Trade-offs in deep learning approaches for precipitation downscaling across weather regimes

Accurate high-resolution precipitation information is critical for assessing hydrological extremes and climate change impacts in complex topographic regions. While machine learning-based downscaling provides a computationally efficient alternative to dynamical models, its performance can depend on regional climate characteristics and terrain complexity. In this study, we systematically benchmarked representative deep learning approaches for daily statistical downscaling over Japan, including quantile-based and stochastic CNN formulations, two- (2D) and three-dimensional (3D) U-Net variants, and a diffusion-based generative model, alongside conventional baselines. Model performance was evaluated using comprehensive statistical and spatial metrics under both climatological and typhoon-affected conditions, enabling us to assess model robustness under extreme precipitation regimes. We further developed a proof-of-concept forecasting configuration for 3D U-Net prediction of next-day precipitation from short predictor sequences, together with an extreme super-resolution experiment to assess scalability to very high target resolutions. The results revealed clear trade-offs: quantile mapping best reproduced distributional properties, deterministic CNNs performed particularly well for standard deterministic metrics, the stochastic CNN achieved the highest average neighborhood-scale skill, and the diffusion model best reproduced the spatial power spectrum, while U-Nets tended to smooth small scales. Nevertheless, the asymmetric-loss U-Net provided the most robust performance during typhoon conditions. These findings provide practical guidance for model selection in complex climatic regions and illustrate how inductive biases interact with regional climate dynamics.