Validation of Aeolus winds using radiosonde observations and NWP model equivalents

In August 2018, the first Doppler Wind Lidar, developed by the European Space Agency (ESA), was launched on-board the Aeolus satellite into space. Providing atmospheric wind profiles on a global basis, the Earth Explorer mission is expected to demonstrate improvements in the quality of numerical weather prediction (NWP). For the use of Aeolus observations in NWP data assimilation, a detailed characterization of the quality and the minimization of systematic errors is crucial. This study performs a statistical validation of Aeolus observations, using collocated radiosonde measurements and 5 NWP forecast equivalents from two different global models, the ICOsahedral Nonhydrostatic model (ICON) of Deutscher Wetterdienst (DWD) and the European Centre for Medium-Range Weather Forecast (ECMWF) Integrated Forecast System (IFS) model, as reference data. For the time period from the satellite’s launch to the end of December 2019, comparisons for the northern hemisphere (23.5− 65◦N) show strong variations of the Aeolus wind bias and differences between the ascending and descending orbit phase. The mean absolute bias for the selected validation area is found to be in the range of 1.8 2.3 m 10 s−1 (Rayleigh) and 1.3 1.9 m s−1 (Mie), showing good agreement between the three independent reference data sets. Due to the greater representativeness errors associated with the comparisons using radiosonde observations, the random differences are larger for the validation with radiosondes compared to the model equivalent statistics. To achieve an estimate for the Aeolus instrumental error, the representativeness errors for the comparisons are determined, besides the estimation of the model and radiosonde observational error. The resulting Aeolus error estimates are in the range of 4.1 4.4 m s−1 (Rayleigh) and 15 1.9 3.0 m s−1 (Mie). Investigations of the Rayleigh wind bias on a global scale show that in addition to the satellite flight direction and seasonal differences, the systematic differences vary with latitude. A latitude based bias correction approach is able to reduce the bias, but a residual bias of 0.4 0.6 m s−1 with a temporal trend remains. Taking additional longitudinal differences into account, the bias can be reduced further by almost 50 %. Longitudinal variations are suggested to be linked to land-sea distribution and tropical convection that influences the thermal emission of the earth. Since 20 April 2020 a telescope 20 temperature-based bias correction scheme has been applied operationally in the L2B processor, developed by the Aeolus Data Innovation and Science Cluster (DISC).


