Globe-CraterSAM: a training-free generalizable framework for mapping sub-kilometer martian craters

Sub-kilometer craters are essential for resolving fine-scale Martian surface chronology, yet their global-scale mapping remains limited by the poor generalizability of existing methods and lack of systematic global-scale validation. Here, we propose Globe-CraterSAM, a training-free and generalized framework for sub-kilometer Martian crater mapping. Central to our framework is an adaptation strategy that identifies craters from Segment Anything Model (SAM)-segmented landform units by exploiting common circular depression characteristics of craters. Before crater identification, we introduce a Buffer-Split Loop Strategy (BSLS) to separate overlapping landform units, preventing adjacent landform units from distorting the circular-depression morphology of craters. A circularity index is then applied to exclude non-circular landform units. We further reconstruct 6-m-resolution quasi-3D terrain from monocular optical images using a Shape-from-Shading-based approach and apply Double-Gaussian Fitting to define quantitative criteria for crater-like depressions. This design enables topographic information to be incorporated into sub-kilometer crater mapping despite the limited resolution of existing global Digital Elevation Models (DEMs). To address the lack of systematic validation for sub-kilometer crater mapping, we construct the first globally comprehensive benchmark dataset, comprising more than 10,000 craters from 50 sub-areas spanning 26 distinct landform types across Mars. Validation results demonstrate the strong generalizability of Globe-CraterSAM, with an average F1 score of 90.09%, outperforming nine state-of-the-art (SOTA) models by 12.38–65.88 percentage points. The framework also demonstrates strong feasibility for large-extent applications, achieving an F1 score of 91.58% across 12 1° × 1° longitude-latitude tiles. Together, Globe-CraterSAM and the benchmark dataset provide a scalable foundation for sub-kilometer crater mapping and open new opportunities for high-resolution planetary investigation.