Change Detection in SAR Images Based on Evolutionary Multiobjective Optimization and Superpixel Segmentation

Abstract

Synthetic aperture radar (SAR) imagery has been widely used in the field of remote sensing image change detection. However, its disadvantage of strong coherent multiplicative noise reduces the accuracy of change detection results. This paper proposes a novel SAR image change detection method, which is mainly comprised of three steps. Firstly, the difference image (DI) which is generated by log-ratio operator is segmented into superpixels by Simple Linear Iterative Clustering (SLIC) Algorithm. Secondly, superpixels are encoded uniformly in order to be utilized as the training samples, and deep neural network is used to extract deep features of DI. Finally, this paper designs an improved clustering algorithm which is optimized by Non-dominated Sorting Genetic Algorithm (NSGA-II). When the deep features of DI are used to cluster, Bhattacharyya distance between two categories of samples is selected as the similarity measurement. Taking the logarithmic likelihood function of clustering algorithm and the Bhattacharyya distance between the two categories as two optimization objectives, NSGA-II algorithm is used to optimize the model, and a set of pareto optimal solutions are thus generated. Compared with various indexes for accuracy evaluation, the map which has the highest accuracy is the final change detection map. Experimental results on real synthetic aperture radar datasets show that the proposed method is superior to other classical change detection methods, which demonstrates its effectiveness, feasibility, and superiority of the proposed method.

This work was supported in part by the Key-Area Research and Development Program of Guangdong Province under Grant 2020B090921001 and in part by the National Natural Science Foundation of China under Grant 62036006 and Grant 61906146, and in part by the Key Research and Development Program of Shaanxi Province under Grant 2018ZDXM-GY-045.