Geospatially enhanced convolutional neural network for HY-2B scatterometer-based marine wind field retrieval

Precise monitoring of sea surface wind fields is imperative for understanding ocean-atmosphere interactions and extreme weather evolution, where spaceborne scatterometers provide unique observational capabilities but face inherent limitations in accuracy and efficiency using the geophysical model function (GMF). We developed convolutional neural network-based models (CNN-WSPD and CNN-WDIR respectively for sea surface wind speed and direction retrievals) integrating HY-2B scatterometer data and ECMWF reanalysis with geographical coordinates. The innovation lies in embedding positional features into multi-angle backscatter processing, enhancing physical consistency across diverse marine environments. Validation employs a global dataset with rigorous statistical comparisons against traditional GMF approaches. Proposed models reduce wind speed root mean square error (RMSE) by 9.7% (with 91.5% of samples exhibiting wind speed differences within 2 m/s) and wind direction RMSE by 1.6% (with the proportion of wind direction differences within 20° increasing to 90.5%), demonstrating superior global accuracy. The model presented in this study helps to revolutionize satellite scatterometer technology by advancing product development, enabling real-time high-resolution wind mapping to enhance typhoon forecasting accuracy, and pioneering next-generation designs that support climate resilience initiatives.