Sea Surface Wind Speed Retrieval from the First Chinese GNSS-R Mission: Technique and Preliminary Results

: Launched on 5 June 2019, the BuFeng-1 A / B twin satellites were part of the ﬁrst Chinese global navigation satellite system reﬂectometry (GNSS-R) satellite mission. In this letter, a brief introduction of the BF-1 mission and its preliminary results of sea surface wind retrieval are presented. Empirical fully developed sea (FDS) geophysical model functions (GMFs) relating the normalized bistatic radar cross-section to the sea surface wind speed are proposed for the BF-1 GNSS-R instruments. The FDS GMFs are derived from the collocated BF-1 observations, the European Center for Medium-Range Weather Forecasts (ECMWF) reanalysis data, and the advanced scatterometer (ASCAT) satellite observations. The preliminary tests reveal that the root-mean-square error (RMSE) between the derived wind speed and the reanalysis is 2.63 m / s for wind speeds in the range of 0.5–40.5 m / s. Further comparisons with the ASCAT observations and mooring buoys show that the RMSEs are 2.04 m / s and 1.77 m / s, respectively, at low-to-moderate wind speeds. This study demonstrates the e ﬀ ectiveness of BF-1 and provides a basis for the future GMF development of the BF-1 A / B mission.


Introduction
The global navigation satellite system (GNSS) continuously emits L-band microwaves to the Earth's surface, which are then reflected into space. By using a GNSS-reflectometry (GNSS-R) receiver, the Earth-reflected signals carrying geophysical information can be collected. In 1988, this GNSS-R technique was first applied to ocean scatterometry [1]. In 1993, Martin-Neira [2] proposed the PAssive Reflectometry and Interferometry System (PARIS) concept to use the reflected GPS signal to perform ocean altimetry. After this, GNSS-R signals were received on land-based, coastal, and airborne platforms to prove the sufficient power budget [3][4][5][6][7][8][9][10][11][12][13]. Then, this technique was first tested in orbit by the UK-Disaster Monitoring Constellation (UK-DMC) mission in 2003 [14,15]. The first delay Doppler map (DDM) from low Earth orbit was obtained by this mission. In July 2014, a technology demonstration GNSS-R satellite, TechDemoSat-1 (TDS-1), was launched and served for 4.5 years. The onboard GNSS-R

BF-1 Mission and Instruments
"Bu Feng (捕风)" means "chasing the winds" in Chinese; BF-1 A/B is the first Chinese GNSS-R mission comprising two demonstration satellites. The payloads of the BF-1 A/B twin satellites, which are the main payloads of the carrier, were developed by the China Aerospace Science and Technology Corporation (CASC). Each platform has two nadir GNSS-R antennas, one navigation antenna, one auxiliary antenna, four low-noise amplifiers with blackbody calibration, and a GNSS-R receiver. The two nadir antennas are directed at the left and right sides of the platform with inclination angles of 26°. In general, the number of specular points located in the antenna footprint of the Global Positioning system (GPS) is approximately four. Both the instruments have four digital channels to receive the best four signals out of the specular points by different GNSS satellites. For every second, the onboard receivers can simultaneously process the four reflected signals into DDMs. The specifications of the receiver are shown in Table 1. Each GNSS-R receiver collects four DDMs per second for 24 h each day. In this mission, the world's first in-orbit BDS DDM was mapped on 10 June 2019 and its first DDM was acquired on 9 June 2019. The in-orbit-measured DDM has dimensions of 128 × 21, corresponding to 32 delay chips in quarter chip increments and a 5000 Hz Doppler frequency with 250 Hz intervals. Examples of DDMs measured by the BF-1A instrument are shown in Figure 2.

BF-1 Mission and Instruments
"Bu Feng (捕风)" means "chasing the winds" in Chinese; BF-1 A/B is the first Chinese GNSS-R mission comprising two demonstration satellites. The payloads of the BF-1 A/B twin satellites, which are the main payloads of the carrier, were developed by the China Aerospace Science and Technology Corporation (CASC). Each platform has two nadir GNSS-R antennas, one navigation antenna, one auxiliary antenna, four low-noise amplifiers with blackbody calibration, and a GNSS-R receiver. The two nadir antennas are directed at the left and right sides of the platform with inclination angles of 26 • . In general, the number of specular points located in the antenna footprint of the Global Positioning system (GPS) is approximately four. Both the instruments have four digital channels to receive the best four signals out of the specular points by different GNSS satellites. For every second, the onboard receivers can simultaneously process the four reflected signals into DDMs. The specifications of the receiver are shown in Table 1.

