Sea Surface Wind Speed Retrieval Using Gaofen-3-02 SAR Full Polarization Data

Abstract

The primary payload onboard the Gaofen-3-02 (GF3-02) satellite is a C-band Synthetic Aperture Radar (SAR) capable of achieving a maximum resolution of 1 m. This instrument is critical to monitor the marine environment, particularly for tracking sea surface wind speeds, an important marine environmental parameter. In this study, we utilized 192 sets of GF3-02 SAR data, acquired in Quad-Polarization Strip I (QPSI) mode in March 2022, to retrieve sea surface wind speeds. Prior to wind speed retrieval for vertical-vertical (VV) polarization, radiometric calibration accuracy was analyzed, yielding good performance. The results showed a bias and root mean square errors (RMSEs) of 0.02 m/s and 1.36 m/s, respectively, when compared to the European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis V5 (ERA5) data. For horizontal–horizontal (HH) polarization, two types of polarization ratio (PR) models were introduced based on the GF3-02 SAR data. Combining these refitted PR models with CMOD5.N, the results for HH polarization exhibited a bias of −0.18 m/s and an RMSE of 1.25 m/s in comparison to the ERA5 data. Regarding vertical–horizontal (VH) polarization, two linear models based on both measured normalized radar cross sections (NRCSs) and denoised NRCSs were developed. The findings indicate that denoising significantly enhances the accuracy of wind speed measurements for VH polarization when dealing with low wind speeds. When compared against buoy data, the wind speed retrieval results demonstrated a bias of 0.23 m/s and an RMSE of 1.77 m/s. Finally, a comparative analysis of the above retrieval results across all three polarizations was conducted to further understand their respective performances.

1. Introduction

The sea surface wind field plays a pivotal role in the research of marine environments. Conventional approaches of recording this parameter rely on shore-based observatories, ships, and buoys, which offer limited measurement ranges and a small number of data points. These methods are often constrained by climate states [1]. With advances in remote sensing technology, there has been a growing interest in using airborne or satellite-based sensors to remotely detect the sea surface wind field. Among these technologies, the spaceborne microwave scatterometer stands out as the most important sensor with respect to global marine wind field remote sensing, with its accuracy for wind speed and direction measurements reaching 2 m/s and 20°, respectively [2]. However, the spatial resolution of a microwave scatterometer typically ranges from about 12.5 to 25 km, which falls short of meeting the requirements for some high-resolution wind fields.

Compared to a spaceborne scatterometer, spaceborne SAR has a higher spatial resolution (1 m to 10 m). This capability allows for high-resolution monitoring of the global marine wind field. However, a spaceborne SAR has lower temporal resolution and a smaller overall spatial coverage compared with spaceborne scatterometer. The sea surface wind field alters the texture of the ocean surface through wind stress, which subsequently influences the intensity of ocean surface radar backscatter [3]. By measuring the intensity of radar backscatter from the ocean surface, SAR systems are able to derive the ocean surface wind field. The ongoing development of techniques of wind field retrieval using SAR data has been a persistent, long-standing endeavor.

The most common method for retrieving wind fields using SAR images is to utilize empirical geophysical model functions (GMFs). These functions are established based on measured data from ocean buoys or numerical prediction models that align with radar observations. The GMFs correlate with the NRCS, wind speed and direction at the 10 m height above the sea surface, radar incidence angle, azimuthal angle, frequency, and polarization mode [4]. In contrast to scatterometer, which have multiple azimuth viewing angles, SAR instruments offer a single azimuth view. Consequently, to obtain the sea surface wind speed, it is necessary to independently determine the sea surface wind direction from an alternative data source, such as the forecasting model, scatterometer data, or buoy measurements [5]. In this discussion, we focus solely on the retrieval of wind speed. Currently, the predominant operational models for retrieving sea surface wind speed are the CMOD series models, which have been tailored for C-band SAR imagery with VV polarization. These models include CMOD4 [6], CMOD-IFR2 [7], CMOD5 [8], CMOD5.N [9], C2013 [10], CMOD5.H [11], and CMOD7 [12]. The CMOD series models have proven effective in processing C-band SAR satellite images, including data from satellites like ERS-1/2, ENVISAT, Radarsat-1/2, Sentinel-1, and Gaofen-3 (GF-3). Fan et al. [13] employed the CMOD-IFR2 model to derive sea surface wind speeds from ENVISAT images. When these derived wind speeds were compared with the NCEP/QSCAT hybrid and daily average scatterometer wind speeds, the RMSEs were all found to be less than 2 m/s. Zhang et al. [14] applied the CMOD series models to retrieve wind speeds from Radarsat-2 images, ultimately determining that CMOD5 is the most suitable C-band GMF for retrieving wind speeds from Radarsat-2 VV polarization data, and the RMSE of the retrieval results is 1.02 m/s compared with the buoy data. Similarly, Zhang et al. [15] verified that CMOD5 is optimal for Sentinel-1 VV polarization data, and the RMSE of the retrieval results is 1.17 m/s compared with the buoy data. China successfully launched its inaugural C-band multi-polarization SAR in 2016, named GF-3 [16]. Numerous studies have confirmed the effectiveness of GF-3 imagery for retrieving sea surface wind speeds, with CMOD5.N identified as the most effective C-band GMF for retrieving wind speeds from GF-3 VV polarization data [17,18,19,20].

