Unsupervised Classification of Polarimetric SAR Images Using Bayesian Nonparametric Model and Markov Random Field

A Bayesian Nonparametric (BNP) approach such as the Dirichlet process mixture model (DPMM) is a potential method for classification tasks. It provides a framework for estimating both the number of components in a mixture model and the parameters of the individual mixture components simultaneously from PolSAR data. The Markov random field (MRF) is commonly used for incorporating spatial correlation of image structure and has been widely applied to the classification of remote sensing images. In this paper, by combining the DPMM model and the MRF framework, a novel unsupervised classification method for multilook Polarimetric Synthetic Aperture Radar (PolSAR) images is developed. The prior distribution of the class label is considered as the product of a sitewise term and an interaction term. The expectation–maximization/maximization of the posterior marginals (EM/MPM) algorithm is used to update the class labels and the parameters of the proposed DPMM-MRF model for high resolution multilook PolSAR images. The performance of the proposed method EM-DPMM-MRF is evaluated with PolSAR images and the preliminary results on classification are satisfactory and are encouraging.