An Unsupervised Change Detection Based on Automatic Relationship Analysis

In general, the quality of difference image plays an important role in determining the overall performance of synthetic aperture radar (SAR) image change detection. However, owing to various reasons, there is inevitable loss of information for whatever difference image and actually helpful to change detection. To address such drawbacks, in this paper, an unsupervised change detection approach for SAR images is proposed based on clustering with the relationship of different scale. Primarily, the similarity of difference images is defined by radial basis function (RBF). In order to use complementary information from a mean-ratio image and a log-ratio image, the third difference image is generated by fusion processing. Before all the pixels are classified into changed and unchanged regions by k-means, the difference images are considered as features of the ground truth and their contributions to classification are given by RBF. Furthermore, making use of the statistical analysis and Markov Random Fields (MRF) of pixels, the most likely change pixels are classified to change class and the most likely unchanged pixels to another, the final change detection map is obtained. Experimental results obtained on two real SAR multi-temporal image sets confirm the effectiveness of the proposed technique.