An hierarchical approach for model-based classification of SAR images

We propose an unsupervised classification algorithm for high resolution Synthetic Aperture Radar (SAR) images based on Classification Expectation-Maximization (CEM). We combine the CEM algorithm with the hierarchical agglomeration strategy and a model order selection criterion called Integrated Completed Likelihood (ICL) to get rid of the initialization and the model order selection problems of the EM algorithm. We exploit a mixture of Nakagami densities for amplitudes and a Multinomial Logistic (MnL) latent model for class labels to obtain spatially smooth class segments. We test our algorithm on TerraSAR-X data.