**1. Introduction**

With the continuous development of global navigation satellite systems (GNSSs), spaceborne GNSS reflectometry (GNSS-R) technology has become a hot research direction in the field of remote sensing. In 1993, Martín-Neira proposed the concept of the Passive Reflectometry and Interferometry System (PARIS) and the use of GNSS-R for ocean altimetry [1]. Since then, GNSS-R has been utilized for a range of ocean and land applications, including sea surface altimetry [2], sea surface wind speed measurements [3], sea ice detection [4], and soil moisture measurements [5]. Over the past few decades, a number of ground-based GNSS-R experiments have been conducted. Many airborne experiments have also been conducted to investigate this new remote sensing technology. Notwithstanding some technological challenges, satellite-based GNSS-R technology has the advantages of low cost and great coverage in some applications [6]. Currently, there are more than 14 satellites in operation carrying a GNSS-R payload.

UK-DMC (United Kingdom—Disaster Monitoring Constellation), the first satellite carrying a GNSS-R receiver, was launched on 27 September 2003; data from this system have been used to sense ocean roughness. UK TDS-1 (TechDemoSat-1), the second GNSS-R satellite, was launched on the 8 July 2014. On the 15 December 2016, NASA launched eight microsatellites to form the cyclone GNSS (CYGNSS) constellation with the initial objective of monitoring hurricane intensity [7,8]. Both TDS-1 and CYGNSS have generated a large amount of data which can be downloaded for scientific research [9]. On the 5 June 2019, the BuFeng-1 A/B twin satellites, developed by CASTC (China Aviation Smart Technology Co., Shenzhen, China), were launched from the Yellow Sea. One focus of the satellite mission is on the sensing of sea surface wind velocities, and especially typhoons, using GNSS-R [10].

Sea surface wind speed is an important and commonly used ocean geophysical parameter [11]. The stability of the wind field plays an important role in ocean circulation and global climate [12,13]. Traditional sea surface wind field monitoring methods generally use buoys or coastal meteorological stations, but these methods can only cover small areas with low spatial resolution and expensive equipment [14]. Microwave scatter meters and synthetic aperture radars can also monitor the global sea surface wind field [15,16]. Compared with these traditional wind measurement methods, spaceborne GNSS-R has several advantages, such as rich signal sources and all-weather, all-day, low cost, and large coverage [17,18].

GNSS-R technology is basically mature in retrieving sea surface wind speeds. Zavorotny and Voronovich proposed the scattering model theory in 2000 [19], which can simulate different waveforms of GNSS reflection signals, thus inverting sea surface wind speeds by delayed waveform matching methods [20]. Since then, observations extracted from DDMs (Delay Doppler Maps) have been widely used. DDM is the basic observation data of airborne and spaceborne GNSS-R receivers [21]. Some DDM observations, such as DDM average (DDMA), are directly related to sea surface roughness [21]. Other DDM observations can be used as variables for retrieving sea surface parameters. The normalized bistatic radar cross-section (NBRCS), leading edge slope (LES) and signal-to-noise ratio (SNR) have good correlations with the mean square slope (MSS) of the sea surface. Generally, the MSS is mainly affected by the sea surface wind speed [22].

In recent years, many spaceborne GNSS-R wind speed retrieval models have been developed. Jing et al. demonstrated the effectiveness of NBRCS by proposing some geophysical model functions (GMFs) related thereto [10]. Bu et al. proposed double- and triple-parameter GMFs with higher retrieval accuracy [14]. Machine learning methods have also been used to improve the performance of spaceborne GNSS-R wind speed retrieval. Liu Y. et al. proposed a machine learning algorithm based on a multi-hidden layer neural network. The accuracy of their models was significantly higher than that of GMFs [23]. Many subsequent studies have adopted similar algorithms and obtained results with RMSE of about 1.5–2.0 [24–26]. However, most of the above studies observed that it is difficult to use their algorithms to accurately retrieve high sea surface wind speeds [27,28]. A few studies have tried to enhance the ability of GNSS-R to retrieve high wind speeds. For instance, Zhang et al. developed machine learning-based models to retrieve wind speeds (20–30 m/s) with an RMSE of 2.64 and a correlation coefficient of 0.25 [29].

With high wind speed intervals, the Spaceborne GNSS-R data present different distributions and physical characteristics compared to when low wind speed intervals are applied, which leads to the inconsistent performance of different machine learning models. Therefore, this study analyzes the performance of various machine learning models in different wind speed intervals using the following methods: Regression trees (Binary Tree (BT), Ensembles of Trees (ET), XGBoost (XGB), LightGBM (LGBM) ), ANN (Artificial neural network), Stepwise Linear Regression (SLR), and Gaussian Support Vector Machine (GSVM). In this research, the selection of the input parameters for machine learning methods was significant. In this article, a range of variables are considered and evaluated, which are directly or indirectly relevant to sea surface wind speed. The main contributions of the article are as follows:

(1) Seven machine learning methods are used to retrieve sea surface wind speed, and their performance is evaluated under two different wind speed ranges.


The rest of the paper is organized as follows. Section 2 introduces the GNSS-R variables and then describes the basic principles of the machine learning methods used in this study. Section 3 provides details of the applied data preprocessing strategies, the data filtering algorithm, and the construction of the machine learning-based model; the experimental results are also presented. Section 4 discusses the effects of the variables on wind speed retrieval. Section 5 presents the conclusions.
