Generative Adversarial Networks for Downscaling Hourly Precipitation in the Canadian Prairies

Abstract Developing robust downscaling methods is essential for maximizing the applicability of climate model outputs in engineering design and climate mitigation, particularly in a changing climate. This study evaluates four deep learning model configurations for downscaling, focusing on their structure, functionality, and ability to capture localized convective events in the Canadian prairies. These model configurations aim to downscale coarse‐resolution climate model outputs (∼200 km) to the finer spatial resolution of regional climate models (∼50 km) for hourly precipitation. We introduce advanced metrics to assess the fidelity of precipitation downscaling, examining both marginal statistics and spatiotemporal dependencies. A U‐Network (UNET) captures spatial and temporal dependencies efficiently while three generative adversarial networks (GANs) configurations incorporate a critic network to enhance the realism of generated fields. The study also evaluates the effects of a thresholding layer to constrain precipitation values and a convolution long short‐term memory layer in the GAN critic to better capture temporal dependencies. Results show that all four model configurations effectively capture spatial dependencies, with the simplest GAN architecture outperforming others in preserving temporal dependencies. Latitudinal correlations are better preserved than longitudinal across all models. While UNET produces overly smoothed fields, GANs generate more detailed outputs when downscaling Coupled Model Intercomparison Project phase 6 projections. By optimizing deep learning models for this region, the study provides key insights into future precipitation trends, enabling the identification of localized storms. These findings are critical for improving infrastructure resilience across catchments in the prairies.