NASA and IBM’s Open-Source Lunar Model
NASA and IBM Research have developed an open-source lunar foundation model that utilizes 17 years of lunar orbiter data, providing a comprehensive platform for lunar science.
Key Features:
- Model Purpose: Supports tasks like crater detection, ice deposit prediction, and segmentation of Irregular Mare Patches, excelling in ice prediction at lunar poles.
- Data Utilization: Trained using SomBench, the largest co-registered multimodal lunar corpus, comprising nearly 2 million data bundles from various missions including NASA’s Lunar Reconnaissance Orbiter (LRO) and JAXA’s Kaguya probe.
Data Breakdown:
- High-Resolution Images: Approximately 1 million from the Narrow Angle Camera with a 1-metre resolution.
- Multispectral Images: About 964,000 from the Wide Angle Camera with a 100-metre resolution.
- Diverse Origins: Incorporates over 30 spatial layers from multiple instruments across four missions.
Model Adaptations:
- Lighting Conditions: The model accounts for lunar lighting, using imaging geometry to distinguish between illumination effects and actual surface properties.
- Multimodal Processing: Employs separate processing paths for different data types, contributing to its performance enhancements.
Performance Outcomes:
- Ice Prediction: Significant improvements with up to 22% reduction in prediction error compared to the baseline model.
- Crater Detection: Achieved a 19% improvement in coarse-scale crater detection while requiring less training data.
Research Implications:
- Analytical Tool: Designed to help researchers integrate various observations rather than replace measurement instruments.
- Limitations: Not ideal for absolute geodetic positioning and requires further testing to assess individual model innovations’ contributions.
Open Access and Collaboration:
- The model is publicly available along with its code and datasets, continuing a partnership under NASA-IBM’s AI for Science initiative. This project builds on previous foundation models like Prithvi, aimed at different scientific applications.
The NASA-IBM Lunar Foundation Model enhances lunar research capabilities, marrying extensive observational data with advanced AI methodologies to foster new scientific inquiries.
