Unsupervised change detection of remote sensing images using superpixel segmentation and variational Gaussian mixture model

Abstract

In this paper, we propose a novel unsupervised change detection approach for multitemporal remote sensing (RS) images based on superpixel segmentation and variational Gaussian mixture model (GMM). Firstly, the generated difference image is segmented into multiple superpixels using entropy rate superpixel (ERS) segmentation, which allows for the spatial contextual information to be taken into account in change detection. As such, we utilize the GMM to model the distribution of these superpixels, and assign each superpixel to one component using variational inference (VI) algorithm. Subsequently, according to mean square error (MSE) criterion, the resulting clusters are further grouped into two classes, respectively representing the changed class and unchanged class. As a consequence, we can achieve the change mask (CM) by assigning the superpixels (and its pixels) to the corresponding classes. Experimental results demonstrate the effectiveness of the proposed method with two real multitemporal RS images.