Go to ICLR 2026 Workshop FM4Science homepageLunarFM: a multimodal representation of the Moon's surface
Keywords: geospatial foundation models, multi-modal learning, remote sensing, planetary science, masked autoencoders, AI for science, lunar science
TL;DR: An open, multi-modal foundation model that unifies heterogeneous lunar remote-sensing data for scalable scientific analysis and resource mapping.
Abstract: The renaissance of lunar exploration is largely driven by the dual goal of assessing the Moon’s resource potential and establishing a sustained human presence.
However, current resource mapping efforts remain fragmented due to heterogeneous, multi-instrument datasets that expose fundamental limitations in existing analysis pipelines. Here, we introduce LunarFM, a first-generation foundation model that takes a step toward unifying heterogeneous lunar remote-sensing observations from multiple instruments in a shared representation space, integrating data from six instruments across three orbital missions
with a total of 18 input channels. It is trained via multimodal self-supervised pre-training. The LunarFM framework consists of (1) a machine-learning-ready dataset of co-registered multimodal chips spanning 0.5°$\times$0.5° of lunar latitude and longitude, covering $\pm$70° latitude; (2) a pre-trained multimodal masked autoencoder trained on these inputs, and a companion embedding dataset providing a joint 768-dimensional representation of lunar surface properties; and (3) three illustrative downstream applications, including similarity search and few-shot resource mapping (e.g., ilmenite) to demonstrate the framework's potential.
LunarFM is intended to lower the barrier to entry for lunar scientific investigation and resource-oriented analysis, with the expectation that future work will extend and rigorously benchmark the framework.
All code and data is available at \url{https://lunarfm.spaceml.org}
Submission Number: 129
Loading