AMV-STECNet: A Deep Learning Framework for Spatiotemporal Error Correction of Atmospheric Motion Vectors to Enhance Numerical Weather Prediction

Atmospheric motion vectors (AMVs), vital for upper-level wind observations in numerical weather prediction, face assimilation challenges due to inherent retrieval errors. This study develops a AMVs Spatiotemporal Error Correction Network (AMV-STECNet) that employs multi-task learning to address both temporal and spatial error characteristics. The architecture incorporates three innovative components: 1) Enhanced UNet with attention mechanisms for spatial feature extraction; 2) Transformer-based temporal module capturing multi-scale circulation patterns; 3) Adaptive fusion for spatiotemporal error correction. The model reduces U-component RMSE by 38% (5.82→3.59), outperforming ConvLSTM and UNet while improving fluid consistency by 10%. Assimilation experiments using WRFDA and YH4DVAR systems during Typhoon Nesat (2022) reveal 20% track error reduction in later forecast stages, particularly showing improved agreement with observed trajectories. The corrected AMVs data demonstrates operational forecasting enhancements through a 10% reduction in 24-hour upper wind RMSE and greater than 15% error decrease in surface elements, quantitatively validating its implementation value for numerical weather prediction systems.