Maximum likelihood signal processing techniques to detect a step pattern of change in multitemporal SAR images

In this paper, we address the problem of deriving adequate detection and classification schemes to fully exploit the information available in a sequence of SAR images. In particular, we address the case of detecting a step reflectivity change pattern against a constant pattern. Initially we propose two different techniques, based on a Maximum Likelihood approach, that make different use of prior knowledge on the searched pattern. They process the whole sequence to achieve optimal discrimination capability between regions affected and not affected by a step change. The first technique (KSP-detector) assumes a complete knowledge of the pattern of change, while the second one (USP-detector) is based on the assumption of a totally unknown pattern. A fully analytical expression of the detection performances of both techniques is obtained, which shows the large improvement achievable using longer sequences instead of only two images. By comparing the two techniques it is also apparent that KSP achieves better performance, but the USP-detector is more robust. As a compromise solution, a third technique is then developed, assuming a partial knowledge of the pattern of change, and its performance is compared to the previous ones. The practical effectiveness of the technique on real data is shown by applying the USP-detector to a sequence of 10 ERS-1 SAR images of forest and agricultural areas, which is also used to validate the theoretical results.