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

2026-07-24

AI Summary

This work addresses the challenges of heterogeneous lunar remote sensing data, sparse annotations, and fragmented modeling that hinder global representation learning and resource analysis. We propose the first multimodal foundation model covering the Moon’s mid-to-low latitudes (70°S–70°N), built upon a multimodal masked autoencoder architecture. The model integrates co-registered, 18-channel data from six instruments across three lunar missions—including multispectral, thermal infrared, and radar modalities—to learn a unified 768-dimensional shared embedding space. Accompanying this model, we release a machine learning–ready dataset, pretrained weights, and an embedding library to facilitate cross-instrument and cross-mission knowledge transfer. Evaluated on downstream tasks such as mineral abundance regression, geological unit classification, and similarity retrieval, our approach demonstrates significant performance gains under few-shot conditions.

Abstract

The renewed global focus on lunar exploration, driven by the prospect of in-situ resource utilization and a sustained human presence on the Moon, has created growing demand for accurate, large-scale characterization of the lunar surface. Although vast quantities of orbital remote-sensing data have been collected, scientific analysis and resource mapping remain fragmented by heterogeneous multiinstrument observations, sparse labels, and bespoke task-specific modelling workflows. Here we introduce LunarFM, a multimodal foundation model that learns a general representation of the lunar surface from diverse orbital measurements. LunarFM assimilates observations from six instruments across three lunar missions, mapping 18 input channels to a shared embedding space. We demonstrate that this embedding space supports a diverse range of downstream applications, including similarity search, few-shot resource mapping, mineral abundance regression, and geological unit classification, enabling efficient scientific investigation and resource-oriented analysis. We provide a machine-learning-ready dataset of co-registered multimodal observations spanning latitudes from 70°S to 70°N, a pretrained multimodal masked autoencoder, and a companion embedding dataset providing a joint 768-dimensional representation of lunar surface properties. All code and data are available at https://lunarfm.trillium.tech/

Problem

• lunar surface characterization
• multimodal remote sensing
• resource mapping
• heterogeneous data
• foundation model

Innovation

• multimodal foundation model
• lunar surface representation
• masked autoencoder
• cross-mission data fusion
• resource mapping