In April 2001, the Japan Meteorological Agency (JMA) started the operation of a wind profiler (WPR) network, WInd profiler Network and Data Acquisition System (WINDAS; Ishihara et al., 2006). WINDAS consists of 33 1.3 GHz band wind profilers as of August 2021 (black squares in Fig. 1). The specifications of WPR are listed in Table 1. WINDAS can operate continuously, acquiring vertical profiles of horizontal wind speed, wind direction, vertical velocity, and signal-to-noise ratio (SNR) over the wind profilers using five beams (one vertical beam and four oblique beams). The horizontal wind speed and wind direction are calculated from radial wind speeds by the four-beam method under strict data quality control (Adachi et al., 2005). WINDAS provides a profile of wind data with high accuracy. In operational mode, the temporal and vertical resolutions of WINDAS data are 10 min and 291 m, respectively. The minimum and maximum detection heights are 294 m and 11.6 km above the wind profiler, respectively. There are 40 range bins for one wind profile. The wind measurement accuracy of the WPRs was evaluated by comparisons with winds forecasted by the NWP model and radiosondes (Tada, 2001). From the comparisons, the wind measurement accuracy of the WPRs was comparable to that from radiosonde observations. The random error (root mean square error) σWPR of zonal winds was determined to be about 3 m s-1. The comparison of wind data between Aeolus and the WPRs is useful for assessing wind measurement performance and the spatio-temporal variation in the wind field.

Considering the different spatial and temporal resolutions of the WPRs and the Aeolus, data-matching procedures are necessary before comparing the data obtained by the two sensors. First, the WPR data and Aeolus data need to be matched in both space and time. To achieve geographical matching, the distance between the mean positions of an Aeolus measurement and the WPR was set to be less than 100 km. To achieve temporal synchronization, we used averages of WPR wind data from 30 min before to 30 min after the passage of Aeolus. There is also a difference in the vertical resolution between Aeolus measurements and WPR measurements. The horizontal wind speed and wind direction measured by the WPRs were averaged to the Aeolus bin using the top and bottom altitudes given in the Aeolus L2B data product. After temporal and spatial collocation, the Aeolus L2B wind product closest to each WPR measurement was adopted for comparison. The horizontal wind speed and wind direction measured by the WPRs during the periods from 1 October 2018 to 15 May 2019 (baseline 2B02) and from 28 June to 31 December 2019 and from 20 April to 8 October 2020 (baseline 2B10) were used to compare Aeolus HLOS wind data.

