Classification of Polarimetric SAR Image based on Improved Fuzzy Clustering

Abstract

This paper presents an improved fuzzy clustering approach for Polarimetric SAR image by incorporating neighborhood information. Firstly, polarimetric scattering characteristics of the terrain in PolSAR image are used to generate appropriate initial centers to avoid the issue that FCM is sensitive to random class centers. Then to further enhance the robustness to speckle noise, the conventional robust fuzzy C-mean clustering approach is improved. The work mainly exists in two aspects: (1) The revised Wishart distance is adopted as the data distance measure instead of Euclidean distance to assign a label to each pixel. (2) A weighted fuzzy membership is established by considering local spatial distance and class membership between the central pixel and its neighborhood simultaneously. Finally, the real polarimetric SAR data is utilized for the validation of the proposed unsupervised classification method. Experimental results demonstrate the superiority of the proposed method over the comparisons.