We combine Earth observation, AI, and physical science to understand our planet and how we can live better on it.
Lab News
August 2026: Our paper “Conformal Prediction Sets for Instance Segmentation” will appear at the Conference on Uncertainty in Artificial Intelligence (UAI). This work was inspired by the lack of rigorous uncertainty quantification in agricultural field boundary delineation, which we realized extended to other domains as well. In instance segmentation, ambiguity often comes not from where an object’s boundary lies, but from whether neighboring regions should be merged or split into separate objects. We introduce the first conformal prediction method that captures this structural uncertainty by producing calibrated sets of diverse segmentation masks. Congratulations to PhD student Kerri Lu and postdoc Dan Kluger!
July 2026: I’m honored to receive an NSF CAREER Award in support of our lab’s research. This project will investigate the principles governing Earth observation foundation models and develop open benchmarks and educational resources for the community. I’m very grateful to my students, collaborators, mentors, and NSF for their support. Excited for the work ahead!
June 2026: Congratulations to Matteo Peduto, Qidong Yang, and Jonathan Giezendanner for their paper “Observation‐driven correction of numerical weather prediction for marine winds” being published in the Journal of Geophysical Research: Machine Learning and Computation! We introduce ORCA (Observation-informed Real-time Correction with Attention), a transformer that learns to correct NOAA’s Global Forecast System forecasts using the latest available in situ observations. Unlike previous approaches, ORCA can naturally handle a changing set of observations and produce predictions at arbitrary locations across the ocean, reducing GFS wind forecast error by 45% at 1-hour lead time and by 13% at 48 hours.