NICT has installed 1.54 µm CDWLs (WINDCUBE 400S manufactured by LEOSPHERE; Cariou et al., 2006) in Kobe (34.66∘ N, 135.16∘ E; magenta circle in Fig. 1) and Okinawa (26.50∘ N, 127.84∘ E; yellow circle in Fig. 1). The specifications of the CDWLs are listed in Table 2. The CDWL in Kobe was placed on the rooftop of a building managed by Kobe City. The CDWL in Okinawa was placed on the fifth floor (25.1 m a.m.s.l.) of the steel tower in Okinawa Electromagnetic Technology Center of NICT (hereafter, NICT Okinawa). In this experiment, their range bins had a length of 150 m, with the centre of the first bin at 300 m. With 159 range bins per beam, adjacent range bins were overlapped by 83.1 m, and the maximum range was about 13.4 km depending on the aerosol load and/or cirrus clouds present. The vertical profiles of horizontal wind speed and wind direction were acquired by the Doppler beam swinging (DBS; Röttger and Larsen, 1990) technique from four inclined beams (north, east, south, and west) with an elevation angle of 70∘. The Doppler velocity spectra for all range bins of each beam were obtained 100 000 times on average. Since the PRF was 10 kHz, the accumulation time of each beam was 10 s. The Doppler wind speed at each bin was estimated from the averaged Doppler-shifted frequency spectra using the maximum likelihood estimator (Levin, 1965). We evaluated the bias and random error for wind measurements of the CDWLs using the methods described by Iwai et al. (2013). Bias was estimated at 0.02 m s-1 using measurements from a stationary hard target for single LOS measurements. Random errors were 0.02 to 0.10 m s-1 from -10 to -30 dB wideband SNR and the CDWLs operated near a theoretical Cramer–Rao lower bound (Aoki et al., 2016; Rye and Hardesty, 1993). On the basis of the comparison with collocated radiosonde data, the systematic error and random error (root mean square error) σCDWL of horizontal wind speed acquired by the DBS technique were determined to be about 0.2 and 2 m s-1, respectively (Aoki et al., 2015). Therefore, the CDWL measurements act as a reference owing to their low systematic and random errors that result from the coherent measurement principle of the system. As for the WPR data, the CDWL data and Aeolus data need to be matched in both space and time. To achieve geographical matching, the distance between the mean position of an Aeolus measurement and the CDWL should be less than 100 km. As mentioned earlier, we averaged Doppler velocity spectra for all range bins of each beam from 30 min before to 30 min after the passage of Aeolus, and then the vertical profiles of horizontal wind speed and wind direction were acquired by the DBS technique. As with the WPR, the horizontal wind speed and wind direction measured by the CDWLs were averaged to the Aeolus bin. In Okinawa, the vertical profiles of horizontal wind speed and wind direction measured during the periods from 18 October 2018 to 11 May 2019 (baseline 2B02) and from 28 June to 31 December 2019 and from 20 April to 8 October 2020 (baseline 2B10) were obtained to compare Aeolus HLOS wind data. In Kobe, the vertical profiles of horizontal wind speed and wind direction measured during the periods from 16 October 2018 to 15 May 2019 (baseline 2B02) and from 3 September to 31 December 2019 and from 20 April to 15 July 2020 (baseline 2B10) were obtained to compare Aeolus HLOS wind data.

Twelve GPS radiosondes (GPS-RSs) of type RS41-SGP produced by Vaisala were launched from NICT Okinawa (26.50∘ N, 127.84∘ E; yellow circle in Fig. 1) from October to December 2018 (baseline 2B02). The specifications of the RS41-SGP are listed in Table 3. From September to December 2019 (baseline 2B10), six GPS-RSs were also launched from NICT Okinawa. An overview of the 18 obtained validation cases is given in Table 4. The GPS-RSs transmit observed data every 2 s to an MW41 ground receiver unit. The observed data are processed using Vaisala proprietary software (DigiCORA version). The vertical resolution is about 10 m at the typical ascending speed of 5 m s-1. The horizontal wind speed and direction are calculated using changes in the GPS location. According to the estimated Global Climate Observing System Reference Upper-Air Network (GRUAN), the measurement uncertainties of the horizontal wind speed σGPS-RS and direction are assumed to be 0.7 m s-1 and 1∘, respectively (Dirksen et al., 2014). Although the measurement uncertainties are derived from the radiosonde of type RS92 and not RS41, there is no significant difference in the uncertainty as both radiosonde types use the same technique to obtain wind speed and direction (Jensen et al., 2016; Kawai et al., 2017). Since the GPS-RS wind data are obtained by direct in situ measurements, the GPS-RS observations are generally very accurate, and the instrument errors are small. The GPS-RS measurements are suitable for use as a reference data set for the validation of Aeolus HLOS winds. Furthermore, the observation errors can be assumed to be uncorrelated between different GPS-RSs. However, other errors arise due to the GPS-RS drift during its ascent. The averaged ascent time of the GPS-RSs was about 45 min when they reached an altitude of 25 km. The GPS-RSs launched from NICT Okinawa drifted by a horizontal distance of up to about 120 km. These values were considered when defining collocation criteria for comparisons of Aeolus and GPS-RS measurements. In this study, the GPS-RS measurements that were within 120 km horizontal distance and 60 min temporal difference from the Aeolus measurements were used for the validation. As with the WPR, the horizontal wind speed and wind direction measured by the GPS-RSs were averaged to the Aeolus bin.

