Validation of Aeolus Level 2B wind products using wind profilers, ground-based Doppler wind lidars, and radiosondes in Japan

The first space-based Doppler wind lidar (DWL) onboard the Aeolus satellite was launched by the European Space Agency (ESA) on 22 August 2018 to obtain global profiles of horizontal line-of-sight (HLOS) wind speed. In this 10 study, the Raleigh-clear and Mie-cloudy winds for periods of baseline 2B02 (from 1 October to 18 December 2018) and 2B10 (from 28 June to 31 December 2019 and from 20 April to 8 October 2020) were validated using 33 wind profilers (WPRs) installed all over Japan, two ground-based coherent Doppler wind lidars (CDWLs), and 18 GPS-radiosondes (GPSRSs). In particular, vertical and seasonal analyses were performed and discussed using WPR data. During the baseline 2B02 period, a positive bias was found to be in the ranges of 0.46–1.69 m s–1 for Rayleigh-clear winds and 1.63–2.42 m s–1 for 15 Mie-cloudy winds using the three independent reference instruments. The biases of Rayleigh-clear and Mie-cloudy winds were in the ranges of –0.82–+0.45 m s–1 and –0.71–+0.16 m s–1 during the baseline 2B10 period, respectively. The systematic error for the baseline 2B10 was improved as compared with that for the baseline 2B02. The vertical analysis using WPR data showed that the systematic error was slightly positive in all altitude ranges up to 11 km during the baseline 2B02 period. During the baseline 2B10 period, the systematic errors of Rayleigh-clear and Mie-cloudy winds were improved 20 in all altitude ranges up to 11 km as compared with the baseline 2B02. Immediately after the launch of Aeolus, both Rayleigh-clear and Mie-cloudy biases were small. Within the baseline 2B02, the Rayleigh-clear and Mie-cloudy biases showed a positive trend. For the baseline 2B10, the Rayleigh-clear wind bias was generally negative at all months except August 2020, and Mie-cloudy wind bias gradually fluctuated. The systematic error was close to zero with time in 2020 and did not show a marked seasonal trend. The dependence of the Rayleigh-clear wind bias on the scattering ratio was 25 investigated, showing that the scattering ratio had a minimal effect on the systematic error of the Rayleigh-clear winds during the baseline 2B02 period. On the other hand, during the baseline 2B10 period, there was no significant bias dependence on the scattering ratio. Without the estimated representativeness error associated with the comparisons using WPR observations, the Aeolus random error was determined to be 6.71 (5.12) and 6.42 (4.80) m s–1 for Rayleigh-clear (Miecloudy) winds during the baseline 2B02 and 2B10 periods, respectively. The main reason for the large random errors is 30 probably related to the large representativeness error due to the large sampling volume of the WPRs. Using the CDWLs, the Aeolus random error estimates were in the range of 4.49–5.31 (2.93–3.19) and 4.81–5.21 (3.30–3.37) m s–1 for Rayleighhttps://doi.org/10.5194/amt-2021-243 Preprint. Discussion started: 10 August 2021 c © Author(s) 2021. CC BY 4.0 License.

radiosondes, wind profilers (WPRs), ground-based Doppler wind lidars (DWLs), and aircrafts. The current global observation network provides accurate and precise vertical wind profiles. However, the observational coverage is limited from the global perspective. Wind vectors can be measured by satellite-borne microwave scatterometers and polarimetric 45 microwave radiometers, and the multiple-layer wind vector, called the atmospheric motion vector (AMV), can be retrieved from cloud and water vapour motions derived from geostationary and polar-orbit satellite images. Although these sensors have a large coverage area and high temporal and horizontal resolutions, they lack or have significantly limited vertical sounding capability.
A space-based DWL is a powerful remote sensing instrument for global wind profiling. The European Space Agency (ESA) 50 launched on 22 August 2018 the first space-based DWL, Aeolus, for obtaining global wind profiles (Kanitz et al., 2019;Reitebuch et al., 2020). Aeolus carries a single payload, named Atmospheric Laser Doppler Instrument (ALADIN).
