Authors: Jiangtao Liu – Pennsylvania State University
Title: Geospatial Foundation Embeddings as Transferable Catchment Descriptors for Rainfall-Runoff Modeling
Presentation Type: Poster Presentation
Abstract: Deep learning hydrological models typically rely on static catchment attributes to characterize basin heterogeneity, but these traditional descriptors often suffer from regional inconsistency and limited temporal representativeness. We evaluate geospatial foundation model embeddings as transferable alternatives for rainfall-runoff modeling, testing both the corporate-developed AlphaEarth (AE) and a custom-trained representation-learning model, StefaLand. Across 531 CAMELS basins, AE embeddings alone achieved performance comparable to traditional attributes, while their value was amplified under spatial extrapolation. In a global prediction in ungauged basins (PUB) scenario with 3,434 basins, combining embeddings with traditional attributes substantially improved predictive skill. A dual-path gated architecture revealed structured spatiotemporal patterns of embedding reliance, with models preferentially utilizing AE embeddings during snowmelt transitions and in humid basins. Flood-frequency analysis further demonstrated advantages for extreme return period events. These findings highlight foundation model embeddings as globally consistent catchment descriptors, particularly valuable for ungauged basin prediction.