Improvement of Markov Random Field Technique for Unsupervised Change Detection from Multichannel SAR Images

In applications related to environmental monitoring and disaster management, multichannel synthetic aperture radar (SAR) data present a great potential, owing both to their insensitivity to atmospheric and Sun-illumination conditions and to the improved discrimination capability they may provide as compared with single-channel SAR. However, exploiting this potential requires accurate and automatic techniques to generate change maps from images acquired over the same geographic region in different polarizations or at different frequencies at different times. Furthermore, sensitivity to contextual information of each pixel reduces the error rates in labeling process, thus generates accurate change maps. The smoothing effect of despeckling and the isotropic formulation of the Markov Random Field model cause over-smoothing of the spatial boundaries between changed and unchanged areas in the final change maps. In order to reduce this drawback, edge-preserving MRF models could be integrated in the labeling process. This method improves the precision of edges at spatial boundaries and increases the change detection accuracy. In this paper, a contextual unsupervised change-detection technique (based on a data-fusion approach) is proposed for two-date multichannel SAR images. A Markov Random Field model is formulated by using “energy functions” that combines the information conveyed by each SAR channel, the spatial contextual information concerning the correlation among neighboring pixels and the edge information. In order to estimate the model parameters, the expectation–maximization algorithm is combined with the recently proposed “method of log-cumulants.” The proposed technique was experimentally validated with semisimulated data produced by ASAR-ENVISAT images. Experiments illustrate a significant improvement (average 12%) with the proposed technique over the other change detection approaches. Integrating edge information yielded accurate results in exploiting various levels of changes (low-medium-high) whereas contextual information and information conveyed by channels were unable to detect low and medium level changes. Considering the small number of iterations, computation time is reduced considerably. Generally the highest accuracy achieved by the proposed algorithm is 99/67%.