Near-surface weather variables play an important role in production. However, the coarse spatial resolution meteorological data in existing meteorological products limit their applicability in modern high-precision engineering and decision-making scenarios. To address this limitation, we propose a deep learning-based meteorological downscaling framework, Hybrid Spatial-Field Attention Network (HSFANet). HSFANet is built upon a hybrid spatial-field attention module that jointly models local spatial dependencies, global spatial correlations, and cross-field interactions among different meteorological variables. To account for the varying importance of hierarchical features, a dynamic layer information integration module is introduced to adaptively aggregate multi-level representations. Furthermore, the ground-object cross-attention mechanism is incorporated to capture the coupling relationship between meteorological fields and underlying surface characteristics. Experiments on CLDAS V2.0 reanalysis data demonstrate that the proposed method achieves higher accuracy than state-of-the-art methods across multiple scaling factors. In addition, the engineering application experiment shows that HSFANet can effectively reconstruct future high-resolution meteorological fields from coarse-resolution forecast data, highlighting its potential for practical deployment in engineering and production-related applications.
Gao et al. (Tue,) studied this question.