ALADIN uses a single-frequency UV laser and a direct-detection system and provides profiles of a single line-of-sight (LOS) wind speed on a global scale from the ground up to about 30 km in the stratosphere (ESA, 1999;Stoffelen et al., 2005Stoffelen et al., , 2020Reitebuch, 2012;Kanitz et al., 2019). The main purpose of Aeolus is to provide global wind profiles with vertical 55 resolution and wind observation accuracy that meet for the World Meteorological Organization (WMO) observation requirements to improve NWP and to fill the gap of the current global wind observation systems. Its other main purpose is to contribute to research on the energy balance, atmospheric circulation, precipitation system, southern vibration phenomenon, stratosphere/troposphere exchange, and so forth.
The new remote sensing technology and retrieval algorithm require a careful assessment of the quality and validity of the 60 generated data products before releasing them to the user community. ESA released an Announcement of Opportunity (AO) in 2007 and 2014 calling for calibration and validation (CAL/VAL) proposals for Aeolus. The CAL/VAL activities include a full assessment of all aspects of the DWL wind measurement performance and stability. The National Institute of Information and Communications Technology (NICT) has applied to contribute to CAL/VAL activities for Aeolus in East https://doi.org/10.5194/amt-2021-243 Preprint. Discussion started: 10 August 2021 c Author(s) 2021. CC BY 4.0 License.
Asia and the Western Pacific region. Continuous validation of horizontal LOS (HLOS) wind speed after calibration 65 processes is important in order to contribute to the L2C product, which results from the background assimilation of the Aeolus HLOS winds in the European Centre for Medium-Range Weather Forecasts (ECMWF) operational prediction model.
The purposes of the project are to contribute to reducing uncertainty in Aeolus wind measurements, to validate processes for improving HLOS wind speed measured by Aeolus, and to assess the quality of wind data.
The aim of this paper is therefore to validate the quality of the Aeolus HLOS winds over Japan by using measurements from 70 WPRs, ground-based coherent Doppler wind lidars (CDWLs), and GPS-radiosondes (GPS-RSs). The paper is organized as follows. First, an overview of Aeolus and ALADIN is provided. Section 3 describes the WPR, CDWL, and GPS-RS instrument setups and measurement procedures. The procedure of matching the Aeolus measurements with the reference instruments' measurements is also described in Sect. 3. The intercomparison and statistical methods are addressed in Sect. 4.
Section 5 presents statistical comparisons between the Aeolus measurements and the WPR, CDWL, and GPS-RS 75 measurements. In Sect. 6, the main findings are summarized.

Overview of Aeolus and ALADIN
Aeolus flies in a sun-synchronous polar orbit (inclination 97°) at an altitude of about 320 km, with a period of about 90 min and a seven-day repeat cycle. The typical ground tracks of Aeolus over Japan are shown in Fig. 1. The red and blue lines 80 Red and bule lines represent the typical Aeolus ground tracks for ascending and descending orbits, respectively. https://doi.org/10.5194/amt-2021-243 Preprint. Discussion started: 10 August 2021 c Author(s) 2021. CC BY 4.0 License. represent the Aeolus ground tracks for ascending and descending orbits, respectively. The principal components of ALADIN are two fully redundant diode-pumped single-frequency continuous-wave neodymium-doped yttrium-aluminium-garnet (Nd:YAG) lasers and two diode-pumped Q-switched Nd:YAG lasers (Flight Model A (FM-A) and FM-B) with power amplifiers, a 1.5-m-diameter afocal Cassegrain telescope, a direct-detection receiver, and signal processing devices. The single-frequency Q-switched Nd:YAG lasers with a 1064.4 nm operating wavelength emit about 250 mJ output energy with 85 a 20 ns pulse width (full width at half maximum) operating at a pulse repetition frequency (PRF) of 50.5 Hz. Nonlinear lithium triborate crystals are used to generate the UV laser pulses with a 354.8 nm operating wavelength. The singlefrequency Q-switched UV laser emits about 60 mJ output energy at the PRF of 50.5 Hz (Lux et al., 2020a) and a laser beam divergence of 20 µrad. The laser pulses are directed downward to Earth at an off-nadir angle of 35° and enter at an incident angle of about 37.6° at the sea and land surfaces due to Earth's curvature. The FM-A laser was used until the middle of June 90 2019, and the FM-B laser has been used since 28 June 2019. The direct-detection receiver consists of the Cassegrain telescope, two interferometers, and two accumulation charge-coupled devices (ACCDs). The signal backscattered