The near polar-orbiting and Sun-synchronous Aeolus satellite measures global atmospheric wind profiles along the DWL's line of sight (LOS) from the Earth's surface to the lower stratosphere (Straume, 2018). The LOS of Aeolus is perpendicular to its orbital velocity to mitigate contributions from its along-orbit velocity. Its DWL, the Atmospheric LAser Doppler INstrument (ALADIN, Guo et al., 2021), points 35∘ from the nadir and includes two receivers to measure the Doppler shift from the emitting laser along the LOS: a double Fabry–Pérot spectrometer to measure Rayleigh scattering from air molecules and a Fizeau spectrometer to measure Mie scattering from cloud droplets and aerosols. The HLOS wind component can be derived by analysing the Doppler frequency shift and assuming that the vertical component of winds is negligible. The wind retrieval method of the processed and calibrated Aeolus Level-2B (L2B) HLOS wind product can be found in the Algorithm Theoretical Basis Documents (Rennie et al., 2020). For both Mie and Rayleigh channels, each measurement bin is classified into “cloudy” or “clear” using its optical property information from Level-1B scattering ratio estimates (Rennie et al., 2020). “Cloudy” classification occurs when the measurement bins have non-zero particle backscatter, while “clear” classification occurs for predominantly molecular backscatter. Since Mie-cloudy and Rayleigh-clear winds are considered to be of superior quality compared to Mie-clear and Rayleigh-cloudy winds (Martin et al., 2021; Guo et al., 2021; Baars et al., 2020), Mie winds and Rayleigh winds refer exclusively to Mie-cloudy and Rayleigh-clear winds in the rest of this study.

The backscattered signal must be horizontally and vertically averaged to obtain a sufficient signal-to-noise ratio (SNR) (Drinkwater et al., 2016; Reitebuch et al., 2020a; Lux et al., 2020). Prior to 5 March 2019, both Rayleigh and Mie winds were averaged to up to a horizontal resolution of 87 km. Recognizing that Mie scattering in cloudy air yields stronger returns than Rayleigh scattering in clear air, after 5 March 2019, the Mie wind product was provided at a finer horizontal resolution of 12 km. The vertical resolution decreases from 500 m in the planetary boundary layer (PBL, defined here as below 2 km in altitude) to 1 km in the free troposphere (defined here as 2 to 16 km in altitude) and to 2 km in the lower stratosphere (above 16 km in altitude). The Mie channel covers the vertical range up to 16 km in altitude, and the Rayleigh channel covers up to 30 km.

Aeolus switched from the first laser, flight model A (FM-A), to the second laser, flight model B (FM-B), due to a decrease in UV power output from FM-A at the end of June 2019 (Reitebuch et al., 2020a; Lux et al., 2020). Aeolus L2B near-real-time baseline products 2B02 and 2B06/07 are used during the early FM-A period (15 September to 16 October 2018) and FM-B period (2 August to 30 September 2019 and 1 December 2019 to 31 January 2020) respectively. The ECMWF published the first reprocessed data in autumn 2020 (2B10; available at ftp://2018_aeolus_l2b:ecmwf@acquisition.ecmwf.int/, last access: 20 October 2020), which covers the period between 24 June and 31 December 2019. The major improvement in this product is a daily updated bias correction accounting for variability of the temperature gradients across the detector telescope's primary M1 mirror; additional improvements are mentioned below and in other studies (e.g. Rennie and Isaksen, 2020; Laroche and St. James, 2021). A comparison of the statistical results during the overlapping period, (boreal) summer 2019, between 2B06 and 2B10 will be presented in this study.

The following data selection is carried out in this study.

The L2B product provides a validation flag of 1 (valid) or 0 (invalid) (de Kloe et al., 2016) associated with each range bin in an observation, and we therefore screen out validation flag value 0 (Baars et al., 2020).

