Local spatial self-attention based deep network for meteorological data downscaling

High-resolution meteorological data are essential for simulation and decision-making in weather-sensitive industries such as agriculture and forestry. However, existing meteorological products typically have low spatial resolution (coarser than 0.1°), making it difficult to capture the fine-grained spatial distribution of meteorological variables. Most existing deep learning-based downscaling methods treat the task as an image super-resolution problem, overlooking key characteristics of meteorological data, such as multi-scale local spatial correlation, local–global spatial dependency, and the complex relationship between terrain and meteorological fields, thus limiting modeling accuracy. To address this issue, this paper proposes a deep neural network based on local spatial self-attention, LSSANet, for the spatial downscaling of meteorological data. Specifically, the Local Spatial Self-Attention Module (LSAM) is proposed to capture local–global spatial correlations of meteorological fields. The Multi-scale Dynamic Aggregation Module (MDAM) is introduced to handle multi-scale local spatial dependencies. Furthermore, an elevation embedding layer and a two-stage training strategy are developed to integrate the relationship between terrain and the meteorological field. Experimental results show that LSSANet achieves superior performance compared to traditional and state-of-the-art methods. In the 4$\times$ downscaling task, LSSANet reduces MAE by 5.1%–75.8%; in the 8$\times$ task, by 4.3%–59.7%; and in the 16$\times$ task, by 1.9%–53.4%. Engineering application experiments further demonstrate that the proposed method can generate high-resolution future meteorological forecasts based on the GFS product. These results indicate that LSSANet can accurately reconstruct or predict high-resolution meteorological fields in specific regions, providing valuable support for planning and decision-making in meteorology-sensitive industries.