Data
To develop the FDS GMF and test its performance, we introduced three datasets: buoy measurements, satellite scatterometer observations, and European Center for Medium-Range Weather Forecasts (ECMWF) reanalysis data.

Buoy Measurements
The TAO buoy array was obtained from the NDBC, supported primarily by the US National Oceanic and Atmospheric Administration (NOAA). The TAO array comprises 55 moored buoys near the central equatorial Pacific Ocean. It is a sensor network telemetering oceanographic and metrological data in real-time by taking the advantage of the Argos satellite system. Therefore, the measured sea surface winds are appropriate for the instant performance assessment at low and moderate wind speeds.
The NDBC TAO measurements mainly include surface winds, sea surface temperature, ocean currents, precipitation, humidity, salinity, and pressure. Figure 1 shows the distribution of the mooring buoys we used: they are arranged in eight columns from 165°E to 95°W longitudes and 8°S to 9°N latitudes. The temporal resolution is 10 min. The real-time data have a 1-h delay of obtainment and the historic data has a lag of one day. For the wind speed measurements, the NOAA NDBC buoy data have an accuracy of 1.0 m/s or 10%, which is the highest accuracy among the three datasets in this paper.
The wind from the TAO program is at a 4-m reference height, which can be calibrated into the standard U10 by the following equation [40]: where z (m) is the observation height, Uz is the wind speed at the observation height, and U10 is the 10-m reference height wind speed.

ASCAT Observations
The advanced scatterometer (ASCAT) is a C-band active microwave instrument on the EUMETSAT MetOp satellites, operating since 2007. ASCAT measures the sea surface roughness,

Data
To develop the FDS GMF and test its performance, we introduced three datasets: buoy measurements, satellite scatterometer observations, and European Center for Medium-Range Weather Forecasts (ECMWF) reanalysis data.

Buoy Measurements
The TAO buoy array was obtained from the NDBC, supported primarily by the US National Oceanic and Atmospheric Administration (NOAA). The TAO array comprises 55 moored buoys near the central equatorial Pacific Ocean. It is a sensor network telemetering oceanographic and metrological data in real-time by taking the advantage of the Argos satellite system. Therefore, the measured sea surface winds are appropriate for the instant performance assessment at low and moderate wind speeds.
The NDBC TAO measurements mainly include surface winds, sea surface temperature, ocean currents, precipitation, humidity, salinity, and pressure. Figure 1 shows the distribution of the mooring buoys we used: they are arranged in eight columns from 165 • E to 95 • W longitudes and 8 • S to 9 • N latitudes. The temporal resolution is 10 min. The real-time data have a 1-h delay of obtainment and the historic data has a lag of one day. For the wind speed measurements, the NOAA NDBC buoy data have an accuracy of 1.0 m/s or 10%, which is the highest accuracy among the three datasets in this paper.
The wind from the TAO program is at a 4-m reference height, which can be calibrated into the standard U 10 by the following equation [40]: where z (m) is the observation height, U z is the wind speed at the observation height, and U 10 is the 10-m reference height wind speed.

ASCAT Observations
The advanced scatterometer (ASCAT) is a C-band active microwave instrument on the EUMETSAT MetOp satellites, operating since 2007. ASCAT measures the sea surface roughness, which relates Remote Sens. 2019, 11, 3013 5 of 13 to wind speeds 10 m above the water's surface. In this study, we used the ASCAT version 2.1 data, which are reprocessed by the remote sensing systems (RSSs) with the latest GMF, C-2015 [41]. The RSS produces daily wind field products with a spatial resolution of 25 km in the form of ascending and descending swaths. ASCAT winds are known as being reprocessed with high accuracy. Compared with mooring buoys, unbiased global ASCAT wind errors on the scatterometer measurement scale are about 0.7 m/s [42].

