A new statistical downscaling approach for short-term forecasting of summer air temperatures through a fusion of deep learning and spatial interpolation

Reliable early forecasting of extreme summer air temperatures is essential for effectively managing and mitigating the socioeconomic damage caused by thermal disasters. Numerical weather prediction models have become valuable tools for forecasting air temperature; however, they incur high computational costs, resulting in coarse spatial resolution and systematic bias owing to imperfect parameterization. To address these problems, we developed a novel statistical downscaling and bias correction method (named DeU-Net) for the maximum and minimum air temperature (Tmax and Tmin, respectively) forecasts obtained from the Global Data Assimilation and Prediction System with a spatial resolution of 10 km to 1.5 km over South Korea through the fusion of deep learning (i.e., U-Net) and spatial interpolation. In this study, we used a methodology to decompose statistically downscaled Tmax and Tmin forecasts into temporal dynamics over South Korea and spatial fluctuations by pixels. When comparing the proposed DeU-Net with the dynamical downscaling model (i.e., Local Data Assimilation and Prediction System) and support vector regression (SVR)-based statistical downscaling model at the seen and unseen stations for forecasting the next-day Tmax and Tmin, respectively, DeU-Net showed the highest spatial correlation and the lowest RMSE in all cases. In a qualitative evaluation, DeU-Net successfully produced a detailed spatial distribution most similar to the observations. A further comparison extending the forecast lead time to seven days indicated that the proposed DeU-Net is a better downscaling approach than SVR, regardless of the forecast lead time. These results demonstrate that bias-corrected high spatial resolution air temperature forecasts with relatively long forecast lead times in summer can be effectively produced using the proposed model for operational forecasting.