An Adaptive Technique based on Similarity Measures for Change Detection in Very High Resolution SAR Images

This paper presents a novel adaptive technique for change detection in very high geometrical resolution (VHR) Synthetic Aperture Radar (SAR) images that exploits information theoretical similarity measures for modeling the temporal evolution of probability density functions (pdfs). Image statistics for characterizing pdfs are adaptively estimated on a local basis by exploiting the spatial-context information of pixels on small homogeneous regions shared by multitemporal images (i.e., multitemporal "parcels"). The joint analysis of different orders statistics makes the method robust and suitable to the detection of both step changes of the backscattering and texture changes. The use of parcels allows one to model both complex objects in the investigated scene and borders of the changed areas and change details. Experimental results confirm the effectiveness of the proposed approach.