Improving Accuracy of Low-Altitude Three-Dimensional Wind Field Reconstruction via Multi-Task Learning

The demand for high-precision 3D wind fields in low-altitude weather services is increasing. To improve reconstruction accuracy, the high precision of 2D near-surface single-level products and the vertical structure of 3D numerical model outputs are combined to establish a mapping between 2D observations and 3D atmospheric fields, offering a new approach for 3D wind field reconstruction. A multi-task learning-based downscaling model, WindSD-3D, is proposed. Using single-level near-surface wind and terrain data as input, the model reconstructs high-resolution wind fields at multiple heights from 30 to 200 m. The architecture employs a shared feature extractor and two task-specific branches to decouple U and V wind components, and introduces cascaded Laplacian upsampling layers at the output to progressively recover high-frequency details. Ablation experiments and independent validation in the Beijing–Tianjin–Hebei region show that incorporating terrain data and a multi-stage upsampling strategy effectively enhances the model’s ability to reconstruct the vertical structure of wind fields. Compared to the numerical model product (GRAPES_MESO 3 km), WindSD-3D reduces errors at all evaluated heights, with RMSE improvements of 24.534–33.806% and bias reductions of 47.534–76.156%. The reconstructed wind fields exhibit more stable temporal error evolution across heights, with extreme errors suppressed and growing improvements at higher layers. At 120 m and 180 m, the correlation coefficients with observations increase by 44.989% and 43.913%, respectively, while preserving boundary layer vertical coupling. This method successfully extends the accuracy advantage of near-surface data into the vertical direction, offering a new technical approach for developing high-quality 3D wind field products.