Leveraging Earth Embeddings for Generalizable Precipitation Downscaling Across Geographies

As climate change amplifies precipitation extremes and their societal and economic impacts, downscaling precipitation provides valuable local-scale information for risk assessment and adaptation planning.
However, deep-learning based statistical downscaling methods typically rely on high-resolution training data (e.g., radar observations), which are scarce and unevenly distributed globally, making geographic generalization a central challenge. Prior work shows large performance drops of deep-learning based downscaling models under geographic distribution shifts -- effects that remain even when considerably increasing the training data [1].
We view the geographic distribution shift as a form of subpopulation shift, where training and target samples are drawn from the same set of geographic domains but differ in their sampling frequencies. Consequently, the shift is driven primarily by changes in the prevalence of climatic regimes, rather than by changes in the conditional relationship between predictors and targets.
To improve robustness under cross-region transfer, we inject additional geographic context through Earth embeddings from geospatial foundation models (e.g., SatCLIP [2]). Potential strategies for integrating these embeddings into diffusion-based downscaling models include attention-based conditioning, feature modulation, and auxiliary conditioning networks.

[1] Harder, P., Schmidt, L., Pelletier, F., Ludwig, N., Chantry, M., Lessig, C., Hernandez-Garcia, A. and
Rolnick, D. [2025], ‘Rainshift: A benchmark for precipitation downscaling across geographies’, arXiv
preprint arXiv:2507.04930 .

[2] Klemmer, K., Rolf, E., Robinson, C., Mackey, L. and Rußwurm, M. [2025], Satclip: Global, general-
purpose location embeddings with satellite imagery, in ‘Proceedings of the AAAI Conference on Artificial
Intelligence’, Vol. 39, pp. 4347–4355.