The Universities Space Research Association (USRA) has played a key role in the development of the NASA-IBM Lunar Foundation Model, an open-source artificial intelligence model aimed at aiding scientific analysis of lunar datasets. The model was created through a collaboration between NASA and IBM Research to help researchers work with the extensive data collected from lunar missions.

Read More

The Lunar Foundation Model was pretrained using a comprehensive dataset known as SomBench, which includes nearly two million co-registered data bundles across 11 modalities and two spatial scales. It integrates various types of lunar surface data, such as imagery, topography, mineralogy, and environmental information.

Dr. Rachel Slank, an associate scientist at USRA's Science and Technology Institute, contributed her expertise as a planetary science subject-matter expert. Based at NASA's Marshall Space Flight Center, she worked closely with both science and modeling teams to align lunar science priorities with the development and evaluation of the model.

The model was tested on three key benchmarks: crater detection, irregular mare patch segmentation, and lunar polar ice prospectivity. These assessments help evaluate its performance in identifying impact features and mapping volcanic landforms. Across all benchmarks, the NASA-IBM Lunar Foundation Model demonstrated comparable or superior results to models pretrained on ImageNet and a similar model without the lunar-specific pretraining. Notably, it showed high label efficiency in crater detection, indicating the potential to reduce the need for extensive labeled data in some tasks.

Designed to analyze multiple types of lunar observations simultaneously, the model accommodates both regional-scale Wide Angle Camera (WAC) data and meter-scale Narrow Angle Camera (NAC) data, incorporating information like terrain and illumination.

To further support the lunar research community, the NASA-IBM team has made the pretrained model, fine-tuning code, and benchmark datasets freely available. A significant part of this project was the creation of SomBench, which provides standardized datasets to evaluate machine learning applications in lunar science. SomBench features a core dataset used for pretraining and a suite of benchmarks addressing themes like impact processes and volcanic history.

Dr. Slank led efforts to develop a high-resolution crater benchmark utilizing images from the Lunar Reconnaissance Orbiter Camera, identifying over 49,000 lunar craters. She contributed significantly to the SomBench datasets and provided robust scientific reviews of both the SomBench and Lunar Foundation Model studies.

Dr. Slank remarked on the challenge of harmonizing lunar datasets taken from different instruments and resolutions, stressing the importance of maintaining scientific integrity in the process. "I can't wait to see what new science researchers will be able to do with the help of the NASA-IBM Lunar Foundation Model!"

The NASA-IBM Lunar Foundation Model and its datasets can be accessed on Hugging Face.