During the early FM-A period, a global constant bias offset of -1.35 ms-1 was added to the Mie winds to bring them into better agreement with the ECMWF model (Rennie and Isaksen, 2020). The Aeolus observation heights were also systematically increased by 250 m due to a known calibration issue. The biases of FM-B HLOS arising mainly from the telescope primary mirror M1 temperature gradients (Rennie and Isaksen, 2020) should be corrected as much as possible before any use for validation against other wind measurements or data assimilation. Fortunately, these biases vary mostly with the orbital node and latitude and partly with longitude and height, facilitating such bias correction. ECCC developed a bias-correction scheme similar to ECMWF, as described in Rennie and Isaksen (2020); see Laroche and St. James (2021). It is a look-up table bias correction based on the mean observation minus the “background” short-range forecast from ECCC (see Sect. 2.2) from the previous 7 d as a function of orbit phase and latitude. It is applied for both Rayleigh and Mie HLOS winds. For the Rayleigh HLOS winds, the correction is also a function of longitude, binned in 10∘ latitude by 36∘ longitude sectors.

To project the wind vector in a given dataset into the Aeolus HLOS, we use 1vHLOS=-usin⁡φ-vcos⁡φ, where vHLOS is the HLOS wind component, u is the zonal wind component, v is the meridional wind component, and φ is the azimuth of the LOS. Conversely, we can also project the HLOS wind vector into the west–east and north–south directions (Wright et al., 2021) for some analysis (Sect. 3.2), using 2vHLOS,u=-vHLOS⋅sin⁡(φ),3vHLOS,v=-vHLOS⋅cos⁡(φ). To repeat, these quantities do not represent zonal and meridional components of the total wind field but the zonal and meridional projections of the vector component of the wind along the HLOS of Aeolus.

The “background” from ECCC, termed “ECCC-B”, is the 9 h short-range forecast used in the operational four-dimensional ensemble-variational (4D-EnVar) data assimilation scheme (Buehner et al., 2015). The forecast model is the operational Global Environmental Multiscale (GEM) (McTaggart-Cowan et al., 2019) with 15 km horizontal grid spacing and 84 vertical levels. There are over 13 million observations assimilated daily during the periods examined in this study, which include data from infrared (56.1 % of all observations assimilated) and microwave (27.7 %) satellite sounders and imagers, aircraft (9.6 %), atmospheric motion vectors (2.3 %), radiosondes (2.1 %), scatterometers (1.0 %), near-surface observations (0.7 %), and satellite-based radio occultation (0.4 %). For the comparison between ECCC-B and Aeolus winds, the closest short-range forecast field, available every 15 min, is selected. Then, this field is linearly interpolated in space to Aeolus measurement locations, first horizontally and then vertically. For the linear interpolation between the model's grid points, the horizontal grid spacing is 15 km and the vertical grid spacing varies from approximately 100 m in the PBL to 1 km in the stratosphere (McTaggart-Cowan et al., 2019). The linear interpolation in time is between two consecutive model states, 15 min apart.

The ERA5 hourly data on 37 pressure levels from the ECMWF are also used in this study in validating the Aeolus measurements. This dataset is based on a four-dimensional variational (4DVar) data assimilation using Cycle 41r2 of the Integrated Forecast System (IFS), which was introduced operationally in 2016. ERA5 provides hourly estimates of atmospheric, land, and oceanic climate variables, available from 1950 to present. Data are gridded on a regular latitude–longitude grid of 0.25∘. A further discussion of the ERA5 configuration can be found in Hersbach et al. (2018, 2020). The process used to match ERA5 and Aeolus data will be discussed in Sect. 2.5.

The CAWS project, led by ECCC, aims to characterize and improve scientific understanding of Arctic weather, climate, and cryospheric systems through enhanced meteorological observation capacity (Joe et al., 2020; Mariani et al., 2018). It also seeks to improve weather forecasts in the Canadian Arctic, test new technologies, and calibrate and validate space-based observations. ECCC's Iqaluit and Whitehorse sites (Fig. 1), so-called “supersites”, were identified as “hotspots” for both extreme weather and transportation infrastructure that merited additional instrumentation. They provide researchers and forecasters with real-time weather observations which can be used in evaluating NWP models. Connected to ECCC's observational science mission, locating these weather stations at high latitudes also tests the ability of the coordinated instrument suites to operate in extreme cold conditions.

