Forecast calibrations of surface air temperature over Xinjiang based on U-net neural network

Abstract

In this study, a deep learning method named U-net neural network is utilized to calibrate the gridded forecast of surface air temperature from the Global Ensemble Forecasting System (GEFS), with forecast lead times of 1–7 days in Xinjiang. The calibration performance of U-net is compared with three conventional postprocessing methods: unary linear regression (ULR), the decaying averaging method (DAM) and Quantile Mapping (QM). Results show that biases of the raw GEFS forecasts are mainly distributed in the Altai Mountains, the Junggar Basin, the Tarim Basin and the Kunlun Mountains. The four postprocessing methods effectively improve the forecast skills for all lead times, whereas U-net shows the best correction performance with the lowest mean absolute error (MAE) and the highest hit rate of 2°C (HR2) and pattern correlation coefficient (PCC). The U-net model considerably reduces the warm biases of the raw forecasts. The skill improvement magnitudes are greater in southern than northern Xinjiang, showing a higher mean absolute error skill score (MAESS). Furthermore, in order to distinguish the error sources of each forecasting scheme and to reveal their capabilities of calibrating errors of different sources, the error decomposition analysis is carried out based on the mean square errors. It shows that the bias term is the leading source of error in the raw forecasts, and barely changes as the lead time increases, which is mainly distributed in Tarim Basin and Kunlun Mountains. All four forecast calibrations effectively reduce the bias and distribution error of the raw forecasts, but only the U-net significantly reduces the sequence error.