Unsupervised classification for PolSAR images based on multi-level feature extraction

With the development of remote sensing systems, the scale of the imaging data grows rapidly, which highly requires appropriate adaptability for interpretation algorithms. Focusing on this trend, an unsupervised classification algorithm for polarimetric synthetic aperture radar (PolSAR) images is proposed based on multi-level feature extraction. The algorithm firstly generates an initial classification map by multi-level feature extraction, and then introduces Wishart classifier into the iterative classification to refine the initial. At the first level, the PolSAR image is classified into four categories by combining entropy and anisotropy features that are extracted from Cloude-Pottier decomposition. From the scattering mechanisms, the second-level classification is conducted with the surface, double-bounce and volume scattering power obtained from three-component decompression. Accordingly, the PolSAR image is further divided into 13 categories. Finally, to discriminate objects with similar polarimetric characteristics but different scattering power, the total scattering power is adopted to classify the PolSAR image into 26 categories at the third level. Experiments on some real PolSAR images acquired by AIRSAR system demonstrate the effectiveness of the proposed method both qualitatively and quantitatively.