SAM-Guided Attention Maps Fusion for Region Supervised Remote Sensing Image Change Detection

In practical applications, pixel-level annotation samples are both time-consuming and labor-intensive, posing a major challenge for large-scale applications. In this paper, we propose a segment anything model (SAM) guided attention maps fusion network for region supervised change detection (CD), namely SGAMFNet, which integrates a pre-trained foundation model SAM and aware attention map. Before training, SAM is utilized to generate segmentation maps based on region labels, initially separating the foreground and background within the region, which can avoid the high computational costs associated with directly integrating foundation models. Then, a foreground-background separation module (FBSM) is designed to further optimize the separation of foreground and background, generating higher quality pseudo labels. Specifically, it constructs a background attention map by retrieving background prototypes to suppress interference from unchanged pixels, and generates a foreground attention map by normalizing the difference features to highlight changed regions. The experimental results on the LEVIR, BCDD, and S2Looking datasets demonstrate that SGAMFNet performs well, effectively avoiding false detections caused by the mixing of foreground and background within region-level labels, thereby improving detection accuracy. This approach provides a novel perspective on region supervised CD and highlights the potential of foundation models in weakly supervised remote sensing image analysis. The code is available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/yunL719/SGAMFNet</uri>.