Introduction
Aeolus is a European Space Agency (ESA) Earth Explorer mission, launched on 22 August 2018 as part of the Living Planet Programme. With an estimated lifetime of three years, it is expected to pave the way for future operational meteorological 25 satellites dedicated to observing the atmospheric wind fields (ESA, 2008). Aeolus is a polar orbiting satellite, flying in a sunsynchronous dawn-dusk orbit at about 320 km altitude. Within about seven days, the satellite covers nearly the whole globe.
Aeolus carries only one large instrument, a Doppler wind lidar called ALADIN (Atmospheric LAser Doppler INstrument) which is the first European lidar and the first ever Doppler Wind Lidar (DWL) in space (Stoffelen et al., 2005;Reitebuch, 2012;ESA, 2008). ALADIN provides profiles of the line-of-sight (LOS) wind component perpendicular to the satellite velocity at an 30 angle of 35 • off-nadir from the ground up to 30 km.
The Aeolus mission primarily aims to demonstrate improvements in atmospheric wind analyses for the benefit of numerical weather prediction (NWP) and climate studies (Stoffelen et al., 2005;Rennie and Isaksen, 2019). Despite the advancement of the Global Observing System (GOS), there are still major deficiencies, the lack of accuracy are significant limitations of currently used wind observation methods (Källén, 2018). Accurate vertical profiles of the wind field from radiosondes, wind 35 profilers, and commercial aircraft ascents and descents are mainly concentrated over continents in the northern hemisphere, whereas only a few profiles are available over the oceans and on most parts of the southern hemisphere. Atmospheric motion vectors derived from tracking cloud and water vapor structures in consecutive satellite images provide single-level winds with nearly global coverage but exhibit significant systematic and correlated errors due to uncertainties of their height assignment (e.g. Folger and Weissmann, 2014;Bormann et al., 2003). The vast majority of global observations consist of satellite radiances, 40 mainly providing information on the atmospheric mass field (temperature, humidity, other trace gases, and hydrometeors).
Wind information can only be retrieved indirectly from these observations, which is a particularly strong restriction in the tropics in the absence of geostrophic balance (Stoffelen et al., 2005). The actively sensed globally distributed lidar LOS winds are therefore filling a major gap of the GOS, especially in the upper troposphere and the lower stratosphere, in the tropics and over the oceans (Baker et al., 2014;ESA, 2008). It has been shown that improvements are to be expected for short range 45 forecasts of severe weather situations, the analysis of tropical dynamics, and for a better definition of smaller scale circulation systems in midlatitudes (e.g. Marseille et al., 2008;Tan and Andersson, 2005;Weissmann and Cardinali, 2007;Weissmann et al., 2012;Žagar, 2004). A crucial prerequisite for the use of meteorological observations in NWP data assimilation systems is a good knowledge of their statistical errors and the minimization of systematic observation errors. For this purpose, uncertainty assessment and validation through extensive comparisons with reference data is an essential requirement to assimilate these 50 novel observations in NWP models and fully exploit the provided wind information.
The Aeolus direct detection wind lidar (ALADIN) is operating in the ultraviolet spectral region (354.8 nm). The laser emits pulses of about 60 mJ at a frequency of 50.5 Hz. A Cassegrain telescope with a diameter of 1.5 m collects the backscattered signal, which Doppler shift is analyzed by a dual channel receiver to measure backscattered signals from both, molecules (Rayleigh channel) and particles (Mie channel) (ESA, 2008;Reitebuch, 2012). This complementarity of the two channels 55 allows for broad vertical and horizontal data coverage in the troposphere. In preparation of the Aeolus mission, a prototype of the satellite instrument, the ALADIN Airborne Demonstrator (A2D), was deployed to test the wind measurement principles under real atmospheric conditions in several measurement campaigns, and to provide information on quality control algorithms (Lux et al., 2018). Two airborne validation campaigns with operation base at DLR (Deutsches Zentrum für Luft-und Raumfahrt e.V.) Oberpfaffenhofen were already performed within the first ten months after the satellite's launch. Deploying the A2D and 60 a 2-µm DWL as reference, wind data for the first experimental comparisons with the Aeolus wind product and model wind data from the European Centre for Medium-Range Weather Forecasts (ECMWF) were provided. Detailed information and results have been published in Lux et al. (2020) and Witschas et al. (2020). Further, Aeolus wind observations are compared to the direct-detection Doppler lidar LIOvent at the Observatoire de Haute-Provence for a time period at the beginning of 2019 (Khaykin et al., 2020) and to wind profiles obtained from radiosonde launches on-board the German RV Polarstern in 65 autumn 2018 across the Atlantic Ocean (Baars et al., 2020) . Airborne Doppler lidars have been used in several case studies of mesoscale phenomena, such as the French mistral (Drobinski et al. 2005), Alpine foehn (Reitebuch et al. 2003), the sea breeze in southern France (Bastin et al. 2005), or the Alpine mountain-plain circulation (Weissmann et al. 2005). As part of the German initiative EVAA (Experimental Validation and Assimilation of Aeolus observations), this paper presents the evaluation of Aeolus winds using operational collocated radiosonde data from the GOS as reference. Besides, observation The text is structured as follows. First, an overview and description of the data sets used for the evaluation of the Aeolus winds is provided. Collocation criteria are specified and the statistical methods for the comparison are described. Section 3 75 presents a time series of the validation, focusing on the temporal evolution of systematic and random differences between the Aeolus observations and the reference data sets. The derivation of error estimates for the Aeolus instrumental error includes the determination of representativeness errors which is based on analysis data from the regional model COSMO (Consortium for Small-scale MOdeling) of DWD and high resolution ICON Large Eddy Model (LEM) simulations. In Section 4, the Aeolus Rayleigh channel bias is investigated in more detail and two bias correction approaches are evaluated. Finally, the results are 80 discussed and summarized.

Data and method
The Aeolus Level 2B (L2B) product is evaluated using collocated radiosonde observations of the GOS and short-term model forecast equivalents (first guess departure statistics) of the global model ICON of DWD and the ECMWF model as reference.

