LunarRockSAM: A Domain-Adapted SAM with Bright-Spots Prompting and Conditional Screening for Lunar Rock Extraction

Authors:

Abstract
Accurate extraction of lunar rocks and analysis of their size distribution are essential for lunar rover navigation and geological investigations. However, previous methods often suffer from inaccurate boundary delineation for irregularly shaped rocks, and poor performance in complex terrain regions. In this study, we propose an enhanced segmentation anything model (SAM) framework, which integrates optical luminance and gradient variations as prior knowledge to improve lunar rock extraction (LunarRockSAM). The SAM mask decoder is fine-tuned on a manually labeled lunar rock dataset to improve adaption to lunar surface characteristics and significantly enhance segmentation precision. A novel bright-spots extraction algorithm is proposed to generate point prompts for accurate rock localization, reducing miss detections in complex terrain regions. Furthermore, a conditional screening mechanism is incorporated to refine segmentation masks for higher accuracy. Our method achieves pixel-level segmentation results with an average precision of 90% and a recall of 86%. Applied to Chang'e-3 (CE-3) and Chang'e-5 (CE-5) landing sites, the proposed method successfully counts rocks with diameters between 0.1 m and 7.17 m. The results show that the CE-3 site has a higher rock density, smaller sizes (0.1–2.72 m), and a steeper slope in size-frequency distribution curve, while the CE-5 site exhibits lower rock density and larger rocks (1.0–7.17 m). Based on these size and spatial distributions, we conducted a detailed comparative analysis of the geological characteristics and differences between the two regions, providing a valuable support for autonomous rover navigation and lunar geological investigation.

This publication has 40 references indexed in Scilit: