A SAR Remote Sensing Image Change Detection Method Based on DR-UNet-CRF Model

Abstract

A Synthetic aperture radar (SAR) image change detection method based on DR-UNet-CRF iterative structure is proposed by introducing a regional dynamic convolutional network to address the problems of semantic information fading phenomenon and indeterminacy of change boundaries due to differential image computation in remote sensing image change detection. Firstly, a DR-UNet segmentation network based on the dynamic region-aware convolution (DRConv) kernel is conceived to supply a univalent potential function for the conditional random field, and a guide-mask generation method guided mask generation method with feature pyramid network (FPN) based structure is presented to guide an improved dynamic convolutional UNet to obtain accurate remote sensing change regions by learning fine spatial region delineation. Secondly, the pair-wise potential function based on image grayscale features and spatial features is designed to model the inter-pixel relationship. Finally, we use a fully connected conditional random field (CRF) model to iteratively optimize for change regions to achieve semantic compensation, thus defining the boundaries of remote sensing images more precisely. By comparing with the mainstream change detection methods, it can be considered that method in this paper has better detection performance.