應用研究
NASA and IBM Open-Source Lunar Foundation Model, Unifying 11 Types of Observational Data Through Geometric Conditioning and Mixed Resolutions
The NASA-IBM Lunar Foundation Model spans scales from 1 to 100 meters per pixel using nearly 2 million aligned lunar data samples. The model, dataset, and fine-tuning code are all publicly available, but its ice-related outputs remain indirect inferences and must not be used for landing or exploration decisions.
NASA and IBM released the NASA-IBM Lunar Foundation Model (LFM) on September 10, aiming to transform lunar remote-sensing data scattered across different missions, instruments, and resolutions into reusable visual representations. The Apache 2.0-licensed weights, SomBench dataset, and TerraTorch fine-tuning code are now available online.
LFM is a ViT-B encoder-decoder trained from scratch. Its encoder has a width of 768, 12 layers, and 12 attention heads, while pretraining uses 256×256-pixel inputs. Each of the nine image-based modalities is discretized separately using a VQ-VAE and finite scalar quantization (FSQ), then jointly trained with a TerraMind-like masked-token prediction objective. FlexiViT patch embeddings allow downstream tasks to use different patch sizes.
Two lunar-specific design choices are especially noteworthy. First, solar incidence angle, phase angle, observation azimuth, and tile coordinates are serialized as explicit context, helping prevent the model from mistaking illumination-induced shadows for terrain features. Second, Lunar Reconnaissance Orbiter Camera (LROC) Narrow Angle Camera (NAC) tiles at approximately 1 m/px and Wide Angle Camera (WAC) tiles at approximately 100 m/px are trained within the same mixed-batch loop, enabling one set of weights to span a roughly 100-fold difference in scale. Overall pretraining used 16 H100 GPUs, 150,000 steps, and approximately 1,100 GPU-hours.
Official experiments across five random seeds show an RMSE of 0.0293 for polar ice-potential regression, outperforming SwinV2-B at 0.0377. For WAC crater detection, the model also matched or exceeded baselines trained on the full dataset while using only half of the labeled data. LoRA matched full fine-tuning on crater detection, making it the more practical starting point.
The limitations are equally important. Ice-potential labels come from a knowledge-driven fuzzy overlay rather than direct measurements of subsurface ice. The model also lacks an absolute geodetic reference, and generated latitude and longitude values may be inaccurate by tens of degrees. NAC training data are concentrated in the limited regions for which stereo-derived terrain maps are available. Engineering teams should therefore focus next on cross-region generalization, ablation results that isolate the effects of geometric tokens and mixed-resolution training, and whether third parties can reproduce benchmarks currently produced primarily by the release team.