The Iqaluit site is situated in a valley to the north-east overlooking Frobisher Bay in the vicinity of 300 m hills. We focus on two instruments at the Iqaluit site that provide wind profile measurements: the radiosonde and Ka-band radar. We will also briefly mention ground-based Doppler lidar measurements in Sect. 3.1. Vaisala RS92 radiosondes (Mariani et al., 2018) were launched twice daily (45 min before synoptic times 00:00 and 12:00 Coordinated Universal Time – UTC). They measure vector wind profiles with a vertical resolution of roughly 15 m depending on ascent speed, up to about 30 km a.g.l. The data used (available at http://weather.uwyo.edu/upperair/sounding.html, last access: 17 August 2020) are the processed radiosonde data provided at mandatory and significant pressure levels (which have a coarser resolution than 15 m). It takes about 2 h to reach 30 km altitude (around 10 hPa). The instrumental uncertainty for the wind speed is between 0.4 and 1.0 ms-1 and between 0.3 and 0.7 ms-1 for the zonal wind component (Dirksen et al., 2014). The error on the zonal wind component due to drift and elapsed time of the ascending balloon is between 0.5 and 1.0 ms-1 in the troposphere and upper troposphere/lower stratosphere (UTLS) (see Fig. 5b in Laroche and Sarrazin, 2013). As a result, the total error for the zonal wind component from these sources of errors is between 0.6 and 1.2 ms-1. Note that the radiosonde data are assimilated in the ECCC and ECMWF systems, which means that the ECCC-B and ERA5 errors are not independent of the radiosonde observation errors. The ECCC Whitehorse site, situated in a wide valley with large lakes, also has radiosondes that operate similarly to the one at Iqaluit.

The dual-polarization Doppler Ka-band radar at Iqaluit measures the LOS wind speed, fog backscatter, and depolarization ratio every 15 min. The radar measures the LOS wind with 14 m resolution, and the LOS range goes from 5 to 30 km, depending on hydrometeor concentration. The uncertainty of the measurements depends on conditions, SNR, and decibel of the return signal. The average vertical wind profile bias to radiosonde is better than 0.3 ms-1. The horizontal winds are derived using a high-angle plan position indicator (PPI) 75∘ scan using the VAD (velocity–azimuth–display) algorithm (Lhermitte and Atlas, 1961; Wang et al., 2010), whereby the radar scans with a fixed elevation angle (φ), while the azimuth angle (θ) is varied. The radial velocity is given by 4vr=usin⁡θcos⁡φ+vcos⁡θcos⁡φ+wsin⁡φ, where w is the vertical wind component. By fitting the data and assuming uniform winds at each range, these three unknown parameters (u, v, and w) can be derived at each vertical level.

Other than the ECCC supersites, we also validate the Aeolus wind product in comparison to radiosonde measurements over the Canadian Arctic at ground stations in Inuvik, Fort Smith, Hall Beach, Cambridge Bay, Norman Wells Ua, and Baker Lake (Fig. 1). They operate similarly to the radiosondes at Iqaluit and Whitehorse and measure vector wind profiles. Some of the stations launch the radiosondes four times a day at 00:00, 06:00, 12:00, and 18:00 UTC. However, this does not affect the temporal criteria (see Sect. 2.5).

For the ground-based validation, the criterion for coincidence of Aeolus overpasses is that the distance from the sites to the measurements be no more than 90 km (horizontal resolution of Rayleigh winds). Using this coincidence criterion, Aeolus overpasses are selected as targets for validation at Iqaluit three times a week at around 21:50, 11:15, and 22:00 UTC and at Whitehorse twice a week at around 02:25 and 15:30 UTC. The Aeolus measurements are compared to the reanalysis and in situ measurements that are available at the nearest time. Temporal sampling for each product is as follows: Aeolus overpasses at Iqaluit and Whitehorse are as mentioned above, reanalysis data are provided hourly, on the hour, radiosonde data are from launches at 00:00 and 12:00 UTC, with a 2 h time of flight to 30 km as mentioned above, and Ka-band radar data are provided via 15 min scans. For example, if Aeolus overpasses selected as a target for validation at the Iqaluit site occur at 11:15 UTC, the Aeolus HLOS profile would be compared to the reanalysis data at 11:00 UTC, to radiosonde measurements at 12:00 UTC, and to the nearest scan by the radar. On the other hand, if the overpass time is 02:25 UTC, the profile would be compared to the ERA5 data at 02:00 UTC, to the radiosonde measurements at 00:00 UTC, and, again, to the nearest scan by the radar.