Table 1. Statistical analysis of GF-3 wind speed retrieval results for VV polarization.

Researchers	Time Span	Reference Dataset	CMOD	RMSE(m/s)
Li et al. [17]	09/2016–11/2017	WindSat Data	CMOD5.N	1.72
Wang et al. [18]	01/2017–04/2017	Buoy Data	CMOD5.N	2.34
Shao et al. [19]	09/2016–03/2017	ERA5 Data	CMOD5.N	2.00
Zhang et al. [20]	10/2016–05/2018	ERA5 Data	CMOD5.N	2.36

presents the retrieval results of the aforementioned studies using GF-3 data for VV polarization.

The aforementioned CMOD series models are specifically designed for VV polarization. For HH polarization, the general approach involves converting the NRCSs from HH polarization to VV polarization utilizing a PR model, after which the CMOD series of models are applied for wind speed retrieval. Currently, PR models fall into two categories. The first type is related to the incidence angle, such as the Elfouhaily model [21], the Thompson model [22], and the Mouche model [23]. The second type incorporates the incidence angle as well as the relative wind direction, as suggested by Mouche et al. [23]. Several studies have also proposed new PR models based on these existing ones, tailored to specific SAR data used in research [18,19,20].

Table 2. Statistical analysis of GF-3 wind speed retrieval results for HH polarization.

Researchers	Time Span	Reference Dataset	Optimal Model	RMSE (m/s)
Wang et al. [18]	01/2017–04/2017	Buoy Data	Mouche model	2.50
Shao et al. [19]	09/2016–03/2017	ERA5 Data	Mouche model	2.20
Zhang et al. [20]	10/2016–05/2018	ERA5 Data	Mouche model	2.26

presents the retrieval results of the aforementioned studies using GF-3 data for HH polarization. These studies indicate that for GF-3 HH polarization data, the Mouche model has the best retrieval performance. Additionally, it is feasible to use wind speed retrieval models for HH polarization directly without any PR conversion, such as the CSARMOD [24] and CMODH [25] models.

The retrieval of wind speed from co-polarization (VV or HH) data is routinely employed in operational detection. However, its application has certain limitations. Specifically, at very high wind speeds, particularly beyond 33 m/s, the wind speed obtained by the co-polarization model exhibits clear saturation. In contrast, as Zhang et al. [26] demonstrated, the cross-polarization (VH or HV) SAR NRCS maintains a linear correlation with wind speed and remains unaffected by wind direction. This characteristic makes it feasible to utilize cross-polarization NRCSs for estimating the intense wind speeds associated with typhoons and hurricanes through their linear relationship [27,28]. Nevertheless, the echo energy of cross-polarization signals is inherently weaker compared to that of co-polarization. Consequently, cross-polarization NRCS is more vulnerable to noise interference at lower wind speeds [29], highlighting the potential need for denoising procedures to improve low wind speed retrieval accuracy.

China successfully launched the GF3-02, a C-band SAR satellite featuring a one-meter resolution, on 23 November 2021, at the Jiuquan Satellite Launch Center. The primary purpose of this satellite is to safeguard marine rights, oversee maritime emergencies and sea areas, observe polar environments and provide navigational support for polar research vessels, as well as facilitate marine scientific research. As China’s first operational SAR satellite in space, GF3-02 joins its predecessor, GF-3, and its successor, Gaofen-3-03 (launched on 7 April 2022), forming a constellation dedicated to ocean surveillance and monitoring. The advanced spaceborne SAR onboard GF3-02 supports 12 distinct imaging modes and can access fully polarimetric data, encompassing assemblages of VV, HH, VH, and HV polarizations. Building upon earlier work by Yang et al. [30], which preliminarily assessed the wind speed retrieval capability using co-polarization data from GF3-02, this study aims to further understand the satellite’s capabilities. By using the comprehensive dataset of full polarization data from GF3-02, we verified the calibration accuracy for VV polarization and introduced novel models for retrieving HH polarization SAR data. In addition, two models of wind speed retrieval for VH polarization were also proposed. Moreover, we compared three distinct wind speed retrieval results to deepen our insight into the satellite’s capabilities.

The subsequent parts of this paper are arranged in the following manner. Section 2 provides an overview of the SAR data and reference data obtained from external sources. Section 3 details the methodologies, outcomes, and analysis for retrieving wind speeds from the SAR data across VV, HH, and VH polarizations. Section 4 provides the discussion. Section 5 summarizes the findings.

2. Materials

2.1. SAR Data

The GF3-02 SAR, operating in QPSI Imaging Mode, encompasses four polarizations: VV, HH, VH, and HV. Between 14 and 16 March 2022, we gathered a dataset consisting of 192 QPSI mode scenes. Classified as Level-1A products, these scenes consist of single-look complex (SLC) imagery. Each scene is centered on oceanic regions, exhibiting an incidence angle ranging from 20° to 50°. The geographical distribution of these collected scenes predominantly spans the seas adjacent to the Hawaiian Islands and the western seaboard of the United States, illustrated in Figure 1 with red areas indicating the geolocations of data collection.

