High-resolution tropical forest mapping of the Amazon basin: a novel classification approach for the GRFM radar mosaic

The high resolution (100 m) Global Rain Forest Mapping (GRFM) radar mosaics, providing a spatially continuous coverage of entire ecosystems, pave the way to improved estimates of bio-physical parameters related to the tropical vegetation. A new classification scheme for producing a high-resolution regional scale forest/non-forest thematic map of the Amazon is proposed. First, a new wavelet multi-resolution decomposition/reconstruction technique is employed to generate an edge-preserving piecewise constant approximation of the original radar image. Second, a two-stage hybrid learning Nearest Multiple-Prototype (NMP) classifier is applied to the reconstructed radar image. The NMP first stage employs a near-optimal vector quantization algorithm called Enhanced Linde-Buzo-Gray (ELBG). During the training phase, ELBG employs only 1% of the whole data set. At the second stage of NMP, vector prototypes are combined into land cover classes of interest by an expert photo-interpreter. In the pattern recognition phase, each pixel is labeled according to the minimum-distance-to-prototype criterion. Experimental results are reported for a thematic problem involving classes: primary forest, degraded forest, non-forest, and water bodies. Validation is performed using land cover maps provided by the Tropical Rain Forest Information Center (TRIFIC) over three test sites featuring different forest cover disturbance patterns. Main results are that the proposed classifier (i) provides a classification accuracy of 87% in forest/non-forest mapping, (ii) is capable of generalizing over the entire data set, and (iii) requires minor user interaction. It is concluded that the proposed approach responds adequately to the requirements of regional scale high-resolution vegetation mapping