85
The Aeolus L2B product contains the Horizontal LOS wind component (HLOS) observations suitable for NWP data assimilation (Rennie et al., 2020). The majority of wind data are provided by the Rayleigh channel. In clear conditions, the Rayleigh wind coverage is from the surface up to 30 km. The Mie signals are strong within optically thin clouds, on top of optically thick clouds, and cover the atmospheric boundary layer as well as aerosol layers for clear sky conditions. Each Aeolus measurement is an accumulation of 20 laser pulses which corresponds to a horizontal resolution of about 2.9 km. To achieve a sufficient 90 signal to noise ratio to comply with the stringent wind accuracy requirements, observations are processed by averaging up to 30 individual measurements. The resulting HLOS wind observation therefore represents a horizontal average over 86.4 km.
For the Mie channel, the horizontal integration length of the wind measurements was decreased to approximately 10 km after 5 March 2019, taking benefit of the higher signal to noise ratio of cloud returns (Šavli et al., 2019). In addition to HLOS observations, the Aeolus L2B processor developed by ECMWF and the Royal Netherlands Meteorological Institute (KNMI) 95 provides an observation instrument noise estimate. Furthermore, to reduce systematic errors, corrections for the temperature and pressure dependence of the Rayleigh winds are performed using a priori information from the ECMWF model interpolated along the Aeolus track (Dabas et al., 2008). The measurements within an observation are classified into an observation type, clear or cloudy, using measurement-scale (2.9 km) optical properties, such as scattering ratio. Wind retrievals are performed for both channels resulting in four observation products (Rayleigh-clear, Rayleigh-cloudy, Mie-clear and Mie-cloudy). The 100 vertical resolution varies from 0.25 km near surface to 2 km in the highest bins, with a total of 24 bins. The processing at ECMWF is performed in near-real-time, thus measurements are delivered within three hours. More detailed information about the L2B processor retrieval algorithm can be found in Rennie et al. (2020). As Aeolus is a novel mission, the processing algorithms have been evolving since launch. Different processor baselines (in this study 2B02 -2B07) and various updates led to different observation quality in different time periods. A consistent reprocessed data set with unique processor settings is not 105 available yet. Furthermore, the instrument performance varied over the time period assessed in this study, which includes the missions Commissioning Phase (CP) from launch until the end of January 2019, the late Flight Model A (FM-A) laser period until mid of June 2019, and the FM-B laser period until the end of December 2019. Information about the actual performance of the Aeolus wind lidar and a discussion of the systematic and random error sources can be found in  and Rennie and Isaksen (2020). For the validation, the following quality control criteria are applied. Only valid Rayleigh clear 110 and Mie cloudy winds (from now on referred to as Rayleigh and Mie) between 800 and 80 hPa are used. A distinction is made between the ascending orbital pass, when the satellite moves north, and the descending orbital pass when the satellite moves south. Based on a compromise between the quality of the data set and the number of observations that pass the quality control, Rayleigh winds with an estimated error greater than 6 m s −1 and Mie winds with an estimated error greater than 4 m s −1 are excluded. Thus, on average over the validation period about 70 % of the Rayleigh and 76 % of the Mie winds are available 115 for the analysis. On June 14, 2019 a correction scheme for dark current signal anomalies of single pixels (hot pixels) on the Accumulation-Charge-Coupled Devices (ACCDs) of the Aeolus detectors has been implemented into the Aeolus operational processor chain . All measurements before June 14, 2019 affected by hot pixels are excluded from the validation statistic.

Radiosonde data and collocation criteria 120
Radiosonde observations generally provide very accurate information on the true wind conditions. Given that radiosonde wind data are direct in situ measurements, the inherent errors (e.g. instrument errors) are small compared to errors of satellite-based instruments. That makes them well suited to serve as reference data set for the true atmospheric state for the validation of Aeolus HLOS winds. Furthermore, the observation errors can be assumed to be uncorrelated between different radiosondes.
where φ is the L2B azimuth angle, which is defined clockwise from north of the horizontal projection of the target to satellite pointing vector. Since radiosonde observations are rare in the southern hemisphere and polar regions, the analysis concentrates 140 on the midlatitudes of the northern hemisphere (23.5 − 65 • N). To achieve a sufficiently large data set, statistics for one day are based on a running mean over seven days. ods of seven days before they are averaged for the whole globe, to reduce the influence of horizontal and temporal fluctuations of systematic errors on the random errors.