ECMWF Reanalysis Data
The ECMWF supplies both global numerical weather predictions (NWPs) and reanalysis history datasets of the atmosphere, land surface, and oceans. The ECMWF reanalysis data are based on the integrated forecast model and provide spatially continuous wind products globally with an interpolated spatial resolution of up to 0.125 • longitude/latitude and a temporal resolution of 3 h. Two reanalysis datasets are available for sea surface winds until 1 September 2019: ERA-Interim and ERA5.

Delay Doppler Maps (DDMs)
DDMs are the standard products of spaceborne GNSS-R, as shown in Figure 2. The DDM instruments (DDMI) onboard BF-1 A/B create the DDMs by receiving the sea surface-reflected GNSS signal powers in every time delay and the Doppler frequency bins. Similarly, the instruments use a 1-ms coherent integration time and 1000-times noncoherent integration to guarantee the signal-to-noise ratio (SNR). The theoretical DDM calculations are described by Zavorotny-Voronovich [3] as follows: where Υ(τ, f ) 2 is the correlation power of each code delay (τ) and Doppler frequency (f ) in every DDM bin, P t is the transmitter power, G t is the transmitter antenna gain, λ is the L-band wavelength of~19 cm for the GPS L1 carrier, T i is the coherent integration time, G r is the receiver antenna gain, Λ 2 (τ) is the delay code auto-relation function, S 2 (f ) is the Doppler frequency function, A is the effective area of the certain code delay and Doppler frequency, R t and R r are the distances of the effective area to the transmitter and receiver, respectively, σ 0 is the scattering parameter sensitive to the sea surface roughness, which highly correlates with the sea surface wind speed. Once the σ 0 is determined, the DDM can be simulated by Equation (2).

NBRCS Calculation
The parameter σ 0 mentioned in Equation (2) above is also the NBRCS. In this study, the NBRCS is applied as the sea surface roughness-sensitive parameter, also known as the observable, to retrieve the sea surface wind speeds [30]. To calculate the NBRCS (σ 0 ), Equation (2) can be rewritten as follows [33]: where most of the terms are the same as those in Equation (2), however, P g τ, f is the power calibration result of Υ(τ, f ) 2 , and R t and R r are the specular distances to the transmitter and receiver, respectively.
These two range loss terms from the integration are based on the assumption that the ranges are approximately equal near the specular point. Because the averaged NBRCS used to retrieve the winds is only a small area around the specular point, the uncertainty of this assumption regarding the range loss is nearly negligible.
It should be noted that the content in the double integration of Equation (3) is determined only by the geometry. Once the positions and velocities of the receiver and transmitter are determined, the specular point can be calculated numerically. Then, a grid is built around the specular point to compute the contributions in each delay and the Doppler frequency bin. As the other geometry terms, like the antenna gains and range losses, are not considered integrally, σ 0 only has good accuracy near the specular point. Therefore, a 1.5-chip and 2000-Hz delay Doppler domain was selected to calculate the NBRCS (σ 0 ) in this study, which can also be considered as the DDMA.

FDS GMF Development
In this subsection, the ECMWF reanalysis data and ASCAT observations from 1 July 2019 to 1 September 2019 are used as training datasets to develop FDS GMFs. The reanalysis data can supply a spatially continuous wider wind speed range corresponding to a larger amount of matchup pairs. On the other hand, the ASCAT observations have a higher accuracy on wind speed and limited swaths which result in less collocated winds.

