A SAR Image Classification Method Based on Dempster-Shafer Theory and Markov Context with Parametric and Kernel Method Estimation

Abstract

With the raise of resolution of synthetic aperture radar (SAR) image, it is an urgent need to give a precise classification or interpretation of every scene in the high-resolution SAR images. Aiming at solving this problem, our work concentrates on two aspects: finding a more precise statistic model for SAR images, and proposing a novel classification method to provide a more precise classification or interpretation for high-resolution SAR images. Our approach is divided into three main steps, which build the whole image classification gradually: (1) Model finding and pre-classification step. First in this step, we introduce a non-parametric probability density estimate based on kernel method to describe the SAR image. Then two classification methods based on Markov Random Filed (MRF) model with determinate parametric probability distribution and non-parametric density estimate are respectively adopted to form a pre-classification of the high-resolution SAR images. (2) Fusion Step. We fuse the two pre-classification results by the Dempster-Shafer evidence theory in an unsupervised way, in order to eliminate the ambiguity between the pre-classification results. In our case, a decision rule, which maximizes the brief function over all hypotheses, is adopted. (3) Verification and regularization step. In this step, the classification results are verified and a novel iterated maximum selecting approach is proposed to control the fragments and modify the errors of the classification. Experimental results on real SAR images show that, the novel proposed method can provide a more precise classification to every scene of the high-resolution SAR image and illustrate rather an impressive performance.