Adaptive Residual Mamba Network for Remote Sensing Change Detection

Although convolutional neural networks (CNNs) and Transformer-based models have achieved considerable success in the remote sensing change detection (RSCD) field, they often struggle to balance local detail preservation and global dependency modeling. Recently, state-space models such as Mamba have shown promising potential for efficient long-range representation learning. However, most existing Mamba-based RSCD methods are constrained by fixed residual connection mechanisms and selective scanning strategies, resulting in limited feature adaptability in spatial dependency modeling. To address these challenges, we propose Adaptive Residual Mamba (AR-Mamba), a novel Mamba-based framework that introduces adaptive residual and cross-gated scanning mechanisms. Specifically, the adaptive residual state (ARS) block employs adaptive kernel convolver and learnable scaling to dynamically calibrate residual signals. The cross-gated bi-scanning (CGBS) strategy performs bidirectional spatial scanning and cross-directional gating. Experimental results on the LEVIR-CD, SYSU-CD and WHU-CD datasets demonstrate the effectiveness of the proposed AR-Mamba, which achieves F1-score improvements of 0.82%, 1.95% and 1.62% over ChangeMamba, respectively.