Unsupervised change detection for remote sensing images based on object-based MRF and stacked autoencoders

This paper proposes a novel algorithm of unsupervised change detection for remote sensing images based on object-based MRF (OMRF) and Stacked Autoencoders(SAE). To overcome the edge contraction phenomenon of MRF model, we propose an OMRF model, in which we assume that pixels within the same object will be classified into the same category. Then, a network of SAE is introduced to form a detector that can learn how to analyze the images to be detected and recognize the changed pixels and unchanged pixels, with the reference of pre-classified images just obtained by the object-based MRF model. The experiment results show that the overall error rate is decreased and the accuracy of change detection is obviously promoted. We can draw the conclusion that SAE plays a substantial role in improving the effectiveness of change detection because of its powerful ability of features extraction.