Dual-Stream Downscaling and Flow-Constrained Correction of Meteorological Fields over Complex Terrain

High-resolution near-surface meteorological fields are essential, yet conventional numerical simulation is computationally expensive and data-driven methods often lack physical consistency. Here, a lightweight framework for meteorological-field downscaling and observation-driven correction is presented as a cost-effective alternative. In the downscaling stage, multivariate synoptic-scale information is fused with high-resolution terrain by a dual-stream generative adversarial network, and the fields are refined in a reconstruction-driven cascade. In the correction stage, a simplified UNet learns additive corrections supervised directly by station observations, and a flow-constrained regularizer enforces physical consistency. In independent tests, the downscaling stage reduces the 0.25° four-variable error by 54.0–73.5% relative to bicubic interpolation, and the full-field correction reduces the station MAE of T2m, V10, and U10 by 17.81%, 27.59%, and 21.06%, respectively. The full training-to-inference workflow can be completed within a few days on an ordinary personal computer, making it practical for operational fine-scale meteorological services.