DASAM-CD: Dual-attention adaptive SAM for change detection on remote sensing images

High-resolution remote sensing imagery provides detailed spatial information for land-cover change monitoring, but it also increases the difficulty of accurately detecting changes in complex scenes with acceptable computational cost. Existing bi-temporal change detection methods still face challenges in balancing detection accuracy and model complexity. To address this issue, we propose DASAM-CD, a lightweight-oriented change detection framework based on the Mobile Segment Anything Model (MobileSAM). The proposed framework integrates hierarchical feature extraction, dual-attention feature enhancement, and multi-stage Bit_Fusion into a unified pipeline. Specifically, MobileSAM is used to extract multi-level bi-temporal features, from which shallow and deep representations are selected to preserve spatial details and semantic cues. The dual-attention enhancement module further refines channel-wise and spatial change-related responses, while the Bit_Fusion strategy integrates enhanced bi-temporal features for final change prediction. To evaluate the proposed method, we construct the Macao Change Detection Dataset (MCDD) for urban change detection and conduct additional experiments on the CropLand Change Detection Dataset (CLCD) for agricultural land-cover change detection. Experimental results show that DASAM-CD improves IoU by 4.46 and 1.08 percentage points over Meta-CD on MCDD and SAM-CD on CLCD, respectively. In addition, DASAM-CD contains 7.21M parameters and requires 26.46G FLOPs under an input resolution of \(256 × 256\) , indicating a compact model size and moderate computational complexity while achieving competitive detection performance.