An Object-Based Method Based on a Novel Statistical Distance for SAR Image Change Detection

This paper introduces an object-based method based on a new statistical distance for SAR image change detection. Firstly, multi-temporal segmentation is carried out to segment two temporal SAR images simultaneously. It considers the homogeneity in two temporal images, and could generate homogeneous objects in spectral, spatial and temporal. In addition, through setting different segmentation parameters, the multi-temporal images can be segmented in a set of scales. This process exploits the advantages of OBIA that could effectively reduce spurious changes, and considers the scale of change detection task. Secondly, a multiplicative noise model called Nakagami-Rayleigh distribution is employed to describe SAR data, and then applied to Bayesian formulation. Thus, a new statistical distance that is insensitive to speckles is derived to measure the distances between pairs of parcels. Then, cluster ensemble algorithm is utilized to improve accuracy of individual result in each scale to obtain the final change detection map. Finally, multi-temporal Radarsat-2 images are employed to verify the effectiveness of the proposed method compared with other four methods.