Data Science Talk Series 

Geospatial Diffusion for Land Cover Imperviousness Change Forecasting

Abstract: Debvrat Varshney is a Postdoctoral Research Associate with the GeoAI Group at Oak Ridge National Laboratory (ORNL). He works on computer vision and machine learning for multimodal geospatial applications. At ORNL, he's using diffusion models for landcover forecasting, generating synthetic satellite imagery, and developing their detection techniques. 

Bio of the speaker: Land-use and land-cover (LULC) have a significant effect on several Earth system processes. For example, impervious surfaces reduce water infiltration, impacting regional hydrology and flood risk. While Earth System models have improved forecasting hydrologic and atmospheric processes at higher resolutions, the ability to forecast LULC change has lagged. This talk proposes a new paradigm exploiting Generative AI for land cover change forecasting by framing it as a data synthesis problem conditioned on historical and auxiliary data sources. We demonstrate how a domain-informed conditional diffusion model is able to make accurate forecasts at 700m x 700m resolution, beating baseline null models. We also discuss future research to incorporate Earth’s physical properties and enable scenario simulations via driver variables.