An improved deep convolutional residual network for daily precipitation downscaling and future projections in the Western source region of the South-to-North water diversion project

Downscaling is a critical step to bridge the gap between large-scale climate information and local-scale impact assessments. However, steep and complex mountainous terrain induces highly heterogeneous local-scale meteorological conditions, posing significant challenges for accurately downscaling meteorological variables, particularly daily precipitation, which exhibits strong nonlinearity. To improve the daily precipitation downscaling skills in mountainous regions, this study presents a novel deep learning approach: Deep Convolutional Residual Network (DCRN). This approach was constructed based on a deep convolutional neural network with residual blocks and Gamma-distribution-based two-stage loss function combining a classifier for rainfall occurrence and a regression for rainfall amount. The illustrative cases of downscaling daily precipitation from coarse resolutions (2.8° × 2.8°) to fine resolution (0.5° × 0.5°) in mountainous catchments in the Dadu River and Yalong River basins demonstrate that the improved DCRN has enhanced downscaling accuracy and captured the spatial and temporal patterns remarkably well. Furthermore, the downscaled future precipitation projections indicate that the study basins are likely to experience wetter conditions over the next 80 years. The results provide improved high-resolution precipitation information, offering valuable support for policy-making in water resource management and adaptive operation and planning of water diversion projects under future climate change.