Table 3. The details of the acquired QPSI mode data.

Orbit Code	Beam Code	Center Incidence Angle (deg)	Observation Time	Count of Scene
1607	199	36.23	14 March 2022	32
1614	199	36.22	14 March 2022	57
1614	192	26.56	14 March 2022	26
1629	211	46.03	15 March 2022	39
1634	213	47.41	16 March 2022	38

outlines the detailed attributes of each dataset. Notably, the beam code assigned to these scenes corresponds directly to their respective incidence angles. Furthermore, the QPSI mode data offer a spatial resolution of 8 m and span a swath width of 30 km. This high-resolution capability underscores the utility of GF3-02 SAR data for detailed maritime observations.

2.2. ERA5 Data

ERA5 represents the fifth-generation reanalysis from the ECMWF. It provides comprehensive atmospheric, ocean wave, and land surface data spanning from 1940 to the present day. This dataset is organized on a grid of 0.25 degrees in both latitude and longitude, with hourly updates available. To facilitate a precise comparison, we collected ERA5 data that are temporally and spatially aligned with the GF3-02 SAR observations. The temporal resolution of the ERA5 data is one hour. Therefore, ERA5 data with a time interval of less than one hour from the SAR data are selected. In order to validate the wind speed retrieval results derived from the SAR data, we interpolated the ERA5 data to correspond exactly with the center point of each subscene within the SAR imagery. The flow chart of the comparison process between ERA5 data and SAR data is shown in Figure 2.

2.3. Buoy Data

The buoy data utilized in this study are sourced from the National Data Buoy Center (NDBC) under the aegis of the National Oceanic and Atmospheric Administration (NOAA). Encompassing a large span of oceanographic parameters, these data include atmospheric conditions, ocean properties, wave characteristics, sea ice information, and climate variables. This dataset is updated every ten minutes, providing high temporal resolution for analysis. A critical consideration in our methodology pertains to the variation in anemometer heights across different buoys, necessitating a standardization process for wind speed measurements. Specifically, the CMOD-derived wind speeds represent neutral wind conditions at a consistent reference altitude of 10 m over the ocean surface. To harmonize the datasets, we employ a logarithmic wind profile equation [31], which facilitates the conversion of all buoy-measured wind speeds to this uniform height of 10 m. This ensures compatibility and accuracy in our comparative analyses.

$$U\left(h\right)=U\left(h_m\right)\times ln(h/h_0)/ln(h_m/h_0)$$

(1)

where $h_0$ is the roughness length (usually taking the value 1.52 × 10⁻⁴), $h$ is the 10 m height, $h_m$ is the altitude of the anemometer on the buoy, $U\left(h_m\right)$ is the wind speed observed at the buoy, and $U\left(h\right)$ is the wind speed at a height of 10 m after conversion. This equation is derived under the assumption of atmospheric neutrality; that is, the impact of differences in atmospheric stability is neglected. When atmospheric conditions are atypical, it may lead to errors. The temporal resolution of the buoy data is ten minutes. Therefore, buoy data with a time interval of less than ten minutes from the SAR data are selected. Because of the limitations of the SAR data acquired, we selected the nearest buoy for each scene within a 50 km radius. A total of 9 scene–buoy pairs were screened for subsequent wind speed validation. The scene–buoy pairs are shown in Figure 3. In addition, the specifics of NDBC buoys are tabulated in

Table 4. The detailed information of NDBC buoys.

Station Code	Area	Station Geolocation (Lat/Lon(deg))	Anemometer Height (m)
46002	West Coastal Seas of America	42.662N/130.507W	3.8
51000	Seas off the Hawaiian Islands	23.528N/153.792W	4.1
51004	Seas off the Hawaiian Islands	17.538N/152.230W	4.1

, and the flow chart of the comparison process between buoy data and SAR data is shown in Figure 4.

3. Methods and Results

3.1. GF3-02 SAR Wind Speed Retrieval for VV Polarization

3.1.1. Radiometric Calibration Accuracy for VV Polarization

As detailed in Section 1, for retrieving wind speed from VV polarization SAR data, CMOD series models are the most commonly used method, which describe the relationship between the NRCSs, the sea surface wind speed, the relative wind direction (difference between sea surface wind direction and the radar look direction), the radar incidence angle for a fixed wavelength, and the polarization mode. In the case of the CMOD5.N, the model takes the following form:

$${\sigma }_0=A_0\left(v,\theta \right){\left(1+A_1\left(v,\theta \right)\mathit{cos}\left(\varphi \right)+A_2\left(v,\theta \right)\mathit{cos}\left(2\varphi \right)\right)}^p$$

(2)

where ${\sigma }_0$ is the NRCS in linear units, $v$ is the wind speed, $\theta$ is the incidence angle, and $\varphi$ is the relative wind direction. $A_0$ characterizes the wind speed, which is a property that remains consistent in all directions, and $A_2$ describes the asymmetry of the upwind and crosswind directions. $A_1$ captures the asymmetry between the upwind and downwind directions, with its magnitude being minor yet significant. The value of the constant p is 1.6.

