Global Location Transferability of Generative Deep Learning Models for Precipitation Downscaling

Generative deep learning models have shown remarkable skill in the probabilistic downscaling of climate and weather forecasts, with generative adversarial networks (GANs) as a particularly effective approach for precipitation downscaling. However, most existing methods are trained for specific regions, and their performance on unseen geographic areas remains largely unexplored. In our work, we evaluate the transferability of generative models to new locations outside their training domain. Using a global experimental setup, we employ ERA5 as the predictor dataset and IMERG as the high-resolution target dataset at 0.1° resolution. To systematically assess the performance across diverse regions, we design a hierarchical location split with 16 regions. We then train networks independently on the 16 regions and evaluate each of them on all others. Our findings provide insights on the robustness and limitations of generative models for global-scale precipitation downscaling, revealing challenges such as poor generalization to unseen orography and decreased performance in tropical regions, both for models applied in these areas and for those trained in the tropics and transferred elsewhere.