The feature “distance to river” successfully aided the polarimetric features derived by dual-polarization and fully polarimetric data in classifying tropical peatlands and increased the accuracy indicators of the classification results. The category “$<2\mathrm{km}$” of the feature “distance to river” was revealed to represent the actual distance of peatland to exist as measured from the main river stream. Thus, the feature “distance to river” increased the classification accuracy when integrated with the polarimetric features derived by the combination of dual-polarization and fully polarimetric data by means of the decision tree classifier. The accuracy indicators derived by the combination of dual-polarization data, fully polarimetric data, and the additional feature “distance to river” were increased by nearly 10% for OA and 15% for K over the best obtained OA and K derived by the SAR-based tropical peatlands classification using the decision tree classifier.