ECMWF Reanalysis Data
The matchup NWP data, used to train the empirical FDS GMF, are the 10-m referenced sea surface wind speed reanalysis data provided by the ECMWF. In this study, we used the ERA-Interim dataset of the ECMWF reanalysis to conduct a 2-month global sea surface wind speed retrieval from 1 July 2019 to 1 September 2019. The ERA-Interim reanalysis dataset used has an interpolated spatial dimension of 0.125 • longitude/latitude and 3-h increments of temporal resolution. Finally, the matchup volume of the NBRCSs was 12,632,225 with alignment criteria of 1 h and the nearest longitude/latitude grid. The matchup maximum wind was up to 40.8 m/s.
After the 3-dB beam-width antenna pattern screening, the scatter density plot was obtained between the NBRCS and ECMWF reanalysis winds, as shown in Figure 3a. The wind speeds range from 0.5 to 40.8 m/s. A significant power function relationship between the ECMWF reanalysis winds and NBRCS extracted from DDMs can be inferred. Therefore, the empirical FDS GMF was developed based on this relationship. As shown in Figure 3a, 16.86% of the NBRCSs from 0 to 20 correspond to the moderate wind speeds of 5 to 15 m/s, indicating that a large amount of NBRCSs will correspond to this wind speed range. Consequently, the wind speeds above 15 m/s will be underrated by these extra data. Accordingly, a quality control method is necessary to screen out the low NBRCSs that relate to the low and moderate wind speeds mentioned above.
Therefore, after the antenna pattern gains control of a 3-dB beam-width and SNR of 4-dB quality control, 80% of the data were employed as the training set; the remaining 20% was used to validate the model accuracy. The scatter density plot after quality control of the antenna pattern and SNR is shown in Figure 3b. Notably, by setting up the filters of the antenna gain control and SNR threshold, the NBRCS ambiguity problem of moderate wind speeds between 5 and 15 m/s can be largely mitigated from the rate of 16.86% to 1.59%, which effectively improves the performance of the empirical GMF.
Next, we further developed the empirical FDS GMF based on the NBRCS (σ 0 ), wind speed (U 10 ), and elevation angle (θ ele ). Here, the elevation angle is the angle between the horizontal plane and the line extending from the spacecraft to the specular point. First, we built a lookup table (LUT) for the elevation angle. It is known that the common sea Fresnel coefficient varies as the elevation angle increases. The elevation angle of 28 • is the inflection point at which the left-handed circle polarization takes a larger portion compared with the right-handed circle polarization. Thus, the elevation angle is divided into 63 bins, and the bin center is incremented by 1 • from 28 • to 90 • with a 1 • bin width. Therefore, all the DDMs and auxiliary information within each elevation bin compute the optimal coefficients of the power function (U 10 = Aσ B 0 + C), as shown in Figure 4. The coefficients of the power functions of different elevation angles then fill the 2-D LUT. Figure 4 presents examples of the empirical GMF that correspond to elevation angles of 35-84 • , indicating that the elevation angle has a greater influence on the wind speed. and NBRCS extracted from DDMs can be inferred. Therefore, the empirical FDS GMF was developed based on this relationship. As shown in Figure 3a, 16.86% of the NBRCSs from 0 to 20 correspond to the moderate wind speeds of 5 to 15 m/s, indicating that a large amount of NBRCSs will correspond to this wind speed range. Consequently, the wind speeds above 15 m/s will be underrated by these extra data. Accordingly, a quality control method is necessary to screen out the low NBRCSs that relate to the low and moderate wind speeds mentioned above.  Therefore, after the antenna pattern gains control of a 3-dB beam-width and SNR of 4-dB quality control, 80% of the data were employed as the training set; the remaining 20% was used to validate the model accuracy. The scatter density plot after quality control of the antenna pattern and SNR is shown in Figure 3b. Notably, by setting up the filters of the antenna gain control and SNR threshold, the NBRCS ambiguity problem of moderate wind speeds between 5 and 15 m/s can be largely mitigated from the rate of 16.86% to 1.59%, which effectively improves the performance of the empirical GMF.
Next, we further developed the empirical FDS GMF based on the NBRCS (σ ), wind speed (U10), and elevation angle (θele). Here, the elevation angle is the angle between the horizontal plane and the line extending from the spacecraft to the specular point. First, we built a lookup table (LUT) for the elevation angle. It is known that the common sea Fresnel coefficient varies as the elevation angle increases. The elevation angle of 28° is the inflection point at which the left-handed circle polarization takes a larger portion compared with the right-handed circle polarization. Thus, the elevation angle is divided into 63 bins, and the bin center is incremented by 1° from 28° to 90° with a 1° bin width. Therefore, all the DDMs and auxiliary information within each elevation bin compute the optimal coefficients of the power function ( σ ), as shown in Figure 4. The coefficients of the power functions of different elevation angles then fill the 2-D LUT. Figure 4 presents examples of the empirical GMF that correspond to elevation angles of 35-84°, indicating that the elevation angle has a greater influence on the wind speed.

