A Non-Parametric Texture Descriptor for Polarimetric SAR Data with Applications to Supervised Classification

Abstract

The paper describes a novel representation of polarimetric SAR (PolSAR) data that is inherently non-parametric and therefore particularly suited for characterising data in which the commonly adopted hypothesis of Gaussian backscatter is not appropriate. The descriptor is also non-local and can capture image structure in terms of the arrangement of edge-, ridge- and point-like features, to yield a salient characerisation of semi-periodic spatial patterns. The basic approach is based closely on [1] and has been adapted for application to PolSAR data. As an example application, the descriptor is evaluated in the context of supervised classification. The performance is compared with conventional statistical approaches on both simulated and real PolSAR data