Adaptive Model-Based Classification of Polarimetric SAR Image

Abstract

An adaptive classification is developed as a hybrid of the eigenvector- and model-based decompositions of polarimetric SAR (PolSAR) image. It adopts the canonical models that widely used in model-based target decomposition to obtain an improvement for the well-known $H/\alpha$ classification. First, a correspondence principle is developed to adaptively select the matched canonical models. The models are parallelly combined in terms of the scattering similarity for a fine description of the scattering mechanism then. Twelve classes are finally achieved with each one carrying a unique symbol to indicate a specific scattering. The classification does not depend on a particular data set, avoids the hard partitioning, and solves the obscures in $H/\alpha$. Comparison on real PolSAR image with $H/\alpha$ validates the better discrimination of radar targets.