by moving atmospheric molecules (Rayleigh scattering) and aerosol and cloud particles (Mie scattering) is collected by the afocal Cassegrain telescope. One of the two interferometers uses the double-edge technique using two Fabry-Perot interferometers (Chanin et al., 1989;Flesia and Korb, 1999;Flesia and Hirt, 2000;Gentry et al., 2000), which is mainly 95 sensitive to atmospheric molecules (Rayleigh channel). The other one uses a spectrometer based on a Fizeau interferometer (Schillinger et al., 2003;Morancais et al., 2004), which is sensitive to aerosol and cloud particles (Mie channel). Both the Fizeau and Fabry-Perot interferometers act as a narrowband filter. The signals for Rayleigh and Mie channels are imaged on each ACCD after passing through some optics (Weiler et al., 2020). The signals imaged on the two ACCDs are converted to electrical signals and stored. 100 We used the Aeolus Level 2B (L2B) data products of Raleigh and Mie channels during the baseline 2B02 and 2B10 periods. The L2B data products including a primary mirror correction (baseline 2B10; Rennie and Isaksen, 2020) have been available for new observations since April 2020. The homogeneous reprocessed data are also available using the baseline 2B10 from 28 June to 31 December 2019. In this study, we used three different periods to assess L2B data products, from 1 October to 18 December 2018 (baseline 2B02), and from 28 June to 31 December 2019 and from 20 April to 8 October 2020 105 (baseline 2B10), where a consistent re-analysed dataset is available. The first period was within the commissioning phase.
We mainly discuss the measurement performance of Aeolus for Rayleigh-clear and Mie-cloudy winds. Rayleigh-clear winds refer to wind observations in an aerosol-free atmosphere. Mie-cloudy winds refer to winds acquired from Mie backscattered signals induced by aerosols and clouds (Witschas et al., 2020). The quality of the Aeolus wind data is indicated by validity flags. The validity flag (de Kloe et al., 2016) considers the validity of the products. Several different technical, instrumental, 110 and retrieving checks account for this flag. It has the value 1 (valid) or 0 (not valid). We only used Aeolus products with a validity flag of 1. We also used HLOS estimated errors (theoretical) of the L2B data products. The estimated error is a theoretical value, which is estimated on the basis of measured signal levels as well as the temperature and pressure sensitivities of the Rayleigh channel response (Dabas et al., 2008).

Wind profilers
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 120 operate continuously, acquiring vertical profiles of horizontal wind speed, wind direction, vertical velocity, and signal-tonoise 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 125 294m and 11.6 km above the wind profiler, respectively. There are 40 range bins for one wind profile. There is no significant difference between wind profiler winds and radiosonde winds in the biases and root mean square errors (Ishihara et al., 2006). The random error (root mean square error) 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 spatiotemporal variation in the wind field. 130 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 is set to be less than 100 km. To achieve temporal synchronization, we use averages of WPR wind data from 30 min before to 30 min after the passage of Aeolus. After temporal and spatial collocation, the Aeolus L2B wind product 135 closest to each WPR measurement is 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.

Coherent Doppler wind lidars 140
NICT has installed 1.54-µm CDWLs (WINDCUBE 400S manufactured by LEOSPHERE; Cariou et al., 2006) in Kobe (34.66ºN, 135.16ºE; magenta circle in Fig. 1) and Okinawa (26.50ºN,127.84ºE; 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 MSL) 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 145 center 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) 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 10,000 times on average. Since the PRF was 10 kHz, the accumulation time of each beam was 10 s. The 150 Doppler wind speed at each bin was estimated from the averaged Doppler-shifted frequency spectra using the maximum likelihood estimator. 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. Random errors were 0.02 to 0.10 m s -1 from -10 to -30 dB wideband SNR and the CDWLs operate 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 155 systematic error and random error (root mean square error) 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 160 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. 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 165 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.

