A scalable deep learning framework for daily precipitation downscaling: architecture, accuracy, and adaptability

ABSTRACT Global climate models provide essential large-scale climate projections, yet their coarse spatial resolution (0.5°–4°) limits understanding of urbanization's impact on regional climate dynamics. This study presents the Deep Residual Network for Precipitation Downscaling (DRN-PD), a modular neural architecture designed to enhance spatial precision and modeling interpretability in daily precipitation downscaling. By integrating CMIP6 atmospheric predictors, high-resolution land use data, and CHIRPS observations, DRN-PD advances understanding of climate–land surface interactions. Applied to the Yangtze Delta Megalopolis, the model achieves 18-fold spatial refinement, generating precipitation outputs at 5 km resolution. It consistently reproduces seasonal precipitation patterns, with mean absolute errors of 2.81–7.28 mm and root mean square errors below 11.28 mm. Spatial evaluations show strong agreement with observed rainfall distributions, achieving cosine similarity values exceeding 0.914 and effectively capturing rainfall centers and gradients. Incorporating a data augmentation strategy emphasizing extreme rainfall events, the model improves accuracy in high-intensity rainfall zones, with reduced relative errors and enhanced cosine similarity. These results demonstrate that DRN-PD, combined with appropriate training strategies, can effectively capture precipitation patterns across varying climatic and geographic contexts. This research contributes a practical tool for exploring climate–environment dynamics and supports evidence-based decision-making under increasing climate variability.