| Summary: | Deep learning techniques, especially convolutional neural networks (CNNs), have dramatically boosted the performance of statistical downscaling. In this study, we propose a CNN-based 2 m air temperature downscaling model named Terrain-Guided Attention Network (TGAN), which aims at rebuilding 2 m air temperature distribution from 0.0625<inline-formula><tex-math notation="LaTeX">$^{\circ }$</tex-math></inline-formula> to 0.01<inline-formula><tex-math notation="LaTeX">$^{\circ }$</tex-math></inline-formula> over Southwest China. More concretely, TGAN utilizes two upsampling modules to progressively reconstruct the high-resolution temperature data from the low-resolution one. Then, to better recover the spatial detail of the low-resolution temperature data, an attentive-terrain block is proposed to introduce digital terrain model (DEM) information. It aggregates the temperature data and the corresponding-scale DEM information via the attention mechanism in a multiscale manner. Ultimately, the reconstruction module is employed to obtain the high-resolution temperature data. We use the 2019 data for training, and utilize the 2018 data to verify the effectiveness of the proposed TGAN. The experimental results showed that TGAN achieved the lowest root-mean-square error (1.12<inline-formula><tex-math notation="LaTeX">$\,^\circ$</tex-math></inline-formula>C) when incorporating DEM data by attentive-terrain blocks in a multiscale manner, followed by incorporating DEM data in a multiscale manner (TGAN-land, 1.31<inline-formula><tex-math notation="LaTeX">$\,^\circ$</tex-math></inline-formula>C) and only incorporating DEM data (SRCNN-land, 1.36<inline-formula><tex-math notation="LaTeX">$\,^\circ$</tex-math></inline-formula>C). Meanwhile, TGAN showed a competitive performance when compared with several advanced deep-learning-based super-resolution algorithms and reconstructed the texture details of 2 m air temperature fields more clearly. In general, among various deep learning approaches, TGAN achieves better downscaling results for 2 m air temperature reconstruction and provides a practical method and guidance for the back-calculation of high-resolution historical meteorological grid data.
|
|---|