Geospatial Diffusion
Forecasting decades of land-cover change with a generative model.
Forecasting how the landscape will change — where impervious surfaces like roads and buildings will spread over the coming decades — is central to planning for flooding, heat, and urban growth. Most approaches treat this as pixel-wise classification or physical simulation. This project, carried out with Oak Ridge National Laboratory’s GeoAI group, asks a different question: can a diffusion model, the kind normally used to generate images, instead generate plausible futures for the land itself?
The target is imperviousness — the paved, built-up fraction of a place, which is a direct proxy for urban development. Rather than classifying what exists, the model is conditioned on what a region looked like historically together with auxiliary information about it, and learns to generate its likely future state. Trained across the continental United States on the National Land Cover Database, it produces decadal-scale forecasts of imperviousness change, showing that generative models can deliver useful large-scale predictions and opening a path toward physically informed Earth-system forecasting. The work was published at SIGSPATIAL 2025.