Retrieval of Sea Surface Wind Speed From CYGNSS Data in Tropical Cyclone Conditions Using Physics-Guided Artificial Neural Network and Storm-Centric Coordinate Information

Abstract

A novel artificial neural network (ANN) model is introduced for the retrieval of tropical cyclone (TC) sea surface wind speed from the Cyclone Global Navigation Satellite System (CYGNSS) Level 1 data. WindSat TC wind data serves as the “truth” information for the ANN training. Compared to conventional machine learning approaches, the proposed model incorporates specialized information including the storm-centric coordinate information of CYGNSS observations and physical-guided scattering azimuth angle. In addition, a feature selection process is employed, utilizing both XGBoost regressor and Pearson correlation coefficient, to identify the most pertinent input variables for wind speed retrieval. The results show that the proposed model with storm-centric coordinate information and first-order cosine form of scattering azimuth angle as additional inputs demonstrates good retrieval performance. It achieves a bias of -0.59 m/s and an root mean square error (RMSE) of 3.43 m/s, corresponding to a decrease of 60.93% and 20.05% compared to the current CYGNSS baseline wind products for young seas with limited fetch (YSLF). Especially above 35 m/s, the proposed model outperforms the CYGNSS YSLF product, illustrating its advantages under high wind speeds. Moreover, the effects of two special inputs on the model performance are explored. It is found that the RMSEs are reduced by about 25.43% and 9.50%, respectively, after incorporating the two specific inputs, suggesting that considering TC-related inputs is more effective than the physics-guided initialization in improving model performance. Our retrieval results provide valuable guidance for improving the use of GNSS-R data for near real-time retrieval of TC winds.