In order to obtain the wind speed, it is necessary to accurately estimate the NRCSs. The steps encompass radiometric calibration, reduction in speckle noise, and assessment of scene homogeneity. The radiometric calibration of GF3-02 SAR L1A data is as follows:

$${\sigma }_0=10\times \log \left[{\left(I\times {\displaystyle \frac{QV}{\mathrm{32,767}}}\right)}^2+{\left(Q\times {\displaystyle \frac{QV}{\mathrm{32,767}}}\right)}^2\right]-\mathit{K\_const}$$

(3)

where ${\sigma }_0$ denotes the NRCS in nonlinear units, QV refers to the Qualify Value of GF3-02 and $\mathit{K\_const}$ is the calibration constant. I represents the real channel and Q represents the imaginary channel of the same imagery. The reduction in speckle noise in the SAR image is accomplished through the process of pixel averaging. In this investigation, the GF3-02 SAR imagery was subjected to a 5 km averaging process in both the range and azimuth dimensions in order to facilitate the acquisition of subscenes. The assessment of scene homogeneity, as referenced in [32], is employed to eliminate subscenes that have been compromised by oil slicks, upwelling, ships and other phenomenon on the sea surface.

The radiometric calibration accuracy is critical in wind speed retrieval. Yang et al. [30] provided a calibration constant of 13.91 dB for VV polarization and concluded that the CMOD5.N model is the most efficient for retrieving wind speeds from VV polarization GF3-02 SAR imagery, a conclusion we adopted in this study. Regarding the absolute bias of NRCS, the typical precision of spaceborne SAR wind speed retrieval generally falls within 0.5 m/s [20]. Utilizing this benchmark, we employed the CMOD5.N model to determine the required threshold for the absolute bias of NRCS given a wind speed error tolerance of 0.5 m/s. Figure 5 illustrates how the absolute bias of NRCS correlates with different wind speeds, incidence angles, and relative wind directions, under the condition that the GF3-02 SAR data are limited to incidence angles ranging from 20° to 50°. The simulation results reveal a relationship: larger wind speeds and smaller incidence angles necessitate a smaller absolute bias of NRCS to uphold retrieval accuracy. Consequently, under the existing data conditions, the radiometric calibration accuracy, or absolute bias of NRCS, must not exceed 1.43 dB to ensure reliable wind speed retrieval.

To assess the radiometric calibration precision of our SAR dataset, we must examine the disparity between simulated and measured NRCS values. The simulation process incorporates ERA5 wind speed and direction data corresponding to individual subscenes, along with the radar’s incidence angle, into the CMOD5.N model. By comparing this simulated NRCS against the derived measurements from each subscene, we can determine the absolute bias of NRCS. Figure 6 illustrates these findings. The diagram is demarcated by two horizontal red lines, denoting the critical threshold; based on prior analysis, an acceptable deviation limit stands at 1.43 dB or less. Red dots represent the average absolute NRCS bias values computed for 1 m/s wind speed increments, all comfortably within this threshold. Notably, these results reveal that 78% of all assessed NRCS biases comply with this standard, underscoring the high quality of our SAR data calibration with a wind speed bias of 0.5 m/s.

3.1.2. Wind Speed Retrieval for VV Polarization

To further validate both the radiometric calibration accuracy and the wind speed retrieval capability of the VV polarization data, we conducted a wind speed retrieval process utilizing the CMOD5.N model in conjunction with the measured NRCS values of the subscenes. The interpolated ERA5 wind direction datasets served as inputs for the wind direction variable within the CMOD5.N model, while the interpolated ERA5 wind speeds functioned as benchmarks against which to assess the retrieved VV polarization wind speeds. It is important to highlight that, given the greater sampling density of the ERA5 data compared to buoy measurements, only the interpolated ERA5 wind speeds were employed as references to evaluate the performance of VV polarization retrieval. The comparative results depicted in Figure 7 illustrate that the bias and RMSE associated with the wind speed retrieval are minimal at 0.02 m/s and 1.36 m/s, respectively. The strong agreement between the ERA5 reference winds and retrieved wind speeds confirms the capability of the GF3-02 SAR in retrieving wind speeds using VV polarization data.