ASCAT Data
Using the same process, the ASCAT data were collocated with the BF-1 A/B NBRCS data from 1 July to 1 September. The total number is 860,859 with the matchup criteria of 1 h and the nearest longitude/latitude grid. The reason for the lower volume of coincident pairs is that the ASCAT data are limited by their swath and orbit, which are different from the spatially continuous ECMWF models. The matchup maximum wind speed is 33.2 m/s.
After the same quality control method of the 3-dB antenna beam-width and SNR ≥ 4, 80% of the data were used to train the FDS GMF for the sea surface wind speed. The elevation angle bins remained as 28-90 • with a 1 • bin width. Figure 5 shows examples of the best-fitted results using the ASCAT training dataset.

ASCAT Data
Using the same process, the ASCAT data were collocated with the BF-1 A/B NBRCS data from 1 July to 1 September. The total number is 860,859 with the matchup criteria of 1 h and the nearest longitude/latitude grid. The reason for the lower volume of coincident pairs is that the ASCAT data are limited by their swath and orbit, which are different from the spatially continuous ECMWF models. The matchup maximum wind speed is 33.2 m/s.
After the same quality control method of the 3-dB antenna beam-width and SNR ≥ 4, 80% of the data were used to train the FDS GMF for the sea surface wind speed. The elevation angle bins remained as 28-90° with a 1° bin width. Figure 5 shows examples of the best-fitted results using the ASCAT training dataset.

Wind Retrieval Testing
This section describes the validation of the trained empirical FDS GMFs by three datasets: the remaining 20% collocated reanalysis winds, the remaining 20% of the ASCAT observations validation dataset, and the NDBC TAO mooring buoys measurements.

Reanalysis Comparisons
When developing the empirical FDS GMF using the reanalysis data, we only used 80% of the matchup data to tune the model coefficients, while the remaining 20% were used for model testing. The remaining 20% was comprised of 461,932 data points. The aligned ECMWF reanalysis wind speeds and NBRCS derived wind speeds are from 0.6 to 36.7 m/s. As shown in Figure 6, most of the

Wind Retrieval Testing
This section describes the validation of the trained empirical FDS GMFs by three datasets: the remaining 20% collocated reanalysis winds, the remaining 20% of the ASCAT observations validation dataset, and the NDBC TAO mooring buoys measurements.

Reanalysis Comparisons
When developing the empirical FDS GMF using the reanalysis data, we only used 80% of the matchup data to tune the model coefficients, while the remaining 20% were used for model testing. The remaining 20% was comprised of 461,932 data points. The aligned ECMWF reanalysis wind speeds and NBRCS derived wind speeds are from 0.6 to 36.7 m/s. As shown in Figure 6, most of the data are located around the 1:1 line. The retrieved wind speeds had a root mean square error (RMSE) of 2.72 m/s and the coefficient of determination was 0.5. data are located around the 1:1 line. The retrieved wind speeds had a root mean square error (RMSE) of 2.72 m/s and the coefficient of determination was 0.5.
Owing to the large number of reanalysis data pairs, another test of 10 minus the alignment criteria was conducted. The total number of matchup pairs then became 428,832. After the 80% training and 20% validation, the scatter density plot was obtained, as shown in Figure 6b. Finally, the unbiased RMSE was 2.63 m/s, and the coefficient of determination was 0.55.

ASCAT Comparisons
The remaining 20% of the aligned data were applied by the FDS GMF as described in Section 3.3.2. After the quality control filters, the maximum wind speed was 25 m/s and the number of validation data became 34,029. The performance of the ASCAT-trained GMF has an unbiased RMSE of 2.04 m/s, and the coefficient of determination is 0.61, as shown in Figure 7.  Owing to the large number of reanalysis data pairs, another test of 10 minus the alignment criteria was conducted. The total number of matchup pairs then became 428,832. After the 80% training and 20% validation, the scatter density plot was obtained, as shown in Figure 6b. Finally, the unbiased RMSE was 2.63 m/s, and the coefficient of determination was 0.55.