Radiosondes 170
Twelve GPS-radiosondes (GPS-RSs) of type RS41-SGP produced by Vaisala were launched from NICT Okinawa (26.50ºN, 127.84ºE; 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 Upper-Air Network (GRUAN), the measurement uncertainties of the horizontal wind speed − 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 180 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 is about 45 min when they reach 185 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 are considered when defining collocation criteria for comparisons of Aeolus and GPS-RS measurements. In this study, the GPS-RS measurements that are within 120 km horizontal distance and 60 min temporal difference from the Aeolus measurements are used for the validation.

4 Intercomparison and statistical methods
There is a difference in the vertical resolution between Aeolus measurements and WPR, CDWL, and GPS-RS measurements.
The horizontal wind speed and wind direction measured by the WPRs, CDWLs, and GPS-RSs are averaged to the Aeolus bin by using the top and bottom altitudes given in the Aeolus L2B data product. All valid averaged wind speeds To validate the quality of Aeolus HLOS winds ( ), the difference from the corresponding WPR, CDWL, and GPS-RS winds projected onto the Aeolus viewing direction ( Following Witschas et al. (2020), the difference between Aeolus HLOS winds and WPR HLOS winds ( ) can be used to verify the thresholds for the estimated HLOS error provided in the Aeolus L2B data product during the baseline   (Rennie and Isaksen, 2020) and those adopted in other validation studies (e.g., Baars et al., 2020).
To evaluate the results of comparison between Aeolus HLOS winds and reference instruments' HLOS winds, we use mean differences (BIAS) and the standard deviation (STD) of the differences as: 215 where N is the number of available data points. In addition to the STD, the scaled median absolute deviation (scaled MAD) is calculated as MAD is used as a very robust measure for the variability of the Aeolus HLOS winds because it is less sensitive to outliers than the STD (Lux et al., 2020b;Witschas et al., 2020;Baars et al., 2020;Rennie and Isaksen, 2020;Martin et al., 2021).
When a data set follows a normal distribution, the MAD value multiplied by 1.4826 (scaled MAD) is identical to the STD (Ruppert and Matteson, 2015). By assuming independence between Aeolus measurements and reference instruments' measurements, the total variance of the difference between them (squared scaled MAD) ( 2 ) is the sum of the variance 225 resulting from the Aeolus random error ( 2 ) and the variance resulting from reference instruments' random error ). Thus, the Aeolus random error is calculated as where , , and − are assumed to be 3, 2, and 0.7 m s −1 , respectively (see Section 2.2, 2.3, and 2.4). Note that this estimation of includes the representativeness error due to the spatial and temporal mismatch between Aeolus and 230 reference instruments' measurements. In addition to the BIAS, STD, and scaled MAD, the correlation coefficient (R) between Aeolus HLOS winds and reference instruments' HLOS winds, and the slopes and intercepts of the linear regression lines are used to evaluate the results of comparison. The increased number of data pairs can be explained by there being about twice as many periods for the baseline 2B10. The energy decrease in the FM-A laser during the baseline 2B02 period led to fewer Rayleigh-clear winds that can be used for 245 the comparison. Since 5 March 2019, Aeolus Mie-cloudy winds have been processed with a smaller horizontal averaging length of down to 10 km, also leading to more Mie-cloudy winds that can be used for comparison during the baseline 2B10 period. Furthermore, the range-gate settings of Aeolus were changed on 26 February 2019, which also increased the number of available data points during the baseline 2B10 period.
