Calibrated Station-Scale Temperature Downscaling from a Temperature-Agnostic Geospatial Foundation Model

Abstract

We downscale coarse climate variables to station-scale temperature using frozen embeddings from a temperature-agnostic geospatial foundation model (Clay v1.5), and study when the resulting uncertainty can be trusted. Across three climate regimes, embedding-based decoders reduce RMSE by 38 to 68 percent over strong statistical downscalers. Split-conformal prediction intervals are near-nominal in-regime, but their empirical coverage collapses from a nominal 90 percent to about 12 percent under cross-regime transfer (an Arizona-trained model applied to the Pacific Northwest); group-conditional (Mondrian) conformal partially restores valid coverage. Presented as an honest, leakage-free reliability result, fully reproducible on public data. Code and results are included in this record. This is an enhanced successor to the earlier preprint (Zenodo 10.5281/zenodo.17171318).