ASCAT Comparisons
The remaining 20% of the aligned data were applied by the FDS GMF as described in Section 3.3.2. After the quality control filters, the maximum wind speed was 25 m/s and the number of validation data became 34,029. The performance of the ASCAT-trained GMF has an unbiased RMSE of 2.04 m/s, and the coefficient of determination is 0.61, as shown in Figure 7.
Remote Sens. 2019, 11, x FOR PEER REVIEW 9 of 14 data are located around the 1:1 line. The retrieved wind speeds had a root mean square error (RMSE) of 2.72 m/s and the coefficient of determination was 0.5. Owing to the large number of reanalysis data pairs, another test of 10 minus the alignment criteria was conducted. The total number of matchup pairs then became 428,832. After the 80% training and 20% validation, the scatter density plot was obtained, as shown in Figure 6b. Finally, the unbiased RMSE was 2.63 m/s, and the coefficient of determination was 0.55.

ASCAT Comparisons
The remaining 20% of the aligned data were applied by the FDS GMF as described in Section 3.3.2. After the quality control filters, the maximum wind speed was 25 m/s and the number of validation data became 34,029. The performance of the ASCAT-trained GMF has an unbiased RMSE of 2.04 m/s, and the coefficient of determination is 0.61, as shown in Figure 7.

Buoy Comparisons
Applied using the alignment criteria of 1 • longitude/latitude and 10 min, the total number of matchups was 16,552 after the simple antenna 3-dB beam-width quality control. This quantity is insufficient for generating an FDS GMF with multiple elevation bins. Therefore, the NDBC TAO buoy measurements were introduced to validate the robustness of the model functions.
The sea surface wind speeds derived by the proposed empirical FDS GMFs were compared with the NDBC TAO buoy measurements. After a reference transition from 4 m to 10 m, the buoy wind speed measurements were 0.2-11.6 m/s. As shown in Figure 8, compared with the buoy measurements, the reanalysis GMF-derived unbiased RMSE is 1.77 m/s and the coefficient of determination is 0.45. Meanwhile, the ASCAT-derived GMF has an unbiased RMSE of 1.82 m/s and a coefficient of determination of 0.41.

Buoy Comparisons
Applied using the alignment criteria of 1° longitude/latitude and 10 min, the total number of matchups was 16,552 after the simple antenna 3-dB beam-width quality control. This quantity is insufficient for generating an FDS GMF with multiple elevation bins. Therefore, the NDBC TAO buoy measurements were introduced to validate the robustness of the model functions.
The sea surface wind speeds derived by the proposed empirical FDS GMFs were compared with the NDBC TAO buoy measurements. After a reference transition from 4 m to 10 m, the buoy wind speed measurements were 0.2-11.6 m/s. As shown in Figure 8, compared with the buoy measurements, the reanalysis GMF-derived unbiased RMSE is 1.77 m/s and the coefficient of determination is 0.45. Meanwhile, the ASCAT-derived GMF has an unbiased RMSE of 1.82 m/s and a coefficient of determination of 0.41.

Summary and Conclusions
From the process of developing the empirical FDS GMF for the BF-1 A/B twin satellites, some details, such as the effects of the antenna gain control, elevation angle, inconsistency between satellite channels, and SNR, can be further summarized as follows.
The BF-1 A/B DDMs were collected by ranking the antenna gains to the specular points scattered in the antenna footprint. The strongest four antenna gains were selected to be considered closest to the main lobe of the nadir antennas. Without any quality control strategy, the original data had an RMSE above 4 m/s. As Figure 3a illustrates, the percentage of the NBRCS was large at the values between 0 and 20, corresponding to the wind speed range of 5-15 m/s. Consequently, a large portion of the NBRCSs corresponded to this range, so high wind speeds above 15 m/s will be underrated. Therefore, we next explored the following four steps to alleviate this issue. First, a 3-dB half beamwidth filter was introduced to ensure the signal is near the main lobe. Thus, the performance of the retrieved wind speed had an RMSE of below 3.8 m/s. Second, the influence of the elevation angle was considered in the GMF by building 63 bins from 28-90°, providing an accuracy of approximately 3.6 m/s of RMSE. Then, the inconsistency between the satellites and channels was considered. As a result, the RMSE became 3.5 m/s. Finally, the most effective factor, SNR, was found to improve the problem of the NBRCS ambiguity between 0 and 20, as illustrated in Figure 3a. After an SNR ≥4-dB filter, the percentage of NBRCS ambiguity became 1.59% from 16.86%. Figure 6 shows that the final reanalysis

