Dealing with uncertainty in multisensor image change detection based on rough set theory

Post-classification comparison is a common approach used for multi-sensor remote sensed imagery change detection in practice because it not only minimizes the impacts of sensor difference between multi-temporal images but also detect the detailed `from-to´ change while many other techniques can only detect `change/non-change´ information. But this approach also presents some well-known limitations, the most unfavorable one is that the change detection accuracy is seriously impacted by the accuracies of each individual classification, another inconvenient one is the precision of image co-registration will also have a big effect on the overall accuracy of change detection because the images acquired by different sensor have different spatial details. The primary objective of the research described here was to develop a method capable of dealing with spatial and classification uncertainty. In our study, we firstly classify the two remote sensed imageries and get the thematic maps, then each band of these two imageries are combined and segmented into image patches using FNEA (fractal network evolution algorithm). Instead of using the original concepts of rough set theory, here we refer to all the objects of two thematic maps as U in attribute-value system I= (U, A) while all the image patches as A. Under these premises, the change of each two corresponding objects from different thematic maps can be detected using upper and lower approximation. The land cover changes from 2004 to 2006 in a part of Kunming City of China are inferred as an example. The results show that this approach we presented is able to identify the land cover change more precisely than the post-classification method.