Dr. Sarah Berk recently represented Data-Driven EnviroLab (DDL) at the 2nd ESA-NASA Workshop on AI Foundation Models for Earth Observation at NASA’s Marshall Space Flight Center in Huntsville, Alabama. In this blog post, Dr. Berk shares her experience at the conference.
I recently attended the ESA-NASA Workshop on AI Foundation Models for Earth Observation (EO) in Huntsville, Alabama. The workshop was established to bring together foundational model developers and EO scientists and other end users, creating a valuable space for cross-domain collaboration as AI continues to advance EO research.
Satellites collect vast amounts of data every day, far more than can be efficiently analyzed using traditional approaches alone. At DDL, our work frequently combines satellite observations, ground-based measurements, and machine learning techniques to better understand urban environments. Our lab was previously awarded a NASA grant to study urban heat stress in the U.S. through satellite remote sensing data which has also informed our work to tackle heat stress for safer, healthier communities in North Carolina. The development in AI foundation models offer an exciting opportunity to accelerate scientific discovery by helping researchers extract information and insights from these increasingly large and complex datasets.
A recurring discussion throughout the workshop was the need for benchmarking in foundational models. Already, dozens of geospatial foundational models exist, but published work does not currently give answers as to which model a user should choose for their given task or which model is definitively better, given their wide range of downstream uses. If a simpler model achieves similar performance, are foundation models actually more accurate, faster to develop, or more cost-effective? Establishing robust benchmarking frameworks will be critical as the field matures. Many discussions centered on combining physics-based and data-driven approaches, such as embedding physics based models into pipelines, and using technologies such as quantum computing with AI to simulate complex processes like weather.
One highlight was hearing about ESA’s Φsat-2, a recently launched satellite powered by AI that can process imagery directly in orbit rather than transmitting all raw data back to Earth, making EO more efficient. As a speaker noted, “satellites observe but do not adapt or understand”. It was also fascinating to learn about foundation models being developed for solar forecasting and to predict weather on Mars. Whilst not my research area, I look forward to seeing how this field develops in the coming years.
The workshop also included hands-on training sessions. In the “Operational Geospatial AI: Fine-Tuning, Inference, and Scalable EO Model Serving” workshop, I learned how to fine-tune foundation models using IBM’s GeoStudio platform and deploy them for inference on EO tasks. A second workshop, “Quantitative Evaluation and Science-Driven Use of Weather Foundation Models,” focused on evaluating weather foundation models. Topics included appropriate performance metrics, accounting for distortions introduced by latitude-longitude grids through area weighting, and the importance of regional evaluations when assessing model performance.
Overall, the workshop highlighted both the rapid progress being made in AI for EO and the challenges that remain. The hands-on sessions highlighted how these technologies are increasingly being made accessible for a wider audience.
I’m excited to see how this field develops!
