Detection of changes in SAR images based on an improved fully-connected conditional random field

Change detection is the research focus of remote sensing. To overcome the shortcomings of the existing conditional random field (CRF)-based change detection, this study proposed a novel change detection method for synthetic aperture Radar (SAR) images based on an improved fully connected CRF (FCCRF). Firstly, this study summarized the comparative algorithms for generating differential images from SAR images, which were divided into three levels, namely pixel, neighborhood, and super-neighborhood. Then, this study selected three typical comparative algorithms-log ratio (LR), neighborhood ratio (NR), and improved non-local graph (INLG)-to produce three sets of complementary differential images. Finally, this study improved the FCCRF by extending the number of Gaussian kernels of the pairwise potential function of FCCRF and generated the change detection maps using the improved FCCRF model. The change detection method proposed in this study integrated the two-phase original SAR images, three sets of complementary differential images, and the global spatial information of images. In addition, this study presented a simple and effective parameter determination strategy, which allows the FCCRF to perform the change detection automatically. Experimental results on four sets of real SAR image data confirmed the effectiveness of the change detection method proposed in this study.