A multiscale technique for reducing registration noise in change detection on multitemporal VHR images

In this paper a multiscale technique for reducing the impact of residual misregistration on unsupervised change detection in very high geometrical resolution (VHR) images is presented. The proposed technique is based on an analysis of the statistical behaviour of registration noise present in multitemporal remote sensing images at different resolution levels. This characterization is carried out in the polar domain by analyzing spectral change vectors (SCVs) computed according to the change vector analysis (CVA) method. The proposed multiscale approach distinguishes between sectors associated with true changes and sectors associated with false alarms due to registration noise by differential analysis of the direction distributions of pixels at different resolution levels. This information is used at full resolution for computing a change detection map that shows: i) a high geometrical fidelity in the detail representation; and ii) a sharp reduction in false alarms due to the residual misregistration noise. The experimental analysis carried out on real VHR multitemporal images confirms the effectiveness of the proposed approach.