MAGMamba-CD: A Lightweight Multiscale Adaptive Gated Mamba Network for Remote Sensing Change Detection
Abstract
Binary change detection from bitemporal remote sensing images is essential for land-cover monitoring, urban analysis, and disaster assessment. However, high-resolution scenes often contain complex object structures, scale variations, registration noise, illumination changes, and ambiguous boundaries, making accurate change localization challenging. In this letter, we propose MAGMamba-CD, a lightweight multiscale adaptive gated Mamba network for remote sensing change detection (RSCD). The proposed multiscale adaptive gated Mamba encoder introduces local feature enhancement and adaptive gating before global state-space scanning, thereby capturing boundary- and scale-aware, long-range dependencies. A progressive Mamba decoder is then adopted to recover spatial details from hierarchical features. Moreover, a semantic difference interaction module combines prechange and postchange features with explicit absolute-difference features, while spatial attention refinement suppresses irrelevant activations and strengthens discriminative change responses. Experiments on CDD, LEVIR-CD, and SYSU-CD demonstrate that MAGMamba-CD achieves a favorable accuracy–efficiency tradeoff. With only 8.85 M parameters and 15.60 G floating-point operations (FLOPs), MAGMamba-CD obtains $F1$ -scores of 96.10%, 90.86%, and 79.36% on CDD, LEVIR-CD, and SYSU-CD, respectively, showing competitive performance against representative convolutional neural network (CNN)-based, Transformer-based, and Mamba-based methods.