We evaluated the vertical HLOS wind profile observations from coincident Aeolus overpasses for Iqaluit and Whitehorse against ground-based measurements, ECCC-B, and reanalysis. Our evaluation was limited to the early FM-A period of Aeolus because the Ka-band radar at Iqaluit has been turned off for repairs since 1 August 2019. Figure 2 shows examples of wind profile measurements on (a) 22 September 2018 when Aeolus was in its ascending orbit phase and (b) 24 September 2018 when Aeolus was in its descending orbit phase, at the Iqaluit site. The HLOS wind profile is shown along with profiles of the zonal projection of the HLOS component (dashed curves), vHLOS,u from Eq. (2), for ERA5, radiosonde, Ka-band radar, and lidar. When the measured winds are positive, it means the HLOS winds are directed away from the instrument (eastward for the ascending orbit phase and westward for the descending orbit phase). To ease the interpretation, we plot the negative HLOS winds when Aeolus is in its descending orbit phase (panel b). The Ka-band radar's vertical range extends to less than 5 km in both profiles, around where there are Mie-wind measurements from Aeolus, because its vertical range depends on hydrometeor concentration and its sampling of coincident timing is limited for the Aeolus measurement period. The Iqaluit site also hosts a ground-based Doppler lidar whose LOS wind measurements yield horizontal vector winds via VAD up to about 3 km a.g.l. or the cloud base height. This instrument provides very limited coincident measurement opportunities with Aeolus due to its limited vertical range. Observations from this instrument, which have been extensively validated against high-resolution radiosondes (Mariani et al., 2020) and are useful for boundary-layer focused work, are also shown in Fig. 2 for visual comparison only for 22 and 24 September.

In Fig. 2, we see that Aeolus consistently captures some of the basic structure of the wind profiles compared to in situ measurements, ERA5 reanalysis, and ECCC-B. Because the solid lines lie very close to the dashed lines, it is evident that Aeolus is providing predominantly zonal wind information even at high latitudes (63∘ N), where the LOS has a greater meridional component than at low latitudes. On 22 September, Aeolus detects an easterly wind feature in the lower atmosphere and accurately picks up the change in sign around 5 km altitude. On 24 September, although a few of the Rayleigh measurements have a deviation close to 50 % from the other datasets around 8 km a.g.l., Aeolus still measures westerly winds in reasonable overall agreement with the other data.

Although the Ka-band radar offers limited sampling, it is retained in this analysis because it offers an entirely independent and unique set of observations in the Canadian North that are not assimilated in any NWP model. Furthermore, it provides consistent measurements with the radiosondes, as shown in Fig. 3. For the same period of analysis and when the radar observations are within 30 min of radiosonde launch, the bias of the wind speed between the radar and radiosonde is less than 1 ms-1 for measurements above 200 m a.g.l., and the standard deviation of the differences is within 3 ms-1.

Figures 4 and 5 show scatter plots between the different datasets with lines of best fit and their range and frequency distributions in percentage around the Iqaluit (black) and Whitehorse (blue) sites. Figure 4 compares Rayleigh winds against the other products (ECCC-B, ERA5, radiosonde, and the limited number of coincident Ka-band radar profiles) and Fig. 5 compares Mie winds against the other products. Aeolus provides more Rayleigh measurements than Mie winds, because the Rayleigh channel measures winds under clear-sky conditions and has greater vertical extent, while the Mie channel measures winds under cloudy or high-aerosol conditions.