Representativeness errors for the Aeolus wind validation
The knowledge of representativeness errors is a key to determine the Aeolus winds instrumental error. Firstly, representativeness errors arise due to different measurement geometries of the compared data sets. Whereby the Aeolus HLOS wind 165 observations correspond to line measurements, the NWP models are treating the Aeolus HLOS winds as point measurements.
Also, the radiosonde observations can be regarded as point measurements. For the estimation of the representativeness error for the comparison of radiosonde and Aeolus data, three further error sources need to be taken into account: The spatial and temporal difference resulting from the collocation criteria; the spatial and temporal difference resulting from the displacement during the radiosonde ascents when radiosonde data from alphanumeric reports are assimilated (13 % of the radiosonde data); 170 and the temporal offset value for the grouping time interval when accounting for balloon drift in BUFR data (87 % of the radiosonde data).
The different error components are evaluated using analysis data of the regional COSMO-DE model of five seven-day peri-

Statistical metrics
The following outlines the applied statistical metrics. Using the forecast of NWP models as reference, the bias estimate is described by the mean first guess departure: where i represents the time step and N is the number of compared data points. y is the Aeolus HLOS wind observation, x b is 185 the state vector of the short-term model forecast (background) and H(.) the observation operator. Given that the model bias for long validation periods and large scales is usually small in comparison to that of Aeolus observations, the mean difference between the Aeolus observations and the reference data can be referred to as bias. In certain conditions, such as in jet stream regions, the tropical upper troposphere and the stratosphere, however, Aeolus HLOS bias estimates based on NWP monitoring statistics should be treated with caution (Rennie, 2016).

190
The bias using the radiosonde measurements as reference is calculated according to For quantifying random deviations, the standard deviation as well as the scaled median absolute deviation (MAD) is determined for the three reference data sets. The MAD is a very robust measure for the variability of the Aeolus HLOS winds, being more resilient to single outliers compared to the standard deviation. In case of a normally distributed data set, the MAD value multiplied by 1.4826 (scaled MAD) is identical to the standard deviation (Ruppert and Matteson, 2015).
3 Validation results -time series characteristics and error estimation of Aeolus HLOS wind comparisons 200

Systematic and random differences
For the time period from the first available L2B data after the satellite's launch up to January 2020, systematic and random differences between the Aeolus HLOS winds, radiosondes and model fields are calculated. Figure  increase occurs in mid-December. This is caused by a manual L2B processor bias correction of + 4 m s −1 in the Rayleigh wind product to compensate for a global average bias drift. The Mie wind mean differences are only slightly increasing. All three independent reference data show very good agreement for the bias estimation, raising confidence that the results are 245 not determined by model biases. Besides the temporal changes in Aeolus Rayleigh and Mie wind quality, the discrepancies between the ascending and descending orbit, mainly for the Rayleigh channel are a challenging issue for using these data in NWP models. Significant differences occur especially in late summer and autumn. Assessing the mean absolute values, the bias is larger for the descending than for the ascending orbit for both channels. For a more detailed analysis of the Rayleigh bias, see Section 4.

250
The Rayleigh wind random differences calculated based on model O-B statistics vary between 3 and 6 m s −1 within the

Estimation of the Aeolus HLOS wind error
The total variance of the difference between radiosonde observations and Aeolus HLOS winds σ 2 val (squared scaled MAD) is the sum of the variance resulting from the Aeolus wind instrumental error σ 2 Aeolus , the variance resulting from the radiosondes wind observational error σ 2 RS , and the variance caused by the representativeness error σ 2 r_RS (Weissmann et al., 2005) (see

Equation 6a
). For the comparison with model equivalents, the model representativeness error σ 2 r_model is used and σ 2 RS is replaced by the model error σ 2 b (see Equation 6b). For the estimation of the radiosonde representativeness error σ r_RS , error sources caused by spatial and temporal displacements need to be considered, additionally to the different measurement geometries of the radiosonde and the Aeolus observations. Therefore, it is necessary to make a distinction between radiosondes for which the drift is assimilated (87 %), and those 285 reports which only contain the launch position and time (13 %). For both cases, the temporal and the spatial part of the representativeness error, resulting from the collocation criteria, has to be considered. The error due to the spatial displacement is assessed by determining the differences between a point and a line measurement as weighted mean over distances up to 120 km in east-west and north-south direction, and calculating the weighted average over altitude. To account for the temporal displacement, a time offset value is estimated by assessing the representativeness error of the appropriate spatial displacement.

