Deep Learning for Daily 2-m Temperature Downscaling

This study proposes a novel method, which is a U-shaped convolutional neural network that combines non-local attention mechanisms, Res2net residual modules, and terrain information (UNR-Net). The original U-Net method and the linear regression (LR) method are conducted as benchmarks. Generally, the UNR-Net has demonstrated promise in performing a 10x downscaling for daily 2-m temperature over North China with lead times of 1-7 days and shows superiority to the U-Net and LR methods. To be specific, U-Net and UNR-Net demonstrate higher Nash-Sutcliffe Efficiency coefficient values compared to LR by 0.052 and 0.077, respectively. The corresponding improvements in pattern correlation coefficient are 0.013 and 0.016, while the root mean square error values are higher by 0.22 and 0.338, respectively. Additionally, the structural similarity index metric is higher by 0.033 and lower by 0.015. Furthermore, regions with significant errors are primarily distributed in complex terrain areas such as the Taihang Mountains, where UNR-Net exhibits noticeable improvements. In addition, the 12 components-based error decomposition method is proposed to analyze the error source of different models. Generally, the smallest errors are observed during the summer season and the sequence error component is proven to be the main source error of 2-m temperature forecasts. Furthermore, UNR-Net consistently demonstrates the lowest errors among all 12 error components. Therefore, combining the numerical weather prediction model and deep learning method is very promising in downscaling temperature forecasts and can be applied to routine forecasting of other atmospheric variables in the future. This research proposes a new method for downscaling using deep learning. The method uses a specific type of neural network called UNR-Net, which combines attention mechanisms, residual modules, and terrain information. The performance of UNR-Net is compared to two other methods: U-Net and LR. In the study, UNR-Net shows promise in performing a 10x downscaling of the daily 2-m temperature in North China. The UNR-Net demonstrates the best overall performance among all the comprehensive indicators (NSE, pattern correlation coefficient, root mean square error, and structural similarity index metric). Errors in the predictions are mainly found in complex terrain areas like the Taihang Mountains, but UNR-Net shows noticeable improvements in these regions. The study also proposes a 12 components-based error decomposition method to analyze the error sources of different models. All in all, it is found that the smallest errors are observed during the summer season and the main source error is the sequence error component. Additionally, when considering lead times of 1-7 days, UNR-Net consistently shows the lowest errors among all 12 error components. Based on these findings, combining numerical weather prediction models with deep learning methods holds great promise for generating high-resolution temperature forecasts. This paper presents a novel deep learning downscaling method, UNR-Net, capable of downscaling daily 2-m temperature by a factor of 10 The overall performance of the UNR-Net method surpasses the U-Net method and linear regression method The 12 components-based error decomposition method is proposed to analyze the error source of different models