A multidimensional scaling optimization and fusion approach for the unsupervised change detection problem in remote sensing images

It is generally well known that the overall performance of the most widely used types of unsupervised change detection methods, based on the luminance pixel-wise difference, is mainly relied on the quality of the so-called difference image and the accuracy of the classification method. In order to address these two issues, this work proposes to first estimate, a new and robust similarity feature map, playing the same role as the difference image, by specifying a set of constraints expressed for each pair of pixels existing in the multitemporal images. As a consequence, the proposed change detection method does not require any preprocessing step of the multitemporal images such as radiometric correction/normalization. In addition, input data can be acquired from different sensors. The quadratic complexity in the number of pixels of this new similarity feature map, between the multitemporal images, is reduced to a linear complexity procedure thanks to the FastMap-based optimization algorithm. Second, in order to achieve more robustness, changes are then identified, from this similarity feature map, by combining (fusing) the results of different automatic thresholding algorithms. Experimental results confirm the robustness of the proposed approach.