A Spatial Downscaling Approach for Enhanced Accuracy in High Wind Speed Estimation Using Hybrid Attention Transformers

Due to the lack of high-coverage regional ground observation data in China before 2008, the China Meteorological Administration's Land Data Assimilation System (CLDAS) faces challenges in directly generating high-resolution and high-quality land assimilation products prior to that year.To address this issue, this paper proposes a deep learning model, HAT(DEM+RMSE_MR+ERROR), which is based on the Hybrid Attention Transformer (HAT). This model aims to improve the downscaling accuracy of high winds in the CLDAS2.0 10m wind field from 6.25 km to 1 km by incorporating digital elevation information (DEM), enhancing the loss function (RMSE_MR), and employing a prediction error method (ERROR). We utilized data from 2020-2021 for training and validation, and data from 2019 for testing, conducting ablation experiments to verify the effectiveness of each module, while comparing the results with those of the traditional bilinear interpolation method and the UNET_DCA model. The experimental results indicate that when assessing by wind speed categories, HAT(DEM) performs best for wind speeds below 3 Beaufort, while HAT(DEM+RMSE_MR+ERROR) excels for wind speeds above 4 Beaufort. Specifically, for wind speeds above 6 Beaufort, HAT(DEM+RMSE_MR+ERROR) achieves MAE(mean absolute error), POD(probability of detection), and TS(threat score) metrics of 0.825 m·s-1, 0.813, and 0.607, respectively, when evaluated against CLDAS3.0 as the ground truth. The model performs better in March to May and November, while its performance is weakest from June to August; it also performs better during the day than at night, and shows suboptimal performance in plains. The model is closer to the ground truth in reconstructing the structural details of wind fields and outperforms the annual average during most high wind weather events, indicating better predictive capability and adaptability for such events. Overall, the proposed HAT(DEM+RMSE_MR+ERROR) model demonstrates significant progress in downscaling predictions of high winds and provides deeper insights into the inversion of high-resolution historical meteorological grid data.