NASA and IBM's open source lunar model turns 17 years of orbiter data into a foundation for lunar science
decoder ·
NASA and IBM have released the NASA-IBM Lunar Foundation Model, an open-source AI tool designed to make decades of lunar observation data accessible for machine learning. Trained from scratch on nearly two million tile bundles spanning 17 years of data from missions like LRO and GRAIL, the model addresses the scarcity of labeled data in lunar science by leveraging large volumes of unlabeled inputs. A key innovation is its explicit inclusion of lighting geometry as input, which helps the model distinguish between surface properties and illumination effects more effectively than previous methods. The model demonstrates significant improvements in predicting polar ice deposits, reducing prediction error by up to 22 percent compared to baseline models. It also performs competitively in crater detection at various scales and segments Irregular Mare Patches, though it does not replace physical measurement instruments for precise geodetic positioning. By assigning separate processing paths to different data layers and using techniques like FlexiViT, the model adapts to diverse image patch sizes without retraining. This release marks a milestone in the NASA-IBM collaboration under their Space Act Agreement, following earlier Earth observation models like Prithvi. The model, code, and datasets are publicly available via Hugging Face, GitHub, and the TerraTorch toolkit.
decoder