Abstract
This study presents a novel machine learning algorithm for wind speed retrieval using the Haiyang-2B (HY-2B) altimeter. To construct a high-fidelity reference dataset, a wind field over the Northwest Pacific Ocean with a spatial resolution of 0.05° and an hourly temporal resolution was hindcast by the Weather Research and Forecasting (WRF) model for the period from 2022 to 2024. Data from multiple active and passive sensors, specifically the Haiyang-2 (HY-2) constellation, the advanced scatterometer (ASCAT), the soil moisture active passive (SMAP) radiometer and the Advanced Microwave Scanning Radiometer 2 (AMSR-2), were assimilated into the WRF simulation. Validation of the atmospheric simulations against moored buoy observations yielded a root mean square error (RMSE) of 2.06 m/s, a correlation coefficient (r) of 0.73 and a scatter index (SI) of 0.42. Subsequently, the normalized radar cross section (NRCS) and significant wave height (SWH) from the HY-2B altimeter were collocated with the simulated wind speeds to train four machine learning models, namely the light gradient boosting machine (LightGBM), random forest (RF), eXtreme Gradient Boosting (XGBoost) and residual network (ResNet). The ResNet model demonstrated superior performance, achieving an RMSE of 1.60 m/s, an r of 0.86 and an SI of 0.32. This significantly outperforms the operational HY-2B product, which exhibited an RMSE of 2.52 m/s, an r of 0.69 and an SI of 0.50 when validated against independent moored buoy observations in 2025. These results indicate that the proposed algorithm substantially enhances the accuracy of altimeter-derived wind speeds. Furthermore, when applied specifically to tropical cyclones (TCs), the method demonstrated robust capabilities by achieving an RMSE of 4.72 m/s, an r of 0.84 and an SI of 0.44.