Data-Driven Performance Evaluation of Geometric Clustering for PolSAR Data Analysis

We have introduced in [1] a method for unsupervised classification of PolSAR data, on the manifold of Hermitian positive definite matrices obtained from the polar decomposition. In this paper we investigate the polarimetric information preservation of the Hermitian factor using manifold gradient computation. We provide an algorithm to select the optimum number of classes based on the Calinski-Harabasz criterion in the Riemannian geometry context.