Unsupervised Change Detection in Multi-temporal Satellite Images Based on Structural Patch Decomposition and k-means Clustering for Landslide Monitoring.

This paper proposes a new, simple, and effective unsupervised change detection method for multi-temporal satellite imagery in the landslide monitoring system. The method combines a modified version of k-means clustering, so called the adaptive k-means clustering algorithm and a structural consistency map to binarize input multi-temporal satellite images. The absolute differences between pairs of contiguous binary images are used to construct a change map. The structural consistency map is generated by a method based on the structural patch decomposition. The evaluation results show that the proposed method is considerably better than a conventional method based on principal component analysis (PCA) and k-means clustering in terms of pixel error rate, correctness, and robustness against noise addition. For the pixel error rate and the correctness, the improvements are approximately \\(42.61\\%\\) and \\(13.92\\%\\), respectively. Also, when noises are added into the input images, the pixel error rate of the proposed method is improved by \\(98\\%\\) from that of the PCA-based method, and the correctness is improved by \\(77.27\\%\\).