During the baseline 2B02 and 2B10 periods, the linear trend between the Rayleigh-clear (Mie-cloudy) winds and WPR 250  winds is clearly seen for all data and both orbit phases (Figs. 4 and 5). Although the Rayleigh-clear winds for all data and both orbit phases exhibit a slightly positive bias between 1.6 and 1.8 m s -1 during the baseline 2B02 period (Figs. 4a-c), no significant wind-speed-dependent bias is apparent. However, the systematic errors (biases) obtained in this study are higher than those of 0.7 m s -1 stipulated in the mission requirements (Ingmann and Straume, 2016). for ascending and descending orbits, respectively. These results indicate that the performance of Aeolus for Mie-cloudy winds is reliable over Japan. The biases of Mie-cloudy versus WPR winds are slightly negative for all data and both orbit phases (-0.51, -0.73, and -0.29 m s -1 ), but these values are smaller than those of the Rayleigh-clear winds. As with the 300 Rayleigh-clear winds, the absolute bias is slightly larger for the ascending orbit than for the descending orbit. The small bias, slope close to 1, and high correlation coefficient agree well with those reported by Guo et al. (2021). The scaled MADs are relatively large (5.56-5.66 m s -1 ), but the values are smaller than those of the Rayleigh-clear winds.
To summarize, the systematic and random errors of Rayleigh-clear (Mie-cloudy) versus WPR winds for the baseline 2B10 are improved as compared with those for the baseline 2B02. In contrast to the baseline 2B02, the systematic error of Mie-305 cloudy winds is superior to that of Rayleigh-clear winds during the baseline 2B10 period. During the baseline 2B02 period, the systematic error was significantly larger than the strict mission requirement of 0.7 m s -1 specified for Aeolus HLOS winds. During the baseline 2B10 period, both Rayleigh-clear and Mie-cloudy winds have generally met the mission requirements on systematic errors. However, the Aeolus random error of Rayleigh-clear and Mie-cloudy winds is considerably larger than the required precision of 2.5 m s -1 in the free troposphere during the baseline 2B02 and 2B10 310 periods. The main reason for not yet achieving the mission requirement for random errors is probably related to the large representativeness error due to the large sampling volume of the WPR. From the statistical comparisons, we found no significant difference between the ascending and descending orbits with respect to the Rayleigh-clear and Mie-cloudy winds during the baseline 2B02 and 2B10 periods.

Vertical distribution of wind differences
The vertical distributions of the bias and standard deviation of the differences between Aeolus and WPR HLOS winds for baseline 2B02 are shown in Fig. 6. The values are binned into bins of 1 km height. The bias uncertainties estimated at 90 % confidence level for all data are reasonably small up to about 10 km altitude where there are relatively many paired data points for comparison (Fig. 6a). For all data, the biases of Rayleigh-clear and WPR HLOS winds are significantly positive in 320 all altitude ranges and less than 3.53 m s -1 up to 10 km. The larger standard deviations at 0−2 km altitude for ascending and descending orbits are caused by fewer paired data points. For Mie-cloudy winds, the biases for all data are also significantly positive in all altitude ranges except for 10-11 km (Fig. 6d). Although the biases are also positive below 8 km during ascending and descending orbits, the vertical distributions of bias are opposite to each other above 8 km. The mission requirement of 0.7 m s -1 has not been achieved by both Rayleigh-clear and Mie-cloudy biases in all altitude ranges. 325 https://doi.org/10.5194/amt-2021-243 Preprint. Discussion started: 10 August 2021 c Author(s) 2021. CC BY 4.0 License.
The same statistics are shown for the baseline 2B10 in Fig. 7. As with the baseline 2B02, the bias uncertainties estimated at 90 % confidence level are reasonably small up to about 11 km altitude where there are relatively many paired data points for comparison. For all data, the biases of Rayleigh-clear and WPR HLOS winds are slightly negative in all altitude ranges and less than -1.6 m s -1 up to 11 km (Fig. 7a). The systematic error was less than that of the baseline 2B02. Below 2 km altitude, the Rayleigh-clear winds have met the mission requirements for systematic errors. The bias and standard deviation in the 330 altitude range of 0-1 km (atmospheric boundary layer) are almost the same as those in the upper level. This suggests that the correction scheme against the scattering ratio was improved in the L2B processor (see Section 5.1.4). However, this result is different from that in the other validation studies conducted during the baseline 2B10 period (Guo et al., 2021). For the ascending (descending) orbit, the minimum (maximum) bias is -2.34 (0.56) m s -1 in the altitude range of 5-6 (6-7) km. The vertical distributions of bias during ascending and descending orbits are opposite to each other in the altitude range of 3-11 335 km. For all data, the biases of Mie-cloudy and WPR HLOS winds are also slightly negative in all altitude ranges except for 3-4 km (Fig. 7d). As with the Rayleigh-clear winds, the systematic error was improved as compared with that of the baseline 2B02. Below 5 km altitude, Mie-cloudy winds have met the mission requirements on systematic errors. For the ascending (descending) orbit, the minimum (maximum) bias is -1.93 (0.54) m s -1 in the altitude range of 5-6 (4-5) km. As with the Rayleigh-clear winds, the vertical distributions of bias during ascending and descending orbits are opposite to each other in 340 the altitude range of 3-11 km.