Summary and Conclusions
From the process of developing the empirical FDS GMF for the BF-1 A/B twin satellites, some details, such as the effects of the antenna gain control, elevation angle, inconsistency between satellite channels, and SNR, can be further summarized as follows.
The BF-1 A/B DDMs were collected by ranking the antenna gains to the specular points scattered in the antenna footprint. The strongest four antenna gains were selected to be considered closest to the main lobe of the nadir antennas. Without any quality control strategy, the original data had an RMSE above 4 m/s. As Figure 3a illustrates, the percentage of the NBRCS was large at the values between 0 and 20, corresponding to the wind speed range of 5-15 m/s. Consequently, a large portion of the NBRCSs corresponded to this range, so high wind speeds above 15 m/s will be underrated. Therefore, we next explored the following four steps to alleviate this issue. First, a 3-dB half beam-width filter was introduced to ensure the signal is near the main lobe. Thus, the performance of the retrieved wind speed had an RMSE of below 3.8 m/s. Second, the influence of the elevation angle was considered in the GMF by building 63 bins from 28-90 • , providing an accuracy of approximately 3.6 m/s of RMSE. Then, the inconsistency between the satellites and channels was considered. As a result, the RMSE became 3.5 m/s. Finally, the most effective factor, SNR, was found to improve the problem of the NBRCS ambiguity between 0 and 20, as illustrated in Figure 3a. After an SNR ≥ 4-dB filter, the percentage of NBRCS ambiguity became 1.59% from 16.86%. Figure 6 shows that the final reanalysis FDS GMF performance has an RMSE of 2.63 m/s at 32.6 m/s, and 2.04 m/s (RMSE) and 0.61 (R 2 ) against the ASCAT observations, as in Figure 7. Thus, regarding the wind speed accuracy and wind speed range, BF-1 performs better than the existing GNSS-R missions: the RMSE of UK TDS-1 against ASCAT is 2.21 m/s for winds between 3 and 18 m/s [19], and the RMSE of CYGNSS FDS GMF is~4 m/s at 20 m/s, compared with the model reanalysis products [23].
The empirical GMFs were also tested with in-situ buoy measurements. The comparison results show a preliminary agreement at the low-to-moderate wind speed range. The RMSEs of the retrieved wind speeds were 1.77 m/s and 1.82 m/s. Figure 8 provides a clear underestimation of the wind speeds. There are two reasons for this: one is the inconsistency of the ECMWF reanalysis wind data, ASCAT observations, and NDBC TAO buoy wind measurements. Meanwhile, the empirical GMFs were not developed based on the buoy measurements, so the behavior is less impressive than the reanalysis and ASCAT results.
This paper presents the results of sea surface wind speed retrieval using the first Chinese GNSS-R satellite data collected by BF-1 A/B. After months of in-orbit operation, the data were first referenced with the sea surface wind speeds product of the ECMWF reanalysis and ASCAT observations to train the empirical FDS GMFs. The NBRCS extracted from the collected DDM was used as the observable when developing the GMF. Based on the data filter of the 3-dB half-beam-width antenna and SNR ≥ 4 dB, the empirical GMFs were developed using the power function of the NBRCS (σ 0 ), wind speed (U 10 ), and elevation angle (θ ele ). Then, the validation datasets were applied by the empirical GMFs to derive the global sea surface wind speeds to assess the model accuracy. Even without potentially large platform attitude error corrections, the proposed GMF obtained an unbiased RMSE of 2.63 m/s at wind speeds of 32.7 m/s. The comparison results of the buoy measurements and ASCAT observations show a better performance under low-to-moderate wind speed conditions.
This study provides an important foundation for the construction of future Chinese operational GNSS-R constellations for the observation of typhoons. A GNSS-R constellation can largely increase the capability of the spatial and temporal coverage, thereby enabling near-real-time typhoon observations and forecasts. Furthermore, all objectives mentioned above could be realized by the low-cost GNSS-R technique. However, the forerunner BF-1 A/B twin satellites still require further work. In the future, the attitude information should be studied and introduced to optimize the quality control process. This would definitively improve the performance of the empirical GMF on sea surface wind speed retrieval. Moreover, a study on the sea surface wind speed under heavy precipitation is required to improve the NBRCS behavior for high wind speeds. Finally, assessment of other applications is required for the validation, such as the inland performance to guide the future constellation mission design.