Gaussian process approach to change detection for high resolution remote sensing image
基于高斯过程的高分辨率遥感图像变化检测

Gaussian process (GP) represents a powerful theoretical framework for Bayesian classif ication. Despite GP classifier have gained prominence in recent years, it remains an approach whose potentialities are not yet suff iciently known in remote sensing community. This paper gives a thorough investigation of GP CLASSIFIER for high resolution (HR) multi-temporal image change detection. Firstly, we give a detailed analysis of the capabilities of GP classif ier in theory. Secondly, we elaborately explore the advantages and disadvantages of the GP classif iers. Finally, we design several experiments to test the performance of the GP classif ier for HR remote sensing image change detection. Moreover, we propose a novel approach for improving the capacities of GP classif ier in remote sensing image change detection. The proposed context-sensitive change detection method is achieved by analyzing the posterior probability of probabilistic GP classif ier within a markov random f ield (MRF) framework. In particular, the method consists of two steps: (1) A supervised initialization is founded on a probabilistic GP classif ier; (2) A MRF regularization aims at ref ining the posterior probability by employing the spatial context information. Five experiments carried out on HR remote sensing image set validate the power of GP classif ier for change detection and also the effectiveness of our proposed methods.