To measure the consistency of vertical profiles, we calculate the adjusted r-squared statistic, radj2, using 5r2=1-∑i(yi-y^i)2∑i(yi-y‾)2,6radj2=1-(1-r2)(N-1)N-p-1, where yi is the Aeolus measurements (or the other dataset shown on the y axis), y^i is the estimated HLOS wind using linear regression, y‾ is the mean of yi, N is the total number of measurements, and p is the number of profiles. The adjustment avoids overestimating the raw correlation from the scatter plots by accounting for within-profile agreement. The N and p to calculate the adjusted r-squared in Figs. 4 and 5 are shown in Table S1 in the Supplement. The radj2 and slope of the fitted line are shown in Table 1. Along with the consistency between ECCC-B and ERA5, these two datasets are also consistent with the radiosonde data, as expected, because radiosonde measurements are used in the operational ECCC and ECMWF data assimilation systems. All adjusted r-squared values in this comparison are above 0.95 for both sites, and the slopes of the fitted line are all 1±0.1.

Overall, the datasets show strong consistency. ECCC-B and ERA5 are highly mutually consistent (Table 1, with adjusted r-squared greater than 0.97) and therefore show similar consistency to Aeolus (Figs. 4a, b and 5a, b). It can be seen that Aeolus winds are in general less consistent with ECCC-B, ERA5, and radiosondes at Iqaluit than the corresponding observations at Whitehorse. Moreover, at Iqaluit, Rayleigh winds show a higher consistency than Mie winds, while the opposite is true for Whitehorse. One possible reason for this relates to the fact that the Mie channel samples winds in the lower atmosphere, where winds are harder to assimilate or measure due to topography. Since Iqaluit is situated in tundra valleys with rocky outcrops that can cause increased variability in the wind field, while Whitehorse is situated in large valleys with less wind variability due to topography, terrain effects might account for the difference in consistency. In addition, the overall range extent of the HLOS wind samples is between -25 and 25 ms-1 at Iqaluit and between -45 and 45 ms-1 at Whitehorse, and r-squared is sensitive to the range of data (note the denominator of the second term in Eq. 5). Overall, Aeolus data show good agreement, with these three datasets with adjusted r-squared greater than 0.8.

On the other hand, the adjusted r-squared between Aeolus winds and Ka-band radar at Iqaluit is only 0.53 for Rayleigh winds and 0.66 for Mie winds. As mentioned above, this might reflect a sampling bias because the vertical range of the instrument is relatively limited due to the requirement for hydrometeors to be present, so that there are therefore relatively few points to sample. In addition, at larger ranges, the radar measures winds further from the radar, and so the radar's measurement covers a larger volume. The validity of the assumption of uniform winds for the VAD calculation to be correct becomes less accurate as the range increases. However, we are comparing the VAD wind profile to a large distance along the track (87 km) as well, so this might not be the main cause. Terrain effects at Iqaluit, mentioned above, might be another possible issue: Aeolus might average out wind variability due to the topography over the resolution provided by the HLOS products. This reflects the challenge involved when comparing measurements to distinct spatial sampling and limited coincidence. Generally, the sampling for these radar measurements is highly limited, which tends to reduce the agreement compared to the other datasets. Nevertheless, the agreement on the variances between Aeolus and the Ka-band radar is at a 99 % confidence level using the F-test. This analysis highlights the importance of programmes such as CAWS continuing to provide ground-based radar measurements to ensure independent measurements of the winds for future DWL missions.

We broaden the region of analysis to the Canadian Arctic by incorporating all available Canadian Arctic radiosonde stations that provide wind profile observations. Figure 6 shows a comparison of adjusted r-squared (N and p are shown in Table S2 in the Supplement) between 2B02/06/07 Aeolus and ECCC-B, ERA5, and radiosonde measurements, coincident with the radiosonde stations shown in Fig. 1, during early FM-A, early FM-B, and mid FM-B periods. Aeolus wind profiles are less consistent with radiosonde measurements than with ECCC-B and ERA5 for both Rayleigh and Mie winds during all three periods of analysis (adjusted r-squared values in the range 0.01–0.05 and lower for the radiosonde observations). These results are in good agreement with those of Martin et al. (2021), who showed that the representativeness error is significantly larger for radiosonde observations than those for the ECMWF and ICON models.

