Soma: Foundation Model for Multi-Modal Lunar Science using LRO Datasets

Deep learning techniques have advanced significantly in language modeling and computer vision tasks, demonstrating strong capabilities in understanding long sequences and supporting a wide range of Earth science applications, and are now expanding to planetary science. Foundation models (FMs), especially those using transformer architectures, offer powerful tools for diverse data tasks and adapt well to downstream applications. We present Soma, a foundation model built on data from the Lunar Reconnaissance Orbiter (LRO), which has gathered more data in its 16 years than all previous planetary missions combined, much of it still untapped. Lunar FM utilizes multiple lunar modalities and resolutions to reduce time spent sorting data, allowing scientists to focus primarily on analysis and discovery. The model uses an encoder–decoder transformer to extract semantic information from complementary LROC (Lunar Reconnaissance Orbiter Camera) data: high-resolution NAC imagery, stereo-derived DTMs, seven-band WAC strips, and static geophysical maps. One of the challenges is LROC images shift in resolution depending on location and when the images were taken. We also take into account different incidence and emission angles and areas on the Moon (latitude and longitude) to ensure a wide variety of test locations. Resolution-aware patch embeddings and lightweight spectral blocks unify these into a token grid. Pre-training combines masked-token reconstruction across modalities that aligns radiance and topography, producing modality-agnostic features. These transfer effectively, with minimal fine-tuning, to tasks like crater detection, volcanic feature mapping, and volatile deposit identification. We believe this is the first effort to design a FM for the Moon encompassing multiple modalities.