Scatter plots of Aeolus HLOS winds against CDWL HLOS winds for Rayleigh-clear and Mie-cloudy winds during the baseline 2B02 period are presented in Fig. 11. Summaries of the statistical parameters retrieved from the scatter plot analysis for the baseline 2B02 and 2B10 are given in Table 7. While Okinawa is located at the southern edge of the subtropical jet stream, Kobe is located just below the subtropical jet stream. Thus, the CDWL at Kobe sampled a higher wind speed of the subtropical jet stream. It can be seen that the acquired HLOS wind speed range is wider for Kobe than for Okinawa in Fig. 11. Both Rayleigh-clear and Mie-cloudy winds exhibit a slightly positive bias. The different colours indicate whether Aeolus had an ascending orbit (red) or descending orbit (blue). There is no significant difference between the ascending and descending orbits. The slopes of the linear regression lines are 1.05 (Rayleigh) and 1.05 (Mie) at Kobe and 0.99 (Rayleigh) and 1.01 (Mie) at Okinawa. The correlation coefficients are 0.98 (Rayleigh) and 0.98 (Mie) at Kobe and 0.93 (Rayleigh) and 0.97 (Mie) at Okinawa. That is, the slopes of the fit and the correlation coefficients of Rayleigh-clear and Mie-cloudy winds are not significantly different from 1 at Kobe and Okinawa. The intercepts of the linear regression lines are determined to be 0.61 m s-1 (Rayleigh) and 1.76 m s-1 (Mie) at Kobe and 1.07 m s-1 (Rayleigh) and 2.37 m s-1 (Mie) at Okinawa. A similar finding is obtained from the biases that are 0.46 m s-1 (Rayleigh) and 1.63 m s-1 (Mie) at Kobe and 1.08 m s-1 (Rayleigh) and 2.38 m s-1 (Mie) at Okinawa. Both Rayleigh-clear and Mie-cloudy winds exhibit a slightly positive bias. Except for Rayleigh-clear winds measured at Kobe, the systematic error does not achieve the mission requirement of 0.7 m s-1. This result is similar to that in the comparisons of Aeolus and WPR measurements, which provides biases of 1.69 m s-1 (Rayleigh) and 2.42 m s-1 (Mie). The systematic error of CDWL observations is smaller than 0.2 m s-1 (see Sect. 2.3) and thus does not significantly contribute to the biases here. The random errors represented by the scaled MADs of Rayleigh-clear (Mie-cloudy) winds are 4.92 (3.55) m s-1 at Kobe and 5.68 (3.76) m s-1 at Okinawa. The values are smaller than the scaled MADs of Rayleigh-clear (Mie-cloudy) versus WPR winds. The main reason for the difference is probably related to the random error being larger for the WPR (3 m s-1) than for the CDWL (2 m s-1). The σAeolus of Rayleigh-clear (Mie-cloudy) winds is determined using Eq. (5) to be 4.49 (2.93) m s-1 at Kobe and 5.31 (3.19) m s-1 at Okinawa. Witschas et al. (2020) determined σAeolus of 3.9 m s-1 for Rayleigh-clear winds and 2.0 m s-1 for Mie-cloudy winds by excluding the airborne 2 µm CDWL measurement error during the commissioning phase. The discrepancies are probably caused by the smaller representativeness error due to the spatial and temporal displacements between Aeolus and airborne CDWL measurements.

