Hypothesis-test-based landcover change detection using multi-temporal satellite images – A comparative study

Remote sensing images and technologies have been widely applied to environmental monitoring, in particular landuse/landcover classification and change detection. However, the uncertainties involved in such applications have not been fully addressed. In this paper two hypothesis-test-based change detection methods, namely the bivariate joint distribution method and the conditional distribution method, are proposed to tackle the uncertainties in change detection by making decisions based on the desired level of significance. Both methods require a data set of class-dependent no-change pixels to form the basis for class-dependent hypothesis test. Using an exemplar study area in central Taiwan, performance of the proposed methods are shown to be significantly superior to two other commonly applied methods (the post-classification comparison and the image differencing methods) in terms of the overall change detection accuracies. The conditional distribution method takes into consideration the correlation between digital numbers of the pre- and post-images and the effect of the known pre-image digital number on the range of the post-image digital number, and therefore yields the highest change detection accuracy. It is also demonstrated that the class-dependent change detection is crucial for accurate landuse/landcover change detection.