Deep learning-based statistical downscaling for short-term forecasting of summer air temperatures

Reliable early forecasting of summer air temperature is important to effectively prepare and mitigate damage such as heat-related mortality and excessive electricity demand caused by heat waves and tropical nights. Numerical weather prediction (NWP) models have been used for operational forecasting of air temperature. However, NWP models have coarse spatial resolution due to massive computational resources arising from complex forecasting systems and unstable parameterization of NWP models, which make the uncertainty of prediction, consisting of systematic and random biases. Therefore, the objective of this study is to develop a novel deep learning-based statistical downscaling approach for the Global Data Assimilation and Prediction System (GDAPS) model’s summer air temperature forecasts over South Korea. This study developed the proposed statistical downscaling model through the decomposition into the temporal dynamics of daily air temperature forecast and spatial fluctuation by pixels. The daily temperature dynamic was estimated using a daily mean GDAPS temperature forecast with simple mean bias correction. The spatial fluctuation by pixels was obtained using the spatial anomaly of downscaled air temperature forecast by the U-Net model. The GDAPS model’s forecast data, present-day high spatial resolution satellite observations, and topography variables were used as input variables for training the U-Net model. The observations at weather stations were spatially interpolated using the regression-kriging, and then we used it as a target image for the U-Net model. The proposed U-net model was compared with the Local Data Assimilation and Prediction System (LDAPS), the dynamically downscaled model of the GDAPS, and the support vector regression (SVR)-based statistical downscaling model. For next-day Tmax and Tmin forecasts, the suggested U-net model showed better performance, having high coefficient of determination (R²) of 0.76 and 0.74 and root mean square error (RMSE) of 2.5 °C and 1.5 °C for next-day Tmax and Tmin forecasts, respectively. When analyzing the skill score (SS) values by stations of the U-Net model, it had remarkably high SS values at stations where the GDAPS had a high absolute value. For Tmax and Tmin forecasts with 1-7 days forecast lead time, the suggested model consistently provided better performance (higher spatial correlation and lower RMSE) than GDAPS and SVR. In addition, the U-net model showed a detailed spatial distribution most similar to that of the observations. These results demonstrated that the suggested model successfully corrected the bias of the GDAPS, improving not only the forecast accuracy but also the ability to capture the spatial distribution of Tmax and Tmin forecasts. Using the deep learning-based suggested model in this study, bias-corrected high spatial resolution air temperature forecasts with a relatively long forecast lead time in summer seasons can be successfully produced.