Figure 12 shows the correlation plots of the Aeolus HLOS winds against CDWL HLOS winds for Rayleigh-clear and Mie-cloudy winds at Kobe and Okinawa during the baseline 2B10 period. As with the baseline 2B02 period, a linear trend between Aeolus and CDWL measurements is clearly seen from the linear regression. At Kobe, the correlation coefficients are 0.96 and 0.97 for Rayleigh-clear and Mie-cloudy winds, respectively, and close to 1. At Okinawa, the correlation coefficients are 0.79 and 0.86 for Rayleigh-clear and Mie-cloudy winds, respectively, and are smaller than those at Kobe. At Okinawa, 47 % and 62 % of the data pairs for Rayleigh-clear and Mie-cloudy winds versus CDWL winds are obtained below 2 km altitude, respectively. This result is suggested to be linked to the strong convection in the atmospheric boundary layer at Okinawa, especially in summer. The intercepts of the linear regression lines are determined to be -0.88 m s-1 (Rayleigh) and 0.22 m s-1 (Mie) at Kobe and -0.52 m s-1 (Rayleigh) and -0.04 m s-1 (Mie) at Okinawa. The biases are -0.81 m s-1 (Rayleigh) and 0.16 m s-1 (Mie) at Kobe and -0.48 m s-1 (Rayleigh) and -0.26 m s-1 (Mie) at Okinawa. The absolute bias of Rayleigh-clear and Kobe CDWL winds is slightly larger than that for the baseline 2B02, the reason for which is unclear. Except for Rayleigh-clear winds measured at Kobe, the systematic error achieves the mission requirement of 0.7 m s-1. The scaled MADs of Rayleigh-clear (Mie-cloudy) winds are 5.21 (3.92) m s-1 at Kobe and 5.58 (3.86) m s-1 at Okinawa. In contrast to the comparisons of Aeolus and WPR measurements, the random errors are almost the same as those for the baseline 2B02, and no improvement of the random error is evident. As with the scaled MADs, the estimated σAeolus of Rayleigh-clear and Mie-cloudy winds at Kobe and Okinawa is almost the same as that for the baseline 2B02.

For the validation of the Aeolus wind products, we launched 12 and 6 GPS-RSs from NICT Okinawa during the baseline 2B02 and 2B10 periods, respectively (Table 4). The GPS-RSs obtained wind profiles with a vertical range up to 25 km. Thus, the GPS-RSs could measure winds of the upper troposphere and lower stratosphere, which cannot be measured by the WPRs and CDWLs.

Figure 13a shows HLOS wind speed profiles measured by the GPS-RSs with the Rayleigh-clear and Mie-cloudy profiles on 8 November 2018. The Mie-cloudy winds are available below 4.5 km and at high altitudes of 9 to 11.5 km owing to the occurrence of cirrus clouds. A cirrus cloud layer was also observed by the CDWL during the overpass of Aeolus (not shown). There are large deviations between Mie-cloudy and GPS-RS winds below 2 km. Since the horizontal distance between the Mie-cloudy measurements and the GPS-RS is about 100 km in this height region, one can assume that the reason for the large deviations is the spatial heterogeneity of the horizontal wind in the atmospheric boundary layer. The Rayleigh-clear winds show good coverage and closely follow the shape of the wind profile at altitudes higher than 2 km, but there are large deviations between Rayleigh-clear and GPS-RS winds at 3 and 8 km. The scattering ratio on the Rayleigh channel is 1.15, and the relative humidity obtained from the GPS-RS is about 30 % at 3 km. Although there is a valid Mie-cloudy wind at 3 km, it is filtered out due to the HLOS error threshold of 5 m s-1. This suggests that the atmospheric classification in the Rayleigh channel was not working properly, and the crosstalk of Mie signals in the Rayleigh channel could have led to the large deviation. At 8 km, there is no valid Mie-cloudy wind. The scattering ratio and relative humidity are 1.13 and about 20 %, respectively. This suggests that the crosstalk has a small influence on the large deviation. The reason for that is unclear. Since the horizontal distance between the Rayleigh-clear measurements and the GPS-RS is about 80 km in this height region, large horizontal wind gradients in this height region potentially have an influence on the deviation. The subtropical jet stream with westerly winds can be seen in the GPS-RS and Rayleigh-clear observations at around 14 km. A maximum absolute wind speed higher than 50 m s-1 was observed in this height region according to the high-resolution GPS-RS profile. Despite the coarse range resolution (2 km) of the Aeolus measurements in this height region, the Rayleigh-clear winds are able to detect the high wind speed.

