Unsupervised Classification of Non-Gaussian Terrain Based on Polarimetric Decomposition

Abstract

Polarimetric synthetic aperture radar (PolSAR) can provide complete polarization property of terrain. Terrain classification is the most common application of PolSAR data. In this paper, an unsupervised clustering algorithm based on Freeman-Durden decomposition (FDD) and a non-Gaussian K-Wishart distribution classifier is proposed. This algorithm combines an advanced statistical distribution with spatial polarization scattering information of multi-looks PolSAR data. We use the prior probability characteristics of the Markov random field model to adaptively adjust the cluster center to make the classification more accurate. The experiment result shows that the proposed algorithm based on non-Gaussian models can better retain the polarization information of the target and the clustering accuracy was effectively improved on the real SAR images.