Unsupervised Change Detection for Multispectral Remote Sensing Images Using Random Walks

In this paper, the change detection of Multi-Spectral (MS) remote sensing images is treated as an image segmentation issue. An unsupervised method integrating histogram-based thresholding and image segmentation techniques is proposed. In order to overcome the poor performance of thresholding techniques for strongly overlapped change/non-change signals, a Gaussian Mixture Model (GMM) with three components, including non-change, non-labeling and change, is adopted to model the statistical characteristics of the different images between two multi-temporal MS images. The non-labeling represents the pixels that are difficult to be classified. A random walk based segmentation method is applied to solve this problem, in which the different images are modeled as graphs and the classification results of GMM are imported as the labeling seeds. The experimental results of three remote sensing image pairs acquired by different sensors suggest a superiority of the proposed approach comparing with the existing unsupervised change detection methods.