HT-Unet: Integrating Hadamard Attention and Transformer for Improved Speed and Accuracy in SO2 Assimilation

Data assimilation (DA) plays a crucial role in improving weather forecasts by integrating observational data into initial background fields. However, its computational efficiency remains a significant challenge. Deep learning (DL) offers promising advancements in data assimilation (DA) by effectively leveraging its ability to capture complex nonlinear relationships and extract spatiotemporal features, enhancing the integration of observational and model-generated data. This paper proposes a novel approach that combines three-dimensional variational (3D-Var) assimilation with a DL method, integrating the Hadamard attention mechanism and Transformer modules into the Unet framework (HT-Unet) to assimilate sulfur dioxide (SO2) and improve forecasting accuracy. The HT-Unet utilizes the results of 3D-Var assimilation as labels to integrate observational data and three-dimensional model background fields, generating high-quality analysis fields. The Hadamard attention effectively captures key features with low computational cost, while the Transformer module captures long-distance dependencies to further enhance assimilation accuracy. Experimental results demonstrate that the HT-Unet method significantly improves assimilation speed, processing a single analysis field 34 times faster than the 3D-Var method. In terms of assimilation accuracy, the HT-Unet method reduces RMSE by 41% compared to the basic Unet model. This new method shows great potential to replace the traditional 3D-Var method, overcoming its high computational costs and significantly enhancing air pollution forecasting accuracy.