Abstract:

It is difficult to obtain training data in practical synthetic aperture radar (SAR)image change detection tasks, and the detection rates of classical methods are low. A K-means and support vector machine (KMSVM) approach, which aims at extracting the change information in the original feature space without any training data, is proposed. 〖JP〗First, the threshold of difference images is selected by the K-means clustering method. Second, an offset should be obtained for deriving an unlabeled set and a pseudotraining set that is necessary for initializing a binary support vector machine (SVM) classifier. Finally, according to a semisupervised learning algorithm based on statistical characteristics, the SVM performs change detection by considering unlabeled data in definition of the decision boundary between unchanged and changed pixels. Experiment results show that, the proposed approach achieves better performance, higher generalization ability and more stable change detection precision, than the classical GMKI, GGMKI methods. These results prove the efficiency of the proposed approach.