Combined similarity-based spectral clustering ensemble for polSAR land cover classification

Abstract

This chapter proposes a novel spectral clustering ensemble method for unsupervised land cover classification of PolSAR data. This method increases the diversity to overcome the instability results caused by the random sampling during Nystro¨m approximation. Compared with the standard spectral clustering methods, the proposed scheme has the contributions in three aspects: firstly, during the process of spectral clustering, Wishart-derived distance measure and polarimetric similarity are combined to obtain the complementary information from the spatial and polarimetric relations between pairwise pixels. Secondly, a new similar function based on MRF potential function is used to construct the similarity matrix, which improves the robustness of spectral clustering to the scaling parameter. Finally, multiple individual classifications are obtained and integrated by an ensemble strategy. The experimental results demonstrate that the proposed method is superior to the compared methods.