Downscaling of the surface temperature forecasts based on deep learning approaches

Accurate high-resolution temperature forecasting is of great significance for the economic and social development of humanity. Due to the chaotic nature of the atmosphere and the limitations of computational resources, model forecasts often lack sufficient resolution and exhibit systematic biases. Therefore, downscaling methods with smaller computational demands have become a good alternative. This study designed a super resolution generative adversarial network (SRGAN) for temperature downscaling, applying it to the 2 m temperature forecasts for the Southwest region of China from the Global Ensemble Forecasting System (GEFS), with forecast lead times of 1 to 7 days. Meanwhile, linear regression (LR), along with two advanced deep learning downscaling methods, U-Net and super resolution deep residual networks (SRDRNs), were also used as benchmarks. The study shows that both deep learning methods, SRGAN and SRDRNs, can effectively address the issue of blurred temperature fields that may occur when using U-Net. By comparing the Nash-Sutcliffe Efficiency coefficient (NSE), pattern correlation coefficient (PCC), root mean square error (RMSE), and peak signal-to-noise ratio (PSNR), we found that SRGAN demonstrated the best performance among the four methods. In this work, a suitable loss function was set using the VGG network to help SRGAN better capture small-scale details. Additionally, a mean square error decomposition method was used to further diagnose the sources of errors in different models, revealing their ability to calibrate various error sources. The results show that SRGAN, SRDRNs, and LR perform best in correcting the square of the bias (Bias²), while U-Net is most effective in correcting the sequence errors.