Reconstructing Marine Boundary Layer Jets Over the South China Sea Using Machine Learning

Abstract Marine boundary layer jets (MBLJs) over the South China Sea (SCS) play a critical role in coastal heavy rainfall, yet their structure remains uncertain due to limited wind observations over the ocean. Here, we develop a U‐Net–based reconstruction framework to estimate 950‐hPa winds from satellite‐derived sea surface winds, providing a satellite‐driven depiction of MBLJs. The model is trained using four decades of ERA5 data and evaluated against independent shipborne Doppler wind LiDAR observations. Relative to ERA5, the reconstructed winds reduce RMSE from 3.34 to 2.69 m s −1 for wind speed and from 2.87 to 2.45 m s −1 for meridional wind. During identified MBLJ events, the reconstruction produces stronger and more spatially extensive jets than ERA5. These results demonstrate that satellite‐informed machine learning can effectively mitigate systematic underestimation of MBLJs in reanalysis products and improve representation of boundary‐layer winds over data‐sparse oceanic regions.