A Bagged-Tree Machine Learning Model for High and Low Wind Speed Ocean Wind Retrieval from CYGNSS Measurements

This paper presents two empirical models, the low wind bagged trees (LWBT) and high wind bagged trees (HWBT) ensemble models to estimate ocean surface wind speed using spaceborne Global Navigation Satellite System Reflectometry (GNSS-R) data. The models are empirically trained using NASA's Cyclone GNSS (CYGNSS) mission level 1 data (version 2.1). The truth label for the LWBT model is the wind speed product derived from European Centre for Medium-Range Weather Forecasts (ECMWF) ERA-5 and Global Data Assimilation System (GDAS), while the label for the HWBT model is wind speed measurements from Stepped Frequency Microwave Radiometer (SFMR). Testing results show that the LWBT and HWBT models achieved global wind speed retrieval root-mean-square-error (RMSE) of ~1.5 m/s and ~1.4 m/s, respectively, corresponding to an improvement of 29% and 65% with respect to the CYGNSS Level 2 standard wind speed product. The maximum bias is reduced by 65% and 60% for LWBT and HWBT over the Level 2 wind speeds, respectively. Two typhoon case studies are presented to corroborate the model performances and their retrieved wind speeds are consistent with reports from World Meteorological Organization (WMO) and with the measurement provided by the Huangmao Zhou (HMZ) weather station.