Generative Paradigms in Planetary Cartography: Benchmarking Foundation Models and Feature Prototype Filtering for Detecting 50m-Scale Martian Craters

The increasing availability of high-resolution orbital imagery, particularly from the Context Camera (CTX), provides the potential to resolve Martian surface features with unprecedented detail. However, existing crater catalogs are predominantly complete only for diameters larger than 1 km, leaving a critical knowledge gap regarding the distribution of sub-kilometer craters. This study addresses the challenge of mapping these small-scale features (down to ~50 m) by introducing a semi-automated framework that synergizes Generative AI benchmarks with feature space cleaning.

To establish a robust methodology, we systematically benchmarked various automated annotation strategies. We compared emerging unsupervised Foundation Models (including pure vision segmentation models like SAM and Multimodal Large Language Models like Gemini 3, GPT-5, and Qwen-Image) against traditional transfer learning baselines pre-trained on existing Lunar or large-scale Martian catalogs. Our analysis reveals that while transfer learning suffers from domain shifts and resolution mismatches when applied to fine-grained CTX targets, multimodal models demonstrate superior zero-shot generalization capabilities. Through extensive prompt engineering experiments, we found that identifying 50m-scale targets requires geologically contextualized prompts rather than simple geometric descriptions, although this comes with increased label noise.

To mitigate this noise, we developed a "Feature Prototype" cleaning mechanism. Utilizing a self-supervised vision transformer (DINOv2), we mapped candidate detections into a feature space defined by positive prototypes of diverse small-scale crater morphologies and negative prototypes of typical generative errors. By filtering samples based on feature distance, we achieved robust noise reduction.

The resulting dataset comprises 16,000 image tiles sampled from the Mars equatorial region (±30°). Notably, this workflow extends reliable detection capabilities down to the ~50-meter scale, demonstrating a distinct advantage over transfer learning baselines and traditional unsupervised methods in resolving fine-grained topography. This study not only fills a significant gap in small-scale crater records but also establishes a rigorous benchmark for leveraging foundation model knowledge in precision planetary cartography.