Enhancing the Spatial Resolution of Medium-Range Precipitation Forecasts Using Super-Resolution Neural Networks

Abstract Two prominent methods for accurate medium-range precipitation forecasts are European Centre for Medium-Range Weather Forecasts (ECMWF)’s numerical weather prediction model, high-resolution forecast (HRES), and a machine learning–based model, GraphCast. While the super-resolution generative adversarial network (SRGAN) has gained attention for enhancing the resolution of meteorological data, its application to medium-range forecast precipitation has not yet been explored. In this study, we developed the All-SR model to enhance the resolution of medium-range forecast precipitation over Japan from 0.25° to 0.05°, using four forecast meteorological variables: precipitation, temperature, mean sea level pressure, and vertically integrated water vapor transport. The All-SR model demonstrated statistically significant improvements over baseline bicubic interpolation. It reduced the mean absolute error by 11.1% for HRES and 8.2% for GraphCast, with these improvements being statistically significant across all 1–10 lead days. Furthermore, it significantly improved heavy rainfall prediction for GraphCast, increasing the threat score at the 50 mm (24 h) −1 threshold by 47.9% over the first 5 days. In contrast, improvements for HRES at the same threshold were only intermittently statistically significant. However, our analysis also identified key limitations: the model tended to underestimate heavy rainfall, and while it significantly improved the positional accuracy of heavy rainfall objects, it failed to consistently improve their shape or intensity. The paper concludes by discussing the underlying causes of these limitations and proposes promising directions for future development of more robust downscaling models. Significance Statement This study presents the first application of super-resolution generative adversarial networks (SRGANs) to enhance the resolution of medium-range forecast precipitation data. By testing various SRGAN models with different combinations of meteorological inputs, we found that the model incorporating all four variables—precipitation, temperature, mean sea level pressure, and vertically integrated water vapor transport—delivered the best performance, effectively downscaling forecasts from both high-resolution forecast (HRES) and GraphCast. Once trained, the model performs inference almost instantaneously, highlighting its strong potential for real-time forecasting applications. Although our research focuses on Japan, the same approach can be extended to other regions, provided that at least two decades of observed precipitation data are available for model training.