A Novel PolSAR Image Classification Method Based on Optimal Polarimetric Features and Contextual Information

Abstract

Due to the severe speckle noise of a fully polarimetric synthetic aperture radar image and the complex backscattering mechanism at the junction of different land covers, some of the pixels are easily mislabeled, especially on the edge of the land covers. To address this issue, this study presents a novel scheme that selects polarimetric features step by step to participate in classification through different classification mechanisms. Different from previous classification methods where all land covers are described by the same polarimetric features, we make a fine selection of polarimetric features for each land cover with the help of information entropy and contextual information. Among many polarimetric features, elements of the covariance matrix and the coherence matrix are selected to optimize the original classification results. The experimental results show that the proposed method can achieve good classification results, especially in the edge area of land covers. Compared to traditional classification methods, the proposed method is robust and is able to improve the overall accuracy by more than 5.58% and the kappa coefficient by more than 0.0613.