Multispectral change detection using multivariate Kullback-Leibler distance

Abstract Change detection is one of the most critical applications in remote sensing. However, distinguishing between changes and non-changes in images collected at different dates and different imaging platforms is challenging. This is because the image dissimilarities caused by the difference in imaging conditions can mislead the change detection algorithms and result in false alarms. This problem is even more severe in urban areas due to a wide range of urban objects that have different materials and spectral signatures. To overcome this problem, the majority of studies in the recent literature use information-based methods for change detection. However, these methods are limited to using only a single band for change detection, without utilizing the multispectral properties of optical remote sensing images. In this paper, we propose a change criterion that uses the multivariate expansion of the Kullback-Leibler divergence to overcome the non-linear imaging condition differences and to utilize the multispectral properties for optical change detection. The proposed change criterion measures the similarity between the multivariate probability density functions of the corresponding objects in two images. For probability density functions, a Gaussian distribution is used whose parameters are approximated by a maximum-likelihood estimation. The degree of similarity between the two probability density functions is given by the MultiVariate Kullback-Leibler distance. The higher the similarity, the lower the probability of change. We tested the proposed change criterion on four real and one simulated urban datasets. The results demonstrate that the proposed method is robust against excessive imaging condition differences and can significantly improve the change detection results.