LunarFM: A Shared Multimodal Representation of the Moon's Surface

TL;DR - LunarFM is a multimodal foundation model that unifies heterogeneous lunar orbital measurements into a shared surface representation. It could streamline scientific analysis and resource mapping despite sparse labels and fragmented datasets.

• Integrates 18 channels from six instruments across three lunar missions.
• Uses a multimodal masked autoencoder to produce 768-dimensional embeddings.
• Supports similarity search, few-shot resource mapping, mineral abundance regression, and geological unit classification.
• Releases code, pretrained models, embeddings, and co-registered data covering 70°S to 70°N.