A systematic difference between the three measurement periods is apparent. Rayleigh winds could be very sensitive to the solar background radiation (SBR) that contaminates the weak Rayleigh backscatter signal under clear-sky conditions. Random errors caused by the SBR were anticipated. Aeolus points towards the Sun-synchronous night side of its orbit to minimize the impact of SBR on the wind observations (Kanitz et al., 2019; Zhang et al., 2021). However, the impact is greater than expected, especially during summertime over the Arctic, where the Rayleigh random errors can be as high as 8 ms-1 (Zhang et al., 2019; Krisch and the Aeolus DISC, 2020; Reitebuch et al., 2020b). As a result, as will also be shown below, the consistency of Aeolus Rayleigh winds with other datasets markedly worsens during summer. We also note a slight drop in consistency of the Mie winds for the mid FM-B period, which took place in winter 2020: for instance, the adjusted r-squared and its 99 % confidence intervals, between Mie winds and ECCC-B, are 0.92 ± 0.03 during autumn 2018, 0.91 ± 0.01 during summer 2019, and 0.87 ± 0.02 during winter 2020. The cloud cover, number of observations, and estimated error from Aeolus do not seem to control this decrease. Its cause, which could be due to the Aeolus measurement or the wind retrieval, remains unclear.

For more insight into how the behaviour in the Canadian datasets we have examined extends to other Arctic regions, we now evaluate Aeolus wind measurements over the whole Arctic (measurement and profile counts provided in Table S3 in the Supplement), including over the Arctic Ocean, where wind observations are particularly sparse. We evaluate the HLOS winds in relation to the ECCC-B and ERA5 products poleward of 70∘ N. Note that we exclude the measurements over a region that partially covers Greenland, the North Atlantic Ocean, and Iceland (50∘ W to 5∘ E and 52.5 to 80∘ N) in September 2019 because Aeolus had a different range bin setting over this area for AVARTAR-I campaign purposes (Fehr et al., 2020). The time series of the estimated errors from the 2B06 (solid line) and 2B10 (dashed line) datasets and the root-mean-square difference (RMSD) between the Aeolus Rayleigh winds and ECCC-B data are shown in Fig. 7. The estimated errors and RMSD over the excluded region (blue) have a sudden jump on 9 September indicated by the vertical red dashed line, while the rest of the Arctic (black) shows a consistent decrease in estimated errors and RMSD. This decrease in this period also shows how the contribution to the error due to the solar background radiation is decreasing with the transition from summer to autumn conditions. The reprocessed data have improved estimated errors and RMSD over the excluded region; however, the jump is still visible. From 9 September to 6 October 2019, the satellite measured at a finer vertical resolution to compare to research-flight measurements. Thus, the derived winds were averaged over fewer measurements. The Rayleigh winds are particularly noisy due to the loss in optical signal along the atmospheric and internal path (Reitebuch et al., 2020b), which emphasizes the seasonal variation of the solar background noise during boreal summer and perhaps also reflects the attempt to measure finer vertical scales. Thus, the price of having a higher vertical resolution is larger errors for this specific range bin setting. For the consistency of the data quality, we thus exclude the measurements for this period and region from subsequent analysis.