3.2. GF3-02 SAR Wind Speed Retrieval for HH Polarization

3.2.1. Development of the PR Model from GF3-02 HH Polarization Data

PR is necessary for transforming NRCSs from HH polarization to VV polarization, which is expressed as the linear ratio between the NRCSs of VV polarization and that of HH polarization, as shown in the following equation:

$$PR={\sigma }_0^{VV}/{\sigma }_0^{HH}$$

(4)

where ${\sigma }_0^{VV}$ and ${\sigma }_0^{HH}$ represent the NRCS values in linear units for VV and HH polarizations, respectively. From Section 1, numerous PR models were developed in past studies. However, to accurately retrieve the wind speed for HH polarization, a specific PR model applicable to the HH polarization data is necessary. The PR model is influenced by the incidence angle and relative wind direction. The relationships between PR and these factors were examined separately, as shown in Figure 8. Due to the limited amount of data, the statistics do not cover all possible incidence angles and relative wind directions comprehensively. In the figure, black dots represent the mean values of the dataset range, with error bars overlaid for clarity. Figure 8a shows the relationship between the PR and the incidence angle. As illustrated, there is a significant positive correlation; the PR rises with an increase in the incidence angle. Figure 8b illustrates the correlation between the PR and the relative wind direction. Despite the relative wind direction data covering only a range from 150° to 360°, a cosine-like pattern can be observed. The PR values peak at around 180 degrees (downwind) and reach their minimum at approximately 270 degrees (crosswind). These results indicate that the PR values derived from GF3-02 SAR data correlate significantly with both the incidence angle and the relative wind direction.

Drawing from the analysis above, we have formulated two distinct PR models utilizing GF3-02 SAR data. The first type of model relies solely on the incidence angle, while the second type depends on both the incidence angle and the relative wind direction. Models that rely solely on the incidence angle typically fall into three categories: the Elfouhaily model [21], the Thompson model [22], and the Mouche model [23]. The specific formulations for these models are presented below.

$$PR\left(\theta \right)={\left(1+2{tan}^2\theta \right)}^2/{\left(1+\alpha {sin}^2\theta \right)}^2$$

(5)

$$PR\left(\theta \right)={\left(1+2{tan}^2\theta \right)}^2/{\left(1+\beta {tan}^2\theta \right)}^2$$

(6)

$$PR\left(\theta \right)=Aexp\left(B\theta \right)+C$$

(7)

where $\theta$ is the incidence angle. $\alpha$, $\beta$, $A$, $B$ and $C$ are the coefficients of the PR models. Based on the available GF3-02 SAR data, the coefficients were refitted using a least squares estimation method. The refitted coefficients are presented in

Table 5. The refitted coefficients of the PR models that depend on the incidence angle.

Coefficient	Value
$\alpha$	1.9693
$\beta$	1.1106
$A$	0.0166
$B$	0.0980
$C$	0.9279

. Specifically, $\alpha$ represents the coefficient for the refitted Elfouhaily model, $\beta$ corresponds to the refitted Thompson model, and $A$, $B$, and $C$ denote the coefficients for the refitted Mouche model. Figure 9 illustrates the comparison between these three refitted models and the SAR data. Within this figure, the black solid line plots the PR versus the incidence angle for the SAR data, consistent with Figure 8a. Both the refitted Elfouhaily model (blue solid line) and the refitted Mouche model (green solid line) exhibit trends that are largely consistent with each other and align well with the black solid line. Conversely, the refitted Thompson model (red solid line) overestimates the PR value at incidence angles below 39 degrees while underestimating it at angles above 39 degrees. In conclusion, both the refitted Elfouhaily model and the refitted Mouche model appear to be suitable choices for retrieving wind speeds from GF3-02 SAR data in HH polarization.

The PR model, which is contingent upon the incidence angle and relative wind direction, is presented in the following equation [23]:

$$PR\left(\theta ,\varphi \right)=C_0\left(\theta \right)+C_1\left(\theta \right)cos\varphi +C_2\left(\theta \right)cos2\varphi$$

(8)

where $\theta$ represents the incidence angle and $\varphi$ denotes the relative wind direction. For each specific relative wind direction, the correlation between PR and the incidence angle is described by the equation that follows:

$${PR}_{\varphi }\left(\theta \right)=A_{\varphi }\mathit{exp}\left(B_{\varphi }\theta \right)+C_{\varphi }$$

(9)

The coefficients $C_i$ (i = 0, 1, 2) can be expressed by ${PR}_{\varphi }\left(\theta \right)$ with upwind, downwind and crosswind.

$$C_0\left(\theta \right)=({PR}_0\left(\theta \right)+{PR}_{\pi }\left(\theta \right)+2{PR}_{\pi /2}\left(\theta \right))/4$$

(10)

$$C_1\left(\theta \right)=({PR}_0\left(\theta \right)-{PR}_{\pi }\left(\theta \right))/2$$

(11)

$$C_2\left(\theta \right)=({PR}_0\left(\theta \right)+{PR}_{\pi }\left(\theta \right)-2{PR}_{\pi /2}\left(\theta \right))/4$$

(12)

The coefficients $A_{\varphi }$, $B_{\varphi }$ and $C_{\varphi }$ ($\varphi =0,\pi /2,\pi$) were refitted using least squares estimation derived from the available SAR data. The refitted coefficients are displayed in

Table 6. The refitted coefficients of the PR model that rely on the incidence angle and relative wind direction.

Coefficient	Value
$A_0$	12.9948
$B_0$	0.0054
$C_0$	−13.9381
$A_{\pi /2}$	9.9005 × 10−6
$B_{\pi /2}$	0.2413
$C_{\pi /2}$	1.3640
$A_{\pi }$	2.2308 × 10−5
$B_{\pi }$	0.2273
$C_{\pi }$	1.6298

.

