Exploring the Capability of ALOS PALSAR L-Band Fully Polarimetric Data for Land Cover Classification in Tropical Environments

Among different applications using synthetic aperture radar (SAR) data, land cover classification of rain forest areas has been investigated. Previous results showed that L-band is more appropriate for such applications. However, SAR images have limited discriminability for mapping large sets of classes compared with optical imagery. The objective of this study was to carry out an analysis about the discriminative capability of an L-band fully polarimetric SAR complex image, compared to the possible subsets of polarizations in amplitude/intensity, for mapping land cover classes in Amazon regions. Two case studies using ALOS PALSAR L-band fully polarimetric images over Brazilian Amazon regions were considered. Several thematic classes, organized into scenarios, were considered for each case study. These scenarios represent distinct classification tasks with variated complexities. Performing a simultaneous analysis of different scenarios is a distinct way to assess the discriminative capability offered by a particular image. A methodology to organize thematic classes into scenarios is proposed in this study. The maximum likelihood classifier (MLC), with specific distributions for SAR data, and support vector machine were considered in this study. The iterated conditional modes algorithm was adopted to incorporate the contextual information in both methods. Considering a kappa coefficient equal to 0.8 as an acceptable minimum, the experiments show that none subset of polarization or fully polarimetric image allows performing discrimination between forest and regeneration types; single-polarized HV data provide acceptable results when the classification problem deals with the discrimination of a few classes; depending on the classification scenario, the dual-polarized HH+HV image produces similar results when compared to multipolarized (i.e., HH+HV+VV) data; in turn, if the MLC method is adopted, multipolarized data may produce close or statistically indifferent classification results compared to those produced with the use of fully polarimetric data. (AU)