Multimodal-Multiresolution Foundation Model for Lunar Remote Sensing

We present a multimodal foundation model for lunar remote sensing, pretrained from scratch on SomBench, a co-registered corpus of nearly two million tile bundles spanning 11 modalities at two spatial scales (i.e., 1 m/pixel and 100 m/pixel). The dataset is geographically partitioned to prevent leakage across train, validation, and test splits. The model adapts the TerraMind masked-token architecture to the lunar setting with two extensions: per-tile acquisition geometry (i.e., illumination angles, solar-frame anchors, tile footprint) is provided as explicit contextual information rather than requiring the model to infer it implicitly from image content; and tiles anchored at meter and hundred-meter scale are trained jointly in a mixed-batch regime so that a single set of weights covers both resolution families. A FlexiViT patch-embedding scheme allows the same pretrained checkpoint to be reused at different working patch sizes without backbone retraining, and modality-wise tokenization lets downstream users drop or add modalities at fine-tuning time. Qualitative generation experiments indicate that the model captures meaningful cross-modal correspondences (i.e., terrain derivatives from elevation, illumination-consistent reflectance from geometry, and coarse spatial pattern across co-registered modality sets). Evaluated across four downstream benchmarks (i.e., crater detection at 100 m/pixel and NAC 1 m/pixel imagery, irregular mare patch segmentation, and polar ice prospectivity regression) using TerraTorch, the pretrained model matches or outperforms ImageNet-pretrained baselines and an architecturally identical random-initialization control. On multimodal ice prospectivity regression at the poles, the pretrained model variants (i.e., full finetuning, Low-Rank Adaptation (LoRA) and frozen) outperform all baselines by margins that exceed run-to-run variability, with the randomly initialized version of our model that beats all but one baselines. This indicates that part of the gain comes from our architecture’s native token-level multimodal processing rather than from pretraining alone. Label efficiency is particularly pronounced for crater detection: at 50% training data on crater detection at 100 m/pixel, the pretrained model already exceeds the strongest ImageNet baseline trained on the full dataset. Among adaptation strategies, LoRA matches or surpasses full fine-tuning on crater detection and irregular mare patches segmentation, using a small fraction of the trainable parameters with lower run-to-run variance. Full fine-tuning remains best on ice-prospectivity regression, with LoRA that however still beats all baselines. We release the pretrained checkpoint, fine-tuning code, and benchmark datasets to support reproducibility and community reuse.