Edge-guided multi-scale foreground attention network for change detection in high resolution remote sensing images

Abstract

Recently, deep learning methods have shown outstanding performance in remote sensing change detection (CD). To improve the accuracy of edge regions, researchers have introduced edge information as prior knowledge into CD methods. However, there is still room for improvement in utilizing edge information. In this study, an edge-guided multi-scale foreground attention network (EMAF) with an innovative and effective strategy is proposed. Our edge-guided strategy focuses on incorporating edge information into the feature fusion process at all levels. EMAF firstly employs a Siamese Resnet-34 as the encoder, producing four levels of bi-temporal features. Secondly, the Foreground Module (FM) is introduced to separate foreground and background at each level, resulting in foreground change features. Then, the Edge Fusion Module (EFM) is introduced to fuse the features of two adjacent levels with the edge probability map obtained by the Edge Extraction Module (EEM). Finally, after performing EFM and EEM in the three levels, EMAF generates the final change map and edge probability map. During the deep supervision training, the Dynamic Weighted Binary Cross-Entropy (DWBCE) loss is calculated as the edge loss. The comparison testing results on WHU-CD, SYSU-CD, and LEVIR-CD datasets validate the effectiveness of EMAF. Notably, EMAF achieves impressive Intersection over Union (IoU) of 89.03% on WHU-CD and 71.04% on SYSU-CD datasets.