TEXTURE BASED POLARIMETRIC SYNTHETIC APERTURE RADAR IMAGE CLASSIFICATION USING COVARIANCE MATRICES

Abstract

This paper proposes a workflow for polarimetric SAR (PolSAR) image classification based on statistical texture descriptors. The methodology presented in this paper involves spatial interdependence between neighboring pixels as well as multiscale texture representation using wavelet decomposition. The collected features are modeled by zero-mean Multivariate Gaussian Distributions (MGDs). Then, their estimated covariance matrix acts as the texture descriptor and is employed in a k-nearest neighbors (kNN) classifier. Experiments using real PolSAR data validate the proposed approaches' accuracy in land cover categorization, showing their potential for reliable class identification.