Abstract
In this paper, we formulate the problem of change detection in multitemporal Synthetic Aperture Radar (SAR) images as unsupervised learning of finite mixtures. In particular, the statistical parameters related to changed and unchanged classes in the log-ratio image are estimated in an explicit way by the Expectation-Maximization (EM) algorithm. Then, we propose to identify automatically the number of changes present in the log-ratio image using the Bayesian Information Criterion (BIC). Finally, in order to increase the accuracy of the change-detection map we perform the analysis of the log-ratio image using Markov Random Fields (MRF) that exploits the interpixel class dependency in the spatial domain according to the use of a regularization term. The experimental results obtained on multitemporal SAR images characterized by a single change confirmed the effectiveness of the proposed method.