Integrating Thresholding With Level Set Method for Unsupervised Change Detection in Multitemporal SAR Images

Abstract

ABSTRACTIn this study, we present a new approach for unsupervised change detection in multitemporal synthetic aperture radar (SAR) images based on integrating thresholding with level set method (LSM), which is free of any prior assumption about modeling the data distribution in the difference image. The proposed approach exploits a discrete wavelet transform fusion strategy aimed at achieving the optimal difference image from the mean-ratio and log-ratio difference images. The generated binary change map (CM), by applying a thresholding method on the fused difference image, is used as the initial contour to produce a final CM on fused difference image using the LSM. Several non-fuzzy and fuzzy thresholding methods are considered to assess the generation of the initial contour for the LS segmentation. To indicate the effectiveness of the proposed method, experiments are implemented on 2 sets of multitemporal SAR images from TerraSAR-X and ERS–2 satellites, respectively. Results of the proposed method were compared with results of some existing state-of-the-art unsupervised change detection methods. Experimental results prove the competence of the proposed method in terms of computational time and accuracy over the unsupervised change detection procedure.