NASA-IBM Releases Open-Source Multimodal Geospatial AI Model 'Lunar Foundation' for Lunar Exploration

A collaborative project led by NASA and IBM Research has released the 'NASA-IBM Lunar Foundation Model', an open-source multimodal geospatial AI model for analyzing lunar exploration data. This model integrates various physical characteristics of the lunar surface to support planetary science research and the development of AI-based exploration applications.

Integrated Dataset SomBench

The SomBench multimodal lunar dataset was used for model training and evaluation. This dataset contains approximately 2 million aligned data points and integrates 11 modalities, including:

• Imagery (including LROC NAC and WAC data)
• Topography and gravity data
• Thermophysical properties and radar data
• Illumination geometry and other geological information

Data resolution ranges from 1m to 20km/pixel, with the key technical challenge being the effective integration of different instruments and spatial resolutions while preserving the scientific value of each dataset.

Performance and Usage

The model demonstrated performance advantages over ImageNet pre-trained models and existing lunar-specific models. Key evaluation metrics include:

• Crater Detection: Improved accuracy in identifying craters at meter-scale resolution
• Segmentation: Enhanced performance in segmenting Maria (lunar plains) regions
• Label Efficiency: Achieving high performance with small amounts of labeled data (Label Efficiency)

The released resources include pre-trained model weights, fine-tuning code, and a standardized evaluation benchmark. These are accessible via Hugging Face, providing a foundation for planetary scientists and AI developers to reproduce and extend lunar-related ML tasks.