On Global Applicability and Location Transferability of Generative Deep Learning Models for Precipitation Downscaling

TL;DR

This study tests whether generative precipitation downscaling models trained in some regions can generalize to unseen geographic areas using a global ERA5-to-IMERG task. Employing a Wasserstein GAN with a ResNet-based generator and static geographic features, the authors train 15 regional models and evaluate cross-region transfer across nine subregions, focusing on the role of orography. Results show substantial improvements over ERA5 baselines in most transfers (23–42% CRPS reduction), but reveal notable failures in high-relief and tropical-to-extratropical transfers, highlighting out-of-distribution challenges. The findings offer practical guidance on regional training coverage and feature inclusion for global downscaling applications and point toward future work with diffusion models and alternative data sources to enhance transferability.

Abstract

Deep learning offers promising capabilities for the statistical downscaling of climate and weather forecasts, with generative approaches showing particular success in capturing fine-scale precipitation patterns. However, most existing models are region-specific, and their ability to generalize to unseen geographic areas remains largely unexplored. In this study, we evaluate the generalization performance of generative downscaling models across diverse regions. Using a global framework, we employ ERA5 reanalysis data as predictors and IMERG precipitation estimates at $0.1^\circ$ resolution as targets. A hierarchical location-based data split enables a systematic assessment of model performance across 15 regions around the world.

Figures (6)

• Figure 1: Spatial domain and regional partitioning of the dataset. The covered area extends from $60^\circ$N to $60^\circ$S and $130^\circ$W to $170^\circ$E. We first divide the domain into nine rectangular subregions, which are then combined into larger regions (N, T, S, W, M, E) to create 15 training areas in total.
• Figure 2: CRPS scores for all training–evaluation region combinations. Rows correspond to training regions, columns to evaluation regions. Lower scores indicate better performance. The left column shows the overall CRPS performance.
• Figure 3: CRPS score maps for three representative training regions: North (N), Tropics (T), and South (S). Left: raw CRPS. Middle: relative improvement over ERA5. Right: relative performance drop compared to direct training on the target region.
• Figure 4: As in Figure \ref{['nts_maps']}, but for a hierarchical training setup: starting with Northwest (NW) and progressively enlarging the training domain. Enlarging the domain generally improves transfer performance, especially when the target region is included.
• Figure 5: The logarithm of the CRPS score is shown globally, averaged over the whole test period. On the left it shows the performance of the GAN trained on SM on the right trained on SE.