Abstract
Change detection (CD) between optical and synthetic aperture radar (SAR) remote sensing images is a fundamental task in image interpretation. The Mamba has been introduced into CD due to its global modeling capability with linear complexity; but its unidirectional sequential modeling mechanism limits spatial structure perception. Moreover, domain discrepancies between optical and SAR sensors often introduce fake changes, compromising detection accuracy. To address these challenges, a novel cascaded fusion of convolutional neural network (CNN) and multi-scale attention-enhanced Mamba network (CFCMNet) is proposed. The framework first uses a deep image domain-transformation (DIDT) subnetwork with NICE-GAN to improve domain consistency. A CNN-MSAM (Multi-Scale Attention-enhanced Mamba) encoder is then designed for robust feature extraction, while a multi-output supervision (MOS) strategy enhances training stability and detection accuracy. The core MSAM module integrates the global modeling capability of state space models with a multi-scale attention aggregation mechanism, achieving unified representation of long-range dependencies and local details. Experimental results demonstrate that CFCMNet achieves state-of-the-art performance on four publicly available datasets (Gloucester- Ⅰ, Gloucester- Ⅱ, Shuguang, and California), with F1 scores of 94.17%, 91.07%, 96.17%, and 80.54%, respectively. These results confirm the effectiveness and advancement of the proposed method in cross-modal CD tasks.