Unsupervised Bayesian change detection for remotely sensed images

Abstract

The availability of remote sensing images with high spectral, spatial and temporal resolutions has motivated the design of new change detection (CD) methods for surveying changes in a studied area. The challenge of unsupervised CD is to develop flexible automatic models to estimate changes. In this paper, we propose a novel hierarchical Bayesian model for CD. Our main contribution lies in the application of Bernoulli-based models to change detection and transforming it to a denoising problem. The originality is related to the capacity of these models to act as implicit classifiers in addition to the denoising effect since even for changed pixels noise is also removed. The second originality lies in the way inference is conducted. Specifically, the hierarchical Bayesian model and Gibbs sampler ensure building an algorithm with secure convergence guarantees. Experiments performed on real data indicate the benefit that can be drawn from our approach.