Change Detection for Hyperspectral Images Using Extended Mutual Information and Oversegmentation

We propose a change detection algorithm for hyperspectral images by properly extending the description of commonly used mutual information metric in monochrome images to hyperspectral images. The newly extended metric for the additional spectral dimension in hyperspectral images accumulates the effects of all spectral bands to the statistical relation between the pixels of the two images at the same location. In order to avoid blurring kind of distortions in the change maps resulting from the usage of fixed size kernels during the calculation of mutual information in previous literature, the proposed method first applies an oversegmentation to the hyperspectral images and then, computes the extended metric over the produced superpixels. The proposed approach based on joint superpixels and extended mutual information is compared with two basic approaches, whereas the first one uses conventional mutual information of two images and the second method utilizes the extended mutual information over rectangular kernels. The experimental results indicate that the change masks obtained by the proposed method are more accurate compared to the baseline approaches.