Artificial Intelligence

IBM Open Sources the Largest NASA AI Model on Hugging Face

By TFH,

Published on Aug 7, 2023   —   2 min read

IBM, in collaboration with NASA, has taken a significant step towards accelerating climate-related discoveries by open sourcing the largest geospatial model on Hugging Face. This initiative is designed to enhance the accessibility of NASA's extensive satellite data, thereby facilitating quicker analysis and fostering scientific innovation. The collaboration represents a significant stride in leveraging technology to address urgent environmental challenges.

A Collaborative Effort

Approximately six months ago, IBM and NASA initiated a joint mission to develop an artificial intelligence model with the capacity to expedite the analysis of satellite imagery. The primary objective of this endeavor was to gain a comprehensive understanding of Earth's rapidly evolving landscape and to render nearly 250,000 terabytes of NASA's mission data accessible to a broader audience, including researchers, scientists, and policymakers.

The Model's Capabilities: A Technical Overview

IBM's foundation model, now publicly available through Hugging Face, stands as the largest geospatial model hosted on the platform and marks the first open-source AI foundation model created in collaboration with NASA. The model's efficiency is evident in its ability to analyze geospatial data up to four times faster than existing deep-learning models, while requiring only half the amount of labeled data.

The model has undergone fine-tuning to map historical U.S. flood and wildfire events, measurements that hold potential for predicting future risk areas. Moreover, its versatility allows for redeployment in various tasks, such as tracking deforestation, forecasting crop yields, or detecting and monitoring greenhouse gas emissions.

Open Source for Progress: A Global Perspective

Jeff Boudier, head of product and growth at Hugging Face, underscored the significance of open-source AI, articulating that scientific advancement is contingent upon information sharing and collaboration. By open sourcing this model, IBM, Hugging Face, and NASA are extending an invitation to researchers across the globe to contribute to the enhancement and creation of other geospatial models and applications. This move embodies a shared vision for global cooperation and progress in the scientific community.

Technical Insights: Architecture and Training

The model is ingeniously constructed on a vision transformer and a masked autoencoder architecture, specifically adapted to process satellite images by broadening its spatial attention mechanism to encompass time. IBM utilized its AI supercomputer, Vela, and harnessed PyTorch and ecosystem libraries for the training and tuning process, focusing on labeled images of floods and burn scars from wildfires.

Conclusion

IBM's decision to open source this model is in harmony with NASA's Year of Open Science and its decade-long Open-Source Science Initiative. It also mirrors IBM's dedication to democratizing AI and making it accessible to all. This collaboration symbolizes a promising step towards tackling complex scientific issues and accelerating the wider deployment of AI across multifaceted applications. The partnership between IBM and NASA serves as a beacon of innovation, demonstrating the transformative potential of technology when applied with vision, collaboration, and a commitment to the greater good.

Share on Facebook Share on Linkedin Share on Twitter Send by email

Subscribe to the newsletter

Subscribe to the newsletter for the latest news and work updates straight to your inbox, every week.

Subscribe