An unsupervised change detection technique based on Bayesian initialization and semisupervised SVM

This paper presents a novel approach to unsupervised change detection, which is based on the combined use of the change vector analysis (CVA) technique and the semisupervised support vector machine (S3VM) classification method. The proposed approach aims at analyzing the information present in multitemporal images by jointly analyzing their original spectral signatures. This is accomplished by using the CVA technique in a selective way for defining a pseudotraining set necessary for initializing the S3VM binary classifier. Then, starting from these initial seeds, the S3VM performs change detection in the original multitemporal feature space. This is done by gradually involving unlabeled multitemporal pixels in the semisupervised learning procedure for better modeling the decision boundary between changed and unchanged pixels. Experimental results obtained on different multispectral and multitemporal images confirm the effectiveness of the proposed approach.