By expanding the region of analysis, we obtain a larger sample, which allows us to look at the separate ascending and descending orbit phases, frequency distributions in different layers in the atmosphere, correlations along the Aeolus track in different atmospheric layers, and the geographic variation of correlations between vertical HLOS profiles. We define four atmospheric layers: the PBL (in the vertical range up to 2 km), the free troposphere (T, 2–8 km), the upper troposphere/lower stratosphere (UTLS, 8–16 km), and the stratosphere (S, altitudes greater than 16 km). Figure 8a and c show examples of stacked distributions of the Rayleigh and Mie winds for winter 2020 over the Arctic. Rayleigh winds at pressure greater than 850 hPa are ignored, as recommended by Rennie and Isaksen (2020), because they show some indications of degradation in forecasts. The Mie channel measures winds under cloudy conditions and thus has more measurements in the PBL and in the free troposphere than at higher altitudes (e.g. Fig. 8c). Furthermore, the ascending and descending Rayleigh distributions (Fig. 8b) are symmetric about zero due to the symmetric azimuth angle of the instrument with respect to the north when switching from the ascending phase to the descending phase. To avoid this artefact and to add some insight into the wind features being measured, we also compare the projected HLOS wind vector to its zonal (positive to the east) and meridional (positive to the north) components. The stacked distribution of the zonal component of the HLOS winds is shown in Fig. 8e and g for Aeolus Rayleigh and ECCC-B HLOS winds. By doing this projection, the distributions for ascending and descending measurements are brought into better agreement (Fig. 8f). We also notice that the projected zonal component of the HLOS winds can provide some information about the vertical variation of the zonal wind. For example, for Aeolus Rayleigh, the mean values of the zonal projection of the HLOS wind for the stratosphere, UTLS, and troposphere are 11.00, 4.00, and 1.00 ms-1 respectively. These mean values, as well as their standard deviations (see the legend of Fig. 8e and g), agree well with ECCC-B (and ERA5 – not shown). Aeolus-measured positive values of the zonal wind component from the stratosphere into the troposphere are consistent with the known climatological presence of westerlies in this region in polar winter. Analysing the zonal projection of the HLOS winds highlights this feature.

We compare the distributions of the differences between the Aeolus wind measurement data and the ECCC-B and ERA5 data during autumn 2018, summer 2019, and winter 2020 over the Arctic, as summarized in Fig. 9, which shows the bias and standard deviations of the differences for the HLOS winds and for their zonal and meridional projections. To highlight the variations of the means, the standard deviations were divided by a factor of 10. The measurements are separated into Rayleigh (red dots) and Mie (black dots) winds. They are further separated into ascending (indicated with upright triangles) and descending (inverted triangles) measurements. The results, with the biases being smaller than 0.7 ms-1, are consistent with the ECCC bias-correction method. The means and standard deviations of the differences in the ascending and descending measurements do not show a significant difference. The discrepancies in the meridional projections of the HLOS winds are smaller because Aeolus picks up mostly the zonal component of the winds due to the direction of the LOS.

Although Fig. 9 shows an overall agreement between Aeolus, ECCC-B, and ERA5, more analysis is required to bring out the differences between the datasets. One way to do so is to separately investigate the consistency between Aeolus and ECCC-B or ERA5 HLOS winds in the PBL, troposphere, UTLS, and stratosphere. Figure 10 shows normalized Taylor diagrams (Taylor, 2001), with ECCC-B as a reference, for Aeolus Rayleigh and Mie measurements over the Arctic during the three seasons of analysis. The angle indicates the correlation between Aeolus measurements and ECCC-B. The distance to the origin represents the standard deviation, and the distance to the star (reference point (1,0)) represents the RMSD; both statistics are normalized by the standard deviation of reference data. The 1.0 normalized standard deviation is highlighted; data that fall outside the dashed quarter circle are noisier than the reference data. Figure 9 shows that Aeolus data consistently have greater standard deviations than ECCC-B during all three periods and for both Rayleigh and Mie winds: its normalized standard deviations are typically within 1.05 to 1.40 or greater. This might imply that Aeolus provides noisier data, that the ECCC-B is missing some extreme values in its wind-component distribution, or both. However, the RMSD is generally within 1 normalized standard deviation, and correlations are normally greater than 0.8 for Rayleigh winds above the troposphere and for Mie winds below the lower stratosphere, with an exception for stratospheric consistency for Rayleigh winds during the summer. During the boreal summer period, the data in the stratosphere seem to agree less with the ECCC-B data, reflecting reduced sampling, solar background noise that is most effective during summer, changes in the atmospheric environment in summer, thermal changes in the telescope, and other possible errors (Reitebuch et al., 2020b).

