A SAM-Based Small-Scale Attention Network for Lunar Impact Crater Detection with DEM Data

The lunar polar regions, owing to their unique geographical location and the potential presence of water ice, have emerged as primary targets for recent lunar exploration missions by major spacefaring nations. Small impact craters, as the most ubiquitous and prominent surface features, offer invaluable scientific evidence for investigating both the internal and surface geological processes of planetary bodies. However, the inherent low solar elevation in these regions results in extensive permanently shadowed areas, which significantly hampers the acquisition of high-quality optical remote sensing imagery. Consequently, existing crater detection networks designed for optical images exhibit suboptimal performance when applied to digital elevation model (DEM) data. Moreover, the small scale and overlapping distribution of craters further constrain the effectiveness of current object detection methodologies. To address these challenges, this study proposes an object detection approach based on the Segment Anything Model (SAM). In the proposed framework, a SAM image encoder with frozen parameters is first employed to extract robust encoded features, which are then enhanced by a novel Small-Scale Attention (SSA) module designed to decode features specifically for small-scale targets. These refined features are subsequently integrated with a YOLO detection head to achieve high-precision crater detection. Experimental results demonstrate that the proposed method surpasses existing state-of-the-art object detection techniques in terms of detection accuracy. This research not only offers a new perspective for lunar polar region exploration but also broadens the applicability of SAM in remote sensing analysis.