Hybrid precipitation downscaling over coastal watersheds in Japan using WRF and CNN

Study region: Kuma River Watershed in Japan. Study focus: High-quality precipitation information is desirable in hydrological modeling and water resources management. This study aimed to generate long-term fine-resolution precipitation datasets over the study region. A hybrid downscaling framework that integrates a dynamical approach by the Weather Research and Forecasting (WRF) model and a deep learning approach by the Convolutional Neural Network (CNN) model was proposed to derive precipitation information at fine resolutions from ERA-Interim datasets. The proposed hybrid downscaling framework was then applied to a coastal watershed in Japan. The merit of the hybrid downscaling approach in generating precipitation datasets at a 6-km resolution from 80-km ERA-Interim datasets, and 54-km and 18-km WRF simulated gridded datasets was explored as an alternative to pure dynamical downscaling approach by WRF. New hydrological insights for the region: The Nash-Sutcliffe efficiency coefficients of daily basinaveraged precipitation at 6-km resolution obtained by CNN from ERA-Interim, 54-km and 18-km WRF simulated datasets were 0.79, 0.93, and 0.98, respectively for training period; 0.71, 0.85, and 0.96, respectively for validation, when compared to 6-km WRF simulated gridded precipitation. The results demonstrated that CNN can reproduce 6-km WRF simulated precipitation and fine-resolution WRF modeling is needed to further enhance the downscaling performance, especially to capture spatial heterogeneity and extreme events. The hybrid downscaling framework of precipitation is promising to preserve the physics of atmospheric dynamics in precipitation modeling and reduce the computational cost considerably compared to pure dynamical downscaling.

Automated summary

The study focused on generating long-term fine-resolution precipitation datasets over the study region using a hybrid downscaling framework incorporating both dynamical and deep learning approaches. The results showed that the hybrid downscaling approach was more effective in generating precipitation datasets at a 6-km resolution compared to pure dynamical downscaling. The study demonstrated that the Convolutional Neural Network (CNN) could reproduce 6-km WRF simulated precipitation, emphasizing the importance of fine-resolution WRF modeling in enhancing downscaling performance. The hybrid downscaling framework was promising in preserving the physics of atmospheric dynamics in precipitation modeling and significantly reducing computational costs compared to pure dynamical downscaling.