The second case discussed in this study is from 1 December 2018 (Fig. 13b). The Mie-cloudy winds are available below 4 km. As compared with the previous case (8 November 2018), the Mie-cloudy winds agree with the GPS-RS winds in the lowermost 2 km. The reason for the agreement is that the Aeolus ground track was relatively near the radiosonde launching position (about 50 km). The Rayleigh-clear winds are available at altitudes higher than 2 km. The occurrence of cloud was sporadically detected by the CDWL, and the relative humidity obtained from the GPS-RS was about 90 % at 3 to 4 km altitude (not shown). It is assumed that the clouds were partly existent in the Aeolus observational domain. The Rayleigh-clear wind shows a large bias at 3 to 4 km altitude, but the scattering ratio on the Rayleigh channel is 1.15. This suggests that there was an issue with the crosstalk correction of Mie signals in the Rayleigh channel. As with the previous case, the subtropical jet stream with westerly winds is seen in the GPS-RS and Rayleigh-clear observations at around 12 km altitude. The Rayleigh-clear wind measurements can detect the high wind speed, but they are slightly overestimated; the reason for that is unclear. Potentially, large horizontal wind gradients in this height region have an influence on the differences.

Figure 13c and d show the correlation plots of the Rayleigh-clear and Mie-cloudy HLOS winds against GPS-RS HLOS winds during the baseline 2B02 period, respectively. Summaries of the statistical parameters retrieved from the scatter plot analysis for the baseline 2B02 and 2B10 are given in Table 8. A linear trend between Aeolus and GPS-RS measurements is clearly seen from the linear regression. The linear regression line has a slope of 0.99 (0.97), with an intercept of 1.00 (2.07) m s-1 for the comparison of Rayleigh-clear (Mie-cloudy) and GPS-RS winds. Both Rayleigh-clear and Mie-cloudy winds exhibit a slightly positive bias. The different colours indicate whether Aeolus had an ascending orbit (red) or a descending orbit (blue). No significant difference is found between the ascending and descending orbits. The biases of Rayleigh-clear and Mie-cloudy winds are 1.00 and 2.15 m s-1, respectively. These values are almost the same as the intercept of the linear regression line. The random error represented by the scaled MAD of Rayleigh-clear winds (4.77 m s-1) is slightly larger than that of Mie-cloudy winds (4.14 m s-1). Baars et al. (2020) compared the Rayleigh-clear and Mie-cloudy winds with winds obtained from the radiosonde launches on board the German RV Polarstern during cruise PS116 carried out in the Atlantic Ocean west of the African continent from 17 November to 10 December 2018 (i.e. during the baseline 2B02 period). They reported biases of 1.52 and 0.95 m s-1 with random errors of 4.84 and 1.58 m s-1 for Rayleigh-clear and Mie-cloudy winds, respectively. The slope and intercept of the linear regression line were 0.97 (0.95) and 1.57 (1.13) m s-1 for the comparison of Rayleigh-clear (Mie-cloudy) and radiosonde winds, respectively. Therefore, the slightly positive bias of Rayleigh-clear versus GPS-RS winds obtained in this study is almost the same as that obtained by Baars et al. (2020). The bias of Mie-cloudy versus GPS-RS winds is larger than that from Baars et al. (2020). The result that the random error of Mie-cloudy winds is much smaller than that of Rayleigh-clear wind contrasts with our results. The discrepancies are probably caused by different observation location, meteorological conditions, and distance between the measurements.