3.2.2. Validation of the PR Model from GF3-02 HH Polarization Data

To verify the reliability of the improved PR models, these models were employed for transforming an NRCS from HH polarization to VV polarization. The converted NRCS was then compared with measured VV polarization NRCS, as shown in Figure 10. The refitted Elfouhaily and Mouche models demonstrated good performance, showing strong agreement between the converted and measured NRCS values. In contrast, the refitted Thompson model showed less effectiveness, with a slight discrepancy between the converted and measured NRCS values, consistent with Figure 9. Additionally, the results from a PR model that relies on both incidence angle and relative wind direction (refitted Mouche-azimuth model) were included. This model performed slightly worse than the refitted Mouche model. The possible reason is that the GF3-02 SAR data used to develop the PR model did not cover the full range of azimuth angles (0°~360°), leading to errors in model refitment.

Wind speed retrieval for HH polarization was conducted using the converted NRCS in combination with the CMOD5.N model. The retrieved wind speeds were then compared to the ERA5 wind speeds, as depicted in Figure 11. The outcomes from the refitted Elfouhaily and Mouche models were essentially consistent and outperformed those from the refitted Mouche–azimuth and Thompson models. This aligns with the findings from the previous NRCS comparison. The RMSE of the retrieval results using the refitted Mouche model in conjunction with CMOD5.N is 1.25 m/s, indicating that retrieving wind speeds using GF3-02 SAR data in HH polarization is more efficient.

3.3. GF3-02 SAR Wind Speed Retrieval for VH Polarization

3.3.1. Development of the Wind Speed Retrieval Model from GF3-02 VH Polarization Data

The cross-polarization for GF3-02 SAR data includes both VH and HV polarizations. Since their characteristics are fundamentally identical, we focus on analyzing the VH polarization data in this study. To model wind speed retrieval for GF3-02 SAR VH polarization, we created a scatter plot of the measured NRCS values for VH polarization against the corresponding ERA5 wind speeds (Figure 12a). The plot reveals a significant linear correlation between the measured NRCS for VH polarization and the ERA5 wind speeds. Consequently, we developed a linear wind speed retrieval model for VH polarization, as shown in Equation (13).

$${\sigma }_0^{VH}=0.6191U_{10}-38.3293$$

(13)

where ${\sigma }_0^{VH}$ is the NRCS for VH polarization. $U_{10}$ represents the wind speed measured at the 10 m level above the sea level. The red line in Figure 12a represents this linear model. However, the echo energy of VH polarization is weak, making it more susceptible to noise at low wind speeds. Noise equivalent sigma zero (NESZ) is considered in the modeling of VH polarization wind speed retrieval. Shi et al. [33] found a direct correlation between NESZ and beam code. As shown in Table 3, the SAR data used in this study include four beam codes: 192, 199, 211, and 213. The NESZs are estimated for different beam code data. Building on the idea of creating a linear model, we modeled each beam code data linearly with respect to the measured NRCS and the corresponding ERA5 wind speeds. Taking Equation (13) as an example, in the linear model, if the wind speed is 0 m/s, the NRCS at this situation is considered to be NESZ. Therefore, the NESZs for the four beam codes were estimated, as presented in

Table 7. The NESZs of the GF3-02 SAR data for four beam codes.

Beam Code	NESZ Value (dB)
192	−35.0294
199	−41.1854
211	−36.8109
213	−39.7870

.

After the NESZ is known, the formula for calculating the NRCS after denoising is as shown below.

$${\sigma }_0^{denoise}={\sigma }_0^{VH}-NESZ$$

(14)

where ${\sigma }_0^{denoise}$ is the NRCS after denoising. The three parameters in Equation (14) are all in linear units. The scatter plots of the denoised NRCS for VH polarization and the corresponding ERA5 wind speeds are shown in Figure 12b. A wind speed retrieval model based on linear principles, utilizing the denoised NRCS from VH polarization, was developed, as demonstrated in Equation (15). The red line in Figure 12b illustrates the linear model based on the denoised NRCS values.

$${\sigma }_0^{VH}=0.9798U_{10}-42.8917$$

(15)

3.3.2. Validation of the Linear Model from GF3-02 VH Polarization Data

To verify the validity of the two linear models presented in Equations (13) and (15), we used both the measured and denoised NRCSs for VH polarization to perform wind speed retrieval using these models. The derived wind speeds were contrasted with the respective ERA5 wind speeds, as illustrated in Figure 13. Figure 13a,b displays the comparison results before and after denoising, respectively. And Figure 13c,d displays the histogram of wind speed bias before and after denoising, respectively. The wind speed comparisons before denoising show greater scatter than those after denoising. When contrasted with ERA5 wind speeds, the RMSEs before and after denoising are 3.03 m/s and 2.07 m/s, respectively. When the ERA5 wind speed exceeds 9.5 m/s, the scatter points exhibit greater concentration, and the precision of the retrieved wind speed markedly improves. This suggests that VH polarization data are particularly well suited for retrieving high wind speeds. Figure 13 reveals a positive deviation, which could be due to the relatively weaker signal energy of VH polarization under low wind speed conditions (less than 9.5 m/s). This weakness renders VH polarization more vulnerable to noise interference compared to VV and HH polarizations. Furthermore, even with denoising in place, the estimation of NESZ can still introduce errors. Since these linear models were established based on ERA5 wind speeds, using them solely as references for comparison is incomplete. To provide a more robust validation, buoy data mentioned in Section 2 were used as an additional reference, as demonstrated in Figure 14. Contrasted with buoy wind speeds, the RMSEs before and after denoising are 2.80 m/s and 1.77 m/s, respectively. These results closely match those obtained utilizing ERA5 data as a reference. In summary, the denoising process is essential for VH polarization wind speed retrieval from GF3-02 SAR data when dealing with low wind speeds (less than 9.5 m/s), consistent with the findings reported in [34].

