An adaptive multiscale random field technique for unsupervised change detection in VHR multitemporal images

This paper presents a novel multiscale technique for unsupervised change detection in very high geometrical resolution images based on adaptive multiscale random fields (AMSRF). AMSRFs are defined according to hierarchical segmentation applied to multitemporal images. Under the assumption that the relationship between random fields at different scales can be modeled according to a Markov chain, the statistical distribution of classes is sequentially estimated from the finest to the coarsest scale, and class labels propagated from the coarsest to the finest one. The method is developed within the framework of the Bayes decision theory. Experimental results obtained on a SPOT-5 multitemporal data set confirm the effectiveness of the proposed approach.