Dr. Yiqun Xie from CGIS will be presenting a talk, "Harnessing AI Challenges for Earth Science Problems: From Space to Physics" on Tuesday, October 10, 2023, from 4pm-5pm EST at a virtual event sponsored by iHARP: NSF HDR Institute.
Please visit the event's webpage on the UMBC iHARP website for the meeting link and more details.
Advances in deep learning have continued to set new expectations for general tasks (e.g., computer vision, natural language processing) and bring new potential to harness geospatial big data for Earth Science problems. However, direct applications of deep learning often fall short due to challenges posed by geospatial data, including spatial heterogeneity/variability, sparse labels, etc. This talk will start with two general frameworks that explicitly tackle the challenges with: (1) A heterogeneity-aware framework that automatically recognize and handle spatial variability during model training; and (2) A physics-informed meta-learning framework that learns to select ensembles of physical models to assist training and reduce the need of labeled data. Then, the talk will show several examples of use-inspired AI for Earth Science, with applications in ICESat-2 height interpolation, global ecosystem model approximation, and label-free cloud masking. Finally, I will discuss our recent work on a coincidental data discovery platform to facilitate Arctic research.