As detailed in Section 1, SAR cross-polarization data demonstrate superiority over co-polarization data as regards retrieving high wind speeds. However, the QPSI mode data utilized in the previously mentioned studies feature a limited swath width of merely 30 km, which poses challenges for typhoon observations. Consequently, we endeavored to employ data from the GF3-02 TOPSAR mode for the retrieval of typhoon wind speeds. The wide-swath operational mode of TOPSAR is composed of five sub-swaths, boasting a swath width of up to 500 km and a resolution of 100 m. The TOPSAR mode data for VH polarization was captured at 21:34 on 4 September 2022, focusing on Hinnamnor situated in the East China Sea. A linear model, as shown in Equation (13), was applied for wind speed retrieval, with the resultant distribution depicted in Figure 15a. The arrows within the figure denote the direction of the typhoon’s wind, as derived from ERA5 data. The peak retrieved wind speed stands at 35 m/s. The distribution map of wind speed bias, compared with ERA5 wind speed, is presented in Figure 15b. As can be seen from the figure, the wind speed bias is of less than 4 m/s in most locations. However, in some areas of the typhoon wall, the bias reaches up to 10 m/s. This divergence can be attributed to the fact that the ERA5 data encompassed a wind speed range of 0–25 m/s, thereby constraining the model’s capacity to assess high wind speeds. In the future, there is a necessity for further investigation into the observational potential of GF3-02 TOPSAR mode data with regard to typhoons.

3.4. Comparison of GF3-02 SAR Wind Speed Retrieval for VV, HH and VH Polarizations

The distribution of wind speed retrieval results of VV polarization, HH polarization and VH polarization of GF3-02 SAR was statistically analyzed, with ERA5 wind speed also included, as shown in Figure 16. For HH polarization, we used the results from the refitted Mouche model combined with CMOD5.N, while for VH polarization, we utilized data after the denoising process. Figure 16 shows a strong correlation between the wind speed retrieval results from VV and HH polarizations. In contrast, the wind speed retrieval results for VH polarization are different from those for VV polarization. Moreover, Figure 17 presents the spatial distribution of retrieved wind speed for all three polarizations of GF3-02 SAR data, focusing on the seas off the Hawaiian Islands (shown in Figure 1). The distributions of the retrieved wind speeds from VV and HH polarizations are very comparable, whereas the distributions from VV and VH polarizations show some differences. These findings echo the trends observed in Figure 16. The similarity in wind speed retrieval results between VV and HH polarizations can be attributed to the fact that the reference value utilized for refitting the PR model parameters of HH polarization is derived from the NRCS of VV polarization. Consequently, upon conversion of the HH polarization NRCS using the refitted PR model, the resultant NRCS closely aligns in value with that of VV polarization, as illustrated in Figure 10. Furthermore, both polarizations employ the CMOD5.N model for wind speed retrieval, which contributes to the consistency in their retrieval results. However, as detailed in Section 3.3.2, VH polarization exhibits a notably larger error when retrieving low wind speeds (less than 9.5 m/s). This discrepancy results in a divergence of the wind speed retrieval results for VH polarization from those of VV and HH polarizations.

Figure 18 illustrates the spatial distribution of retrieved wind speed bias for the three polarizations when compared to ERA5 wind speed. This figure is complementary to Figure 17, focusing on the seas off the Hawaiian Islands. It is evident from the figure that the maximum bias of retrieved wind speed for VV and HH polarizations is 3 m/s, whereas for VH polarization, this bias can peak at 5 m/s. The larger wind speed bias observed in VH polarization can be attributed to the susceptibility of its low-energy signals to noise interference, which, in turn, results in more significant deviations in the retrieved wind speed under low wind speed conditions. Furthermore, a comprehensive summary of the aforementioned wind speed retrieval results is provided in

Table 8. Summary of retrieved wind speed accuracy for three polarizations.

Polarizations	Reference Dataset	Bias (m/s)	RMSE (m/s)
VV	ERA5 Data	0.02	1.36
HH	ERA5 Data	−0.18	1.25
VH	ERA5/Buoy Data	0.01/0.23	2.07/1.77

. The statistical analysis reveals that the precision of wind speed retrieval for VH polarization is marginally inferior to that of VV and HH polarizations, suggesting that there remains potential for enhancing the precision of wind speed retrieval for VH polarization.

