A variational Bayesian approach to remote sensing image change detection

In this paper, we present a variational Bayesian (VB) approach to multitemporal remote sensing image change detection. The content of the so called `difference image' is modeled by finite Gaussians Mixture Model (GMM), then with the factor analysis techniques, underlying structure of image content is inferred automatically. Compared with the Expectation-Maximization (EM) algorithm, the proposed method can adaptively determine the number of components in the mixture model without usual sub- or over-segmentation problem. Moreover, to overcome the local optimization problem, a component split strategy is employed in inference process. Experimental results confirm the effectiveness of the proposed method.