**5. Conclusions**

By using machine learning methods, this study investigated wind speed retrieval in different wind speed intervals. Through extensive processing of experimental data, it was observed that different machine learning methods have different properties in different wind speed intervals. In particular, a range of multi-variable models was developed and evaluated. The results showed that the LGBM model performs best with an RMSE of 1.419 m/s and a correlation coefficient of 0.849 in the low wind speed interval (0–15 m/s), while the ET model performs best with an RMSE of 1.100 and a correlation coefficient of 0.767 in the high wind speed interval (15–30 m/s). In addition, through experiments, some characteristics of ANN models were found in wind speed retrieval. In the low wind speed interval, the choice of activation function hardly affects the ANN models, while the increase of the number of neurons significantly reduces the accuracy of the model. In the high wind speed interval, the increase in the number of neurons has little effect on Sigmoid and Tanh, but it has an obvious effect on ReLu.

The effects of the variables used in the wind speed retrieval models described in this paper were analyzed. Through processing experimental data, it was observed that the models with all variables (i.e. NBRCS, LES, SNR, DDMA, Noise Floor, sp\_inc\_angle, sp\_az\_body, Instrument Gain, Scatter Area, and SWH\_swell) achieved the highest accuracy. In the low wind speed interval, NBRCS, LES, SNR and SWH\_swell were the most important variables. In the high wind speed interval, the models were mostly affected by SWH\_swell, and the ranking of the effects of variables was very different from that in the low wind speed interval.

Future studies will focus on further performance enhancements of the models developed in this paper. For instance, the accuracy of the model would decrease in the presence of large wind speed and high SWH\_swell. It would thus be useful to develop techniques to handle the retrieval of high wind speeds with minimal performance degradation.

**Author Contributions:** All authors have made significant contributions to this manuscript. C.W. constructed a part of machine learning models of this paper, analyzed the data, wrote the initial version of paper and validated all the models; K.Y. conceived the improved method, wrote the revised version of the paper and provided supervision; F.Q. constructed some of the machine learning models used in this paper; J.B., S.H. and K.Z. checked and revised this paper. All authors have read and agreed to the published version of the manuscript.

**Funding:** This work was supported in part by the National Natural Science Foundation of China under Grants 42174022 and in part by the Programme of Introducing Talents of Discipline to Universities, Plan 111, Grant No. B20046.

**Data Availability Statement:** Not applicable.

**Acknowledgments:** We would like to thank the NASA and European Center for Medium-Range Weather Forecasts (ECMWF) for providing the data. The results (figures and tables) presented in this article are mainly generated by MATLAB software (https://ww2.mathworks.cn/ (accessed on 26 March 2022)). The authors thank the anonymous reviewers for their in-depth reviews and helpful suggestions that have largely contributed to improving this paper.

**Conflicts of Interest:** The authors declare no conflict of interest.