Time series variation of wind differences
The time series variation of the bias and standard deviation of the differences between Aeolus and WPR HLOS winds during the baseline 2B02 period are shown in Fig. 8. Immediately after the launch of Aeolus, the biases of the Rayleigh-clear and 345 Mie-cloudy winds are small for all data and both orbit phases. With time, the Rayleigh-clear and Mie-cloudy biases increase for all data and both orbit phases. The Rayleigh-clear bias reached its maximum in January 2019. For the Mie-cloudy winds, the maximums occurred in January and February 2019 for ascending and descending orbits, respectively. Within the baseline 2B02, the Rayleigh-clear and Mie-cloudy biases show a positive trend.
For the baseline 2B10, the same statistics are shown in Fig. 9. For all data, the biases of Rayleigh-clear and WPR HLOS 350 winds are generally negative at all months except August 2020, but the biases do not show a significant seasonal trend (Fig.   9a). The standard deviations of Rayleigh-clear and WPR HLOS data gradually increase with time (from 6.34 to 8.77 m s -1 ). A possible reason is the decrease in the level of the received signal after passing through the telescope. The biases for the ascending orbit are negative throughout the whole period (Fig. 9b). The absolute biases are generally larger for the ascending orbit than for the descending orbit (Figs. 9b and 9c). For all data, the biases of Mie-cloudy and WPR HLOS winds gradually 355 fluctuate and do not show a significant seasonal trend (Fig. 9c). The bias and standard deviation of Mie-cloudy winds are generally smaller than those of Rayleigh-clear winds. There is no significant increase in the standard deviations of Miecloudy winds with time, because the Mie return signal does not depend on the laser energy (Martin et al., 2021). It is interesting to note that the fluctuation of the bias was stronger for the descending orbit than for the ascending orbit in 2019 ( Figs. 9e and 9f). However, the biases for both orbit phases approached zero in 2020. 360

Rayleigh-clear wind bias dependence on scattering ratio
The scattering ratio on the Rayleigh channel is defined as the ratio of the total scattering signal (particles and molecules) to the molecular scattering signal. When the scattering ratio is large, a strong narrowband Mie return signal partly enters the Rayleigh spectrometer, changing the sensitivity of the Rayleigh channel (Witschas et al. 2020). The dependence of the 365 Rayleigh-clear wind bias on the scattering ratio given in the L2B product is shown in Fig. 10. It can be seen that the scattering ratio varied between 1.1 and 1.4 for baseline 2B02 and between 1.05 and 1.65 for baseline 2B10. This means that the determination of the scattering ratio and the threshold for classifying the Rayleigh-clear winds changed between the baselines 2B02 and 2B10. During the baseline 2B02 period, the bias of Rayleigh-clear and WPR HLOS winds slightly increased as the scattering ratio increased (Fig. 10a). Using the L2B products within the commissioning phase, Witschas et al. 370 (2020) reported that the scattering ratio has a considerable influence on the bias of Rayleigh-clear winds. During the baseline 2B10 period, the Rayleigh-clear winds exhibited a slightly negative bias and there was no significant bias dependence on the scattering ratio (Fig. 10b). This means that the correction scheme of the scattering ratio was improved in the L2B processor.

