Mamba-CD: Mamba-Based Change Detection Network for Remote Sensing Images With Change Region-Aware Attention and Recursive Context Refinement Mechanism

Currently, Remote Sensing Image Change Detection (RSICD) based on Convolutional Neural Network (CNN) and Transformer has achieved significant advances. However, the existing methods often suffer from insufficient global dependency modeling capabilities or low computational efficiency. Furthermore, they struggle to handle complex background interference effectively, leading to imprecise detection of change region boundaries. To address these limitations, in this paper, a Mamba-based CD network for binary CD tasks, named Mamba-CD, has been proposed, in which VMamba is adopted as the backbone, enabling efficient global dependency modeling with linear complexity. Firstly, a Change Region-aware Attention Mechanism (CRAM) is designed to enhance features in change regions while suppressing features in background regions. The generated attention maps are constrained by a Boundary Loss function to strengthen the network's capability for the refined localization of change region boundaries. Subsequently, a Gated Recursive Context Refinement Module (GRCRM) is proposed. It utilizes a gating mechanism to filter out unimportant features and progressively fuses high-level global semantic information with low-level features in a recursive manner, thereby enhancing feature representation capability. Experimental results on three public datasets of LEVIR-CD, WHU-CD, and DSIFN-CD demonstrate that, compared to state-of-the-art CD methods, the proposed Mamba-CD achieves superior detection performance with lower model complexity and yields more precise boundaries of change regions.