Speaker
Description
Terrestrial analog sites, Earth environments resembling extraterrestrial surfaces, are an established methodology for validating instruments, algorithms, and mission concepts in planetary exploration. Poland's Błędów Desert, historically shaped by active aeolian dune fields and now substantially reduced through vegetation encroachment, has served as such an analog site, and recent work has shown the viability of machine learning classification on Sentinel-2 imagery of the region for land cover monitoring.
This study investigates whether feature representations learned from terrestrial analog satellite data generalize to Martian orbital imagery, where independent ground truth is inherently unavailable. Sentinel-2 multispectral imagery is used to isolate the active, vegetation-free dune core of Błędów via NDVI filtering, restricting analysis to terrain most analogous to Martian conditions. Sentinel-1 SAR data is incorporated to characterize surface roughness independently of optical reflectance, reflecting radar's established role in planetary surface characterization. A pretrained self-supervised vision model extracts feature embeddings from both optical and radar data, without task-specific labels or Mars-derived training, and these embeddings are evaluated on public Mars orbital imagery (CTX and HiRISE) for zero-shot transfer performance on terrain differentiation and anomaly detection.
While self-supervised methods have been applied to Mars rover imagery, and transfer learning is well studied within Earth observation, this work addresses a distinct and, to our knowledge, unaddressed question: whether Earth-derived orbital and radar representations transfer meaningfully across planetary domains. We report transfer performance across both modalities, characterize failure modes, and discuss implications for multi-sensor terrestrial analog data as a low-cost proxy for validating AI-assisted planetary surface analysis.