Comparison of Aeolus and CDWL wind data 375
Scatterplots 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 was wider for Kobe than for Okinawa in Fig.  380 11. Both Rayleigh-clear and Mie-cloudy winds exhibit a slightly positive bias. The different colors indicate whether Aeolus had an ascending orbit (red) or descending orbit (blue). There is no significant difference between the ascending and 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-400 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 405 2 km altitude, respectively. This result is suggested to be linked to the strong convection in the atmospheric boundary layer

Comparison of Aeolus and GPS-RS wind data
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 can measure winds of the upper troposphere and lower stratosphere which cannot be measured by the WPRs 420 and CDWLs. 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 Mie-cloudy winds are available below 4.5 km and at high altitudes of 9-11.5 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 430 100 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 second case discussed in this study is from 1 December 2018 (Fig. 13b). The Mie-cloudy winds are available below 4 km. Since the occurrence of cloud was sporadically detected by the CDWL at 3−4 km altitude (not shown), it is assumed that 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. 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 which is unclear. There is a possibility regarding horizontal wind 440 gradients in this height region. Figures 13c and 13d 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  of Rayleigh-clear (Mie-cloudy) and radiosonde winds, respectively. Therefore, the slightly positive biases of Rayleigh-clear and Mie-cloudy versus GPS-RS winds obtained in this study are almost the same as those obtained by 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. 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 _ − , and error sources caused by spatial and temporal displacements need to be considered, in addition to the different measurement geometries of the radiosonde and the Aeolus observations. They determined that the radiosonde representativeness errors _ − is 2.48 m s -1 for the Rayleighclear winds, 2.49 m s -1 for the Mie-cloudy winds with 90 km horizontal resolution (corresponding to the baseline 2B02), and 480 2.66 m s -1 for the Mie-cloudy winds with 10 km horizontal resolution (corresponding to the baseline 2B10). The Aeolus random error is calculated by using the radiosonde representativeness error _ − , was 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, was determined to be 3.

Summary
We validated the Aeolus L2B data product for Rayleigh-clear and Mie-cloudy winds using operational WPRs, ground-based CDWLs, and GPS-RSs in Japan during the periods of the baseline 2B02 (from 1 October to 18 December 2018) and 2B10 (from 28 June to 31 December 2019 and from 20 April to 8 October 2020). Statistical analyses based on the three independent reference instruments were performed to validate the Rayleigh-clear and Mie-cloudy wind data. In the 500 comparisons of Aeolus and WPR measurements, the vertical distribution of wind difference, the wind bias dependence on latitude and orbit phases, the time series variation of wind differences, and the Rayleigh-clear wind bias dependence on the scattering ratio were investigated.
Statistical analyses of Aeolus HLOS winds and WPR HLOS winds for Rayleigh-clear and Mie-cloudy winds were carried out. For the baseline 2B02, the systematic error was determined to be 1.69 m s -1 for Rayleigh-clear winds and 2.42 m s -1 for 505 Mie-cloudy winds. For the baseline 2B10, the systematic error was determined to be -0.82 m s -1 for Rayleigh-clear winds and -0.51 m s -1 for Mie-cloudy winds. The systematic error for the baseline 2B10 was less than that for the baseline 2B02.
For the baseline 2B02, was determined to be 6.71 m s -1 for Rayleigh-clear winds and 5.12 m s -1 for Mie-cloudy winds. For the baseline 2B10, was determined to be 6.42 m s -1 for Rayleigh-clear winds and 4.80 m s -1 for Miecloudy winds. The main reason for the large Aeolus random errors is probably related to the large representativeness error 510 due to the large sampling volume of the WPR.