Figure 14a and b show HLOS wind speed profiles measured by the GPS-RS with the Rayleigh-clear and Mie-cloudy profiles on 19 and 21 December 2019, respectively. The range-bin settings of Aeolus were changed to a resolution of 1 km up to an altitude of 19 km on 26 February 2019. Owing to the high range resolution, the Rayleigh-clear wind measurements of Aeolus can detect the rapid changes in the wind speed profiles in the subtropical jet stream. On 19 December 2019, the CDWL observed a cloud layer at around 1 km under rainy conditions during the overpass of Aeolus (not shown), and the Mie-cloudy winds were detected above the cloud layer. On 21 December 2019, the CDWL observed multiple cloud layers up to about 9 km during the overpass of Aeolus (not shown), and the Mie-cloudy winds were detected at these cloud layers.

Figure 14c and d show the correlation plots of the Rayleigh-clear and Mie-cloudy HLOS winds against GPS-RS HLOS winds during the baseline 2B10 period, respectively. As with the baseline 2B02, a linear trend between Aeolus and GPS-RS observations is clearly seen from the linear regression. The linear regression line has a slope of 1.01 (0.92) with an intercept of 0.38 (-0.22) m s-1 for the comparison of Rayleigh-clear (Mie-cloudy) and GPS-RS winds. The intercepts of Rayleigh-clear and Mie-cloudy winds are smaller than those for the baseline 2B02. As with the baseline 2B02, no significant difference is found between the ascending and descending orbits. The biases of Rayleigh-clear and Mie-cloudy winds are 0.45 and -0.71 m s-1, respectively. Both Rayleigh-clear and Mie-cloudy winds generally meet the mission requirements on systematic errors. These values are almost the same as the intercept of the linear regression line and are smaller than those for the baseline 2B02. The scaled MAD of Rayleigh-clear winds is 3.97 m s-1 and smaller than that for the baseline 2B02. On the other hand, the scaled MAD of Mie-cloudy wind is 3.99 m s-1 and almost the same as that for the baseline 2B02.

Martin et al. (2021) estimated the radiosonde representativeness error σr_GPS-RS by considering spatial and temporal displacements and the different measurement geometries of the radiosonde and the Aeolus observations. They determined that the radiosonde representativeness error σr_GPS-RS is 2.48 m s-1 for the Rayleigh-clear winds, 2.49 m s-1 for the Mie-cloudy winds with 90 km horizontal resolution (corresponding to the baseline 2B02), and 2.66 m s-1 for the Mie-cloudy winds with 10 km horizontal resolution (corresponding to the baseline 2B10) based on the radiosonde and the Aeolus observations. The Aeolus random error σAeolus considering the representativeness error σr_GPS-RS in addition to the radiosonde observational error can be calculated as follows: 7σAeolus=σval2-σr_GPS-RS2-σGPS-RS2. σAeolus is determined using the Eq. (7) to be 4.01 m s-1 for Rayleigh-clear winds and 3.24 m s-1 for Mie-cloudy winds during the baseline 2B02 period. During the baseline 2B10 period, σAeolus is determined to be 3.02 m s-1 for Rayleigh-clear winds and 2.89 m s-1 for Mie-cloudy winds. Martin et al. (2021) estimated σAeolus using the radiosonde observations in the mid-latitudes of the Northern Hemisphere (23.5 to 65∘ N), resulting in σAeolus of 4.23 to 4.37 m s-1 for Rayleigh-clear winds, 2.60 to 2.76 m s-1 for Mie-cloudy winds with 90 km horizontal resolution, and 2.97 to 3.03 m s-1 for Mie-cloudy winds with 10 km horizontal resolution. Given that estimates of the representativeness error exhibit large uncertainties (Martin et al., 2021), the Rayleigh-clear and Mie-cloudy wind random errors during the baseline 2B02 period are consistent with the validation results of Martin et al. (2021). During the baseline 2B10 period, the Mie-cloudy wind random error is also in good agreement with the validation result of Martin et al. (2021), whereas the Rayleigh-clear wind random error significantly decreases. Both Rayleigh-clear and Mie-cloudy wind random errors are close to the mission requirement of 2.5 m s-1 in the free troposphere.