Efficient precipitation downscaling over the Southern Tibetan Plateau with deep learning surrogates

Sub-daily high-resolution precipitation products are crucial for improving mitigation and adaptation of natural hazards. However, such products are often lacking in complex terrains, such as the Tibetan Plateau (TP). While convection-permitting models can accurately capture kilometer-scale sub-daily convective precipitation processes, they are computationally expensive. Here we propose a high-resolution Hourly Precipitation Downscaling Network (HPDNet), a deep-learning surrogate designed to approximate the downscaling behaviors of the Weather Research and Forecasting (WRF) model. HPDNet was trained using 10-year WRF simulations (9 km to 1 km) over the southeastern TP (∼0.11 million km2). To evaluate its transferability, the trained model was applied to two independent products, ERA5-Land and IMERG, without retraining. Results indicate that downscaled IMERG exhibits higher accuracy than downscaled ERA5-Land, consistent with the superior initial quality of raw IMERG observations (bias: 0.86 mm/d) relative to ERA5-Land (bias: 2.66 mm/d). In the training domain, downscaled IMERG achieves an approximately 22.8% reduction in RMSE and a 20.0% improvement in the Critical Success Index (CSI). Furthermore, the model was successfully extended to the entire southern TP (∼1.38 million km2), where downscaled IMERG maintains a 16.2% RMSE reduction, stemming from the model's ability to capture fundamental physical drivers. Interpretability analysis reveals that HPDNet's spatial attention concentrates on steep elevation gradients—the primary drivers of orographic precipitation. This focus allows the model to recover fine-scale spatial heterogeneities and orographic enhancement signals that raw products typically smooth out. HPDNet facilitates the efficient generation of multi-decadal, 1-km hourly precipitation estimates across the complex terrain of the southern TP.