Unsupervised Classification of Polarimetric SAR Image Based on Geodesic Distance and Non-Gaussian Distribution Feature

Polarimetric synthetic aperture radar (PolSAR) image classification plays a significant role in PolSAR image interpretation. This letter presents a novel unsupervised classification method for PolSAR images based on the geodesic distance and K-Wishart distribution. The geodesic distance is obtained between the Kennaugh matrices of the observed target and canonical targets, and it is further utilized to define scattering similarity. According to the maximum scattering similarity, initial segmentation is produced, and the image is divided into three main categories: surface scattering, double-bounce scattering, and random volume scattering. Then, using the shape parameter α of K-distribution, each scattering category is further divided into three sub-categories with different degrees of heterogeneity. Finally, the K-Wishart maximum likelihood classifier is applied iteratively to update the results and improve the classification accuracy. Experiments are carried out on three real PolSAR images, including L-band AIRSAR, L-band ESAR, and C-band GaoFen-3 datasets, containing different resolutions and various terrain types. Compared with four other classic and recently developed methods, the final classification results demonstrate the effectiveness and superiority of the proposed method.