CSGANet: Lightweight Channel-Split Group Attention for High-Resolution Remote Sensing Change Detection

Change detection (CD) in remote sensing imagery is a fundamental tool for Earth monitoring, with wide-ranging applications. Existing methods face an accuracy–efficiency tradeoff—heavy models are costly, lightweight ones erode boundary fidelity and smallchange recall—and lack a unified mechanism to fuse local, directional, and global context. To address these issues, we propose CSGANet, a lightweight Siamese framework featuring two key innovations: a channelsplit group attention module, CSGA, to selectively integrate complementary local, directional, and global contexts, and an Adaptive PathMixing Module (AdmPM) for globally guided, boundarypreserving downsampling. The CSGA module asymmetrically allocates channels to different attention branches, striking a balance between computational cost and representational power. Concurrently, AdmPM stabilizes crossscale semantics and preserves finegrained structures during downsampling. Extensive experiments on three public benchmarks: LEVIR, SYSU, and S2Looking, against nine stateoftheart methods, demonstrate that CSGANet achieves competitive accuracy with a compact footprint, reaching F1 scores of 91.82%, 84.26%, and 66.11%, respectively, with only 4.15 M parameters and 16.39 GFLOPs. Compared to recent methods, our method produces cleaner change maps with fewer false alarms and missed detections while significantly reducing model complexity. These results underscore a favorable accuracy–efficiency balance for largescale, highresolution CD, offering a versatile and lightweight design applicable to broader remote sensing tasks. The code for this work is available athttps://github.com/jxy0413/csganet.