The vertical distributions of differences between Rayleigh-clear or Mie-cloudy winds and WPR winds showed that both Rayleigh-clear and Mie-cloudy biases in all altitude ranges up to 11 km were significantly positive during the baseline 2B02 period. During the baseline 2B10 period, the systematic errors of Rayleigh-clear and Mie-cloudy winds were improved as compared with those during the baseline 2B02 period. The time series of wind speed differences between Aeolus and WPR 515 HLOS winds varied considerably during baseline 2B02 period. Immediately after the launch of Aeolus, both Rayleigh-clear and Mie-cloudy biases were small. With time, the Rayleigh-clear and Mie-cloudy biases increased. Within the baseline 2B02, the Rayleigh-clear and Mie-cloudy biases showed a positive trend. For the baseline 2B10, the biases of Rayleigh-clear and WPR HLOS winds were generally negative at all months except August 2020, but the biases did not show a clear seasonal trend. The biases of Mie-cloudy and WPR HLOS winds gradually fluctuated and did not show a clear seasonal trend. The 520 Rayleigh-clear and Mie-cloudy wind biases were close to 0 m s -1 in 2020. The dependence of the Rayleigh-clear wind bias on the scattering ratio was investigated, showing that the scattering ratio has a minimal effect on the systematic error of the Rayleigh-clear winds during the baseline 2B02 period. On the other hand, during the baseline 2B10 period, there was no significant bias dependence on the scattering ratio.
The statistical analyses based on the ground-based CDWLs at Kobe and Okinawa during the baseline 2B02 and 2B10 525 periods showed that the agreement between the Aeolus winds and CDWL winds is generally good. For the baseline 2B02, the systematic error was determined to be 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. Except for the Rayleigh-clear winds measured at Kobe, the systematic error did not achieve the https://doi.org/10.5194/amt-2021-243 Preprint. Discussion started: 10 August 2021 c Author(s) 2021. CC BY 4.0 License. mission requirement.
was determined to be 4.49 m s -1 (Rayleigh) and 2.93 m s -1 (Mie) at Kobe and 5.31 m s -1 (Rayleigh) and 3.19 m s -1 (Mie) at Okinawa. The Aeolus random errors were larger than those from the validation study 530 using the airborne 2 µm CDWL (Witschas et al., 2020). The discrepancies were probably caused by the smaller representativeness error due to the spatial and temporal displacements between Aeolus and airborne CDWL measurements.
For the baseline 2B10, the systematic error was determined to be -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. In contrast to the baseline 2B02, the systematic error significantly decreased except for the Rayleigh-clear winds measured at Kobe. was determined to be 4.81 m s -1 (Rayleigh) and 535 3.37 m s -1 (Mie) at Kobe and 5.21 m s -1 (Rayleigh) and 3.30 m s -1 (Mie) at Okinawa. In contrast to the comparisons of Aeolus and WPR measurements, the Aeolus random errors were almost the same as those for the baseline 2B02, and no improvement of the Aeolus random error was evident.
With the analyses of results obtained from GPS-RSs launched from NICT Okinawa, it was shown that Aeolus can measure accurately wind profiles with a vertical range up to 25 km and capture the rapid changes in the wind speed profiles such as 540 the subtropical jet stream. The statistical analyses based on the GPS-RSs also revealed the good performance of Aeolus during the baseline 2B02 and 2B10 periods. For the baseline 2B02, the systematic error was determined to be 1.00 m s -1 for Rayleigh-clear winds and 2.15 m s -1 for Mie-cloudy winds. For the baseline 2B10, the systematic error was determined to be 0.45 m s -1 for Rayleigh-clear winds and -0.71 m s -1 for Mie-cloudy winds. Both Rayleigh-clear and Mie-cloudy winds generally met the mission requirements on systematic errors. By taking the radiosonde representativeness error into account, 545 was determined to be 4.01 m s -1 for Rayleigh-clear winds and 3.24 m s -1 for the Mie-cloudy winds during the baseline 2B02 period. During the baseline 2B10 period, was determined to be 3.02 m s -1 for Rayleigh-clear winds and 2.89 m s -1 for the Mie-cloudy winds. The random errors of the Rayleigh-clear and Mie-cloudy winds during the baseline 2B02 period were in line with the other validation results. During the baseline 2B10 period, the Aeolus random errors of the Rayleigh-clear and Mie-cloudy winds were improved as compared with those during the baseline 2B02 period. 550

Data availability
The CDWL and GPS-RS data used in this paper can be provided by the corresponding author (iwai@nict.go.jp) upon request.