Generally, the Rayleigh-clear channel provides consistent data with the ECCC-B through the troposphere (T in Fig. 10), the UTLS (U), and the stratosphere (S), while the Mie-cloudy channel provides consistency from the PBL (B) to the lower stratosphere. This reflects the vertical sampling and instrument characteristics and reveals effective complementarity of the instrument and retrieval design. For this reason, in the next paragraph, where we investigate the spatial distribution of the consistency in the lower- and upper-atmospheric regions, we exclude the Rayleigh winds in the PBL and the Mie winds in the stratosphere.

Figure 11 shows the RMSD between Aeolus and ECCC-B for Rayleigh tropospheric (T) and Mie PBL + tropospheric (B + T) profiles, and Fig. 12 shows the same but for Rayleigh UTLS + stratosphere (U + S) and Mie UTLS (U) measurements. Since the estimated errors and RMSD were consistently decreasing in September 2019 over the Arctic except for the region with a different range bin setting (Fig. 7), to avoid misinterpretation of the results on the maps, we exclude the data during this period over the entire Arctic. Since the measurement density differs depending on the latitude, the RMSDs of the profiles are calculated over nearly equal surface area, using the Equal-Area Scalable Earth (EASE) grids (Brodzik et al., 2012). Each grid cell is around 104 km2, which is approximately the square of the along-path resolution of Aeolus Rayleigh winds. The first two and last columns represent the distributions using the near-real-time 2B02/06/07 datasets; the third column shows the distributions using the reprocessed 2B10 data during the early FM-B period.

The distributions of the RMSD in the lower atmosphere are relatively homogeneous across oceanic, ice-covered, and continental regions. This suggests that the overall good agreement seen for the in situ Iqaluit and Whitehorse data as well as the northern Canadian region in the radiosonde network extends from land to the ocean regions without obvious systematic differences in consistency. The agreement between Aeolus and ECCC-B for the Mie winds is better than for the Rayleigh winds for all three periods of analysis. The RMSD for the Mie winds largely lies between 2 and 4 ms-1, and for Rayleigh winds the RMSD is generally greater than 4 ms-1. This was anticipated because the Rayleigh winds are noisier, for reasons alluded to above. The RMSD is systematically greater in the upper atmosphere than in the lower atmosphere, as shown in Figs. 11 and 12, and the differences could be anticipated from the Aeolus estimated errors from the L2B product (as shown in Figs. S1 and S2 in the Supplement).

In the upper atmosphere, the RMSD is significantly larger for the Rayleigh winds during the FM-B period, especially over Greenland, where the RMSD is greater than 6.0 ms-1. This might be caused by the reflection from the Greenland ice sheet, which leads to enhanced errors. The estimated errors also show a similar pattern: the estimated error over Greenland in summer 2019 is around 6.0 ms-1 (Fig. S2b), the reprocessed product reduced the estimated error to around 5.0 ms-1 (Fig. S2c), and it is around 4.0 ms-1 during winter 2020 (Fig. S2d) due to the lack of solar background noise in the boreal winter. This pattern is however not seen during the FM-A period. Whatever the source of these changes (e.g. summertime solar background noise amplification or calibration errors coinciding with the start of the laser-B stream), it is noteworthy that the estimated error product contains potentially useful information for validation purposes. Because such information is useful for error characterization in NWP, as we will discuss below, a more detailed investigation into the estimated error product, including its seasonal, geographic, and flow dependence, is warranted.

Note that the first reprocessed data, 2B10, only overlap with one of the three periods of study: August to September 2019. The estimated observational errors have decreased compared to the 2B06 data (Figs. S1 and S2) since the bias due to the M1 mirror temperature dependence is updated on a daily basis and the dark current signals have been removed using improved quality control. However, we do not see the same improvement in the O–B statistics between the 2B06 and 2B10 products over the Arctic region, possibly because the reprocessed product had improved the precision (error characterization) of the measurements while not leading to a change in the overall agreement between the products, suggesting that accuracy is not changed.