A Multiscale Change Detection Technique Robust to Registration Noise

Abstract

This paper addresses the problem of unsupervised change detection in multitemporal very high geometrical resolution remote sensing images. In particular, it presents a study on the effects and the properties of the registration noise on the change-detection process in the framework of the polar change vector analysis (CVA) technique. According to this study, a multiscale technique for reducing the impact of residual misregistration in unsupervised change detection is presented. This technique is based on a differential analysis of the direction distributions of spectral change vectors at different resolution levels. The differential analysis allows one to discriminate sectors associated with residual registration noise from sectors associated with true changes. The information extracted is used at full resolution for computing a change-detection map where geometrical details are preserved and the impact of residual registration noise is strongly reduced.