Machine Learning-Based Algorithm for Gaofen-3 Wind Vector Retrieval With Assimilation by COAWST in China Seas

This study develops a novel machine learning-based algorithm for retrieving wind vectors from Gaofen-3 (GF-3) synthetic aperture radar (SAR) imagery. Additionally, we explore the assimilation of SAR-derived wind fields with the Coupled Ocean-Atmosphere-Wave-Sediment Transport (COAWST) model. Our dataset comprises 3000 GF-3 quad-polarized stripmap (QPS) images collected over the China Seas during the past eight years, collocated with European Centre for Medium-Range Weather Forecasts (ECMWF) data. The proposed methodology utilizes SAR intensity spectra and polarimetric correlation coefficients (PCCs) between dual-polarization channels (HH/HV) for wind direction retrieval using eXtreme Gradient Boosting (XGBoost). For wind speed retrieval, the XGBoost model incorporates co-polarized normalized radar cross section (NRCS), retrieved wind direction, incidence angle, and azimuth angle as input features. Validation of the wind vector against the measurements from Haiyang-2 (HY-2) measurements confirms the robustness of XGBoost-based algorithms, i.e., a root mean squared error (RMSE) of 22.99° for wind direction with a correlation (COR) of 0.97 and a scatter index (SI) of 0.20°, and an RMSE of 1.26 m/s for wind speed with a COR of 0.89 and a SI of 2.26 m/s. The XGBoost-retrieved SAR wind vectors derived from 180 fine stripmap (FS) images were applied to assimilate COAWST ocean models. Assimilation using wind components (u, v) outperformed using wind speed/direction (U, ϕ) and non-assimilation, validated against HY-2 measurements. The (u, v)-based method achieved superior accuracy for wind speed (RMSE 1.69 m/s) and significant wave height (SWH) (RMSE 0.41 m), which showed improvements in wind speed and SWH of 0.15 m/s and 0.1 m respectively, compared to the (U, ϕ)-based assimilation, demonstrating SAR's effectiveness in refining ocean simulations.