4. Discussion

This research employs GF3-02 SAR full polarization data to retrieve wind speed and corroborates the accuracy through comparisons with ERA5 and buoy data. Specifically, for VV polarization, the wind speed retrieved by the CMOD5.N model is compared with ERA5 data, yielding an RMSE of 1.36 m/s. In the case of HH polarization, the wind speed obtained using both the refitted Mouche model and the CMOD5.N model is assessed against ERA5 data, resulting in an RMSE of 1.25 m/s. For VH polarization, the retrieved wind speed is compared with ERA5 data and buoy data, resulting in RMSEs of 2.07 m/s and 1.77 m/s, respectively.

Numerous factors can compromise the accuracy of wind speed retrieval using GF3-02 SAR data. Primarily, the radiometric calibration accuracy of GF3-02 SAR is a critical factor. As analyzed in Section 3.1.1, 22% of the NRCS deviations beyond the acceptable threshold persist, suggesting that the employed calibration constants harbor inaccuracies that directly affect the precision of wind speed retrieval. Secondly, inherent inaccuracies within the CMOD5.N model cannot be overlooked. Yang et al. [30] demonstrated through their study that employing diverse CMOD models for wind speed retrieval leads to varying degrees of deviation, underscoring the intimate connection between retrieval accuracy and the specific CMOD model utilized. Additionally, the reference data, encompassing both ERA5 and buoy data, are not immune to deviations. Section 3.3.2 highlights that when ERA5 and buoy data were separately employed to validate the accuracy of wind speed retrieval for VH polarization, a discrepancy of 0.2 to 0.3 m/s emerged in the evaluation results. This discrepancy illustrates that the choice of reference data significantly influences the assessment results. Furthermore, the insufficiency of the current dataset poses a challenge. Section 2.3 reveals that only nine sets of matched buoy data are available, which introduces an element of randomness into the wind speed retrieval accuracy results. Moreover, the SAR data statistics detailed in Section 3.2.1 indicate that the lack of comprehensive coverage across the incidence and azimuth angles of the data compromises the robustness of the refitted PR models. Lastly, for HH polarization, the selection of different PR models can markedly impact the wind speed retrieval results. The verification results presented in Table 2 and Figure 11 clearly demonstrate that for GF-3 and GF3-02 data, the Mouche model proves to be more suitable and achieves a higher degree of wind speed retrieval accuracy. In the context of GF3-02 SAR data’s VH polarization, the relatively weak signal energy of VH polarization means that the accuracy of NESZ estimation becomes a pivotal factor affecting wind speed retrieval accuracy, especially when dealing with low wind speeds (less than 9.5 m/s).

5. Conclusions

This study explores the ability of GF3-02 SAR data to retrieve wind speeds across VV, HH, and VH polarizations. To assess the precision of radiometric calibration for GF3-02 SAR VV polarization data, the threshold for the absolute bias of NRCS is simulated using CMOD5.N with a wind speed bias of 0.5 m/s, yielding a value of 1.43 dB. The absolute bias of NRCS is derived by differencing the simulated NRCSs from the measured NRCSs. The results indicate that 78% of the absolute biases fall within this threshold, suggesting a relatively high degree of radiometric calibration accuracy. Further comparison with ERA5 wind speed data demonstrates good consistency for VV polarization retrieval, with a bias of 0.02 m/s and an RMSE of 1.36 m/s, reinforcing the radiometric calibration’s accuracy.

Considering the influence of incidence angle and relative wind direction on PR values within GF3-02 SAR data, three incidence-angle-dependent models and one combined model were refitted for HH polarization. Among the three models that consider the incidence angle, the refitted Mouche model is the most accurate, showing a bias and RMSE of −0.18 m/s and 1.25 m/s, respectively, when compared to ERA5 wind speeds. The combined model, incorporating both incidence angle and relative wind direction factors, exhibits a bias and an RMSE of −0.13 m/s and 1.31 m/s, respectively, slightly lower in accuracy than the refitted Mouche model due to the lack of relative wind direction data in the acquired GF3-02 SAR data.

For VH polarization, a linear model was established based on measured NRCSs and corresponding ERA5 wind speed data. Considering the weak echo energy of VH polarization, NESZ values were estimated for different beam codes to mitigate noise effects. Consequently, a denoising process was applied to establish a more accurate linear model. Compared with buoy data, the denoising model demonstrated a bias and an RMSE of 0.23 m/s and 1.77 m/s, respectively, proving superior performance over the non-denoised model. These findings underscore the necessity of denoising for accurate wind speed retrieval in VH polarization using GF3-02 SAR data when dealing with low wind speeds.

Additionally, wind speed retrieval results for GF3-02 SAR in both HH and VH polarizations were compared against those for VV polarization. The comparison reveals that HH polarization results are more consistent with VV polarization results than VH polarization results. Future work will focus on gathering more sea surface SAR data matching both buoy and ERA5 datasets. Moreover, enhancements to the PR model for HH polarization and wind speed retrieval for VH polarization are needed to bolster the comprehensive analysis and elevate the overall effectiveness of fully polarimetric GF3-02 SAR wind speed retrieval.