Model error σ b and radiosonde wind observational error σ RS
The ECMWF model error is derived from the ensemble data assimilation first guess error, stored in the ODB. It provides a good measure for spatial and temporal variation of the background error. Table 2  The radiosonde observational error σ RS is assumed to be 0.7 m s −1 , according to the estimated GCOS (Global Climate Observing System) Reference Upper-Air Network (GRUAN) measurement uncertainty (Dirksen et al., 2014) .

Aeolus wind instrumental error σ Aeolus
The Aeolus wind instrumental error is calculated using Equation 6a and Equation 6b. Table 2 shows the values of σ Aeolus for 320 the validation with radiosonde observations and ECMWF model equivalents for the latitudinal band between 23.5 and 65 • N for the Rayleigh and Mie winds, separated for the ascending and descending orbit phase. Additionally, σ Aeolus is derived for the global statistic using the ECMWF O-B values.

355
Representative for winter and spring, Figure 3 (a) shows that the bias is quite constant with latitude in that period. Small differences between the orbital passes occur in the southern hemisphere and in the subtropical region of the northern hemisphere.
From 40 • N up to the north pole, almost no deviation between ascending and descending orbit is visible.
In August 2019 (Figure 3 (b)), the bias varies with latitude with an amplitude of 4 -5 m s −1 . As seen in Section 3 for the summer and autumn season, large differences between the orbit phases exist, in particular outside of the tropics. Around the 360 equator, the sign of the bias is positive for the ascending and descending orbit. Between the subtropical region and the poles, the descending orbit bias is still positive, whereas the bias of the ascending orbit has a negative sign.
The results suggest that the satellites orbit phase and latitude position as well as the season seem to influence the Aeolus Rayleigh wind bias. As the formulation of most data assimilation schemes assumes unbiased incoming observations, the correction of systematic differences is crucial. Thus, it is first tested, if a bias correction approach as a function of latitude based 365 on first guess departures of the preceding week, separately for ascending and descending orbit, can remove the systematic differences for the validation period (see Section 4.1.1).

Rayleigh wind bias correction approach as function of latitude
Based on the previous results, a bias correction approach is evaluated and tested with the ECMWF IFS and the ICON model monitoring data sets. For latitude bins of 10 • , the first guess departures from the previous seven days are averaged using the 370 following weights: with i=0 being the current day. The resulting correction values are subtracted from the first guess departure of the considered day and the residuals are averaged for each month of the validation period (Figure 4). Considering the effect of the orbit phase differences, this is done separately for the ascending and descending satellite pass. To estimate if the model bias matters three 375 different configurations are tested, which differ regarding the correction values: the bias correction values are based on the same model (dark filled markers); the bias correction value is calculated with the other NWP model (unfilled markers); the bias correction value is an average value of the two NWP models (light filled markers).
After applying the bias correction, a temporal variation as seen in Section 3 for the systematic differences is still apparent in the residuals. At the beginning of the Aeolus mission, the correction is quite efficient. In spring 2019, when the latitude dependence 380 is comparably weak and the bias comparably high, a residual up to over 1 m s −1 remains. After the processor update in May 2019, when the Rayleigh ascending wind bias tends to be negative, also the residual bias exhibits a negative sign. Differences between the two models regarding the sign of the remaining bias are visible in September 2018 for the ascending orbit and in December 2019. In total, the correction is able to clearly decrease the systematic differences, but there is a remaining bias, in particular in phases with large temporal changes of the bias. The seasonal variation of the bias and the influence of the   shows worse results when applying information of the ECMWF IFS model to correct for the latitude dependent error. Overall, the bias correction approaches show a statistically significant reduction in bias. However, no significant differences between 395 the individual methods were found (following a Student's t-distribution), which again indicates that model biases do not have a dominant effect on the bias assessment.  Altogether, these results show that a temporally varying latitude dependent bias is present for the L2B Rayleigh wind product.
Results from the evaluation with the two independent NWP models and in situ observations are overall in good agreement.
A latitude dependent bias correction successfully reduces the bias, but on average, a bias of 0.37 -0.59 m s −1 remains. The   most, also a significant longitude dependence is apparent. The land sea fluctuations for the ascending orbital pass on the north-415 ern hemisphere and in the tropical region are more pronounced. For the descending orbit, variability is mainly present in the southern hemisphere and it is not clear whether this is linked to the land sea distribution. The positive bias band in the ITCZ region is still present for both orbits.

Rayleigh wind bias correction approach as function of latitude and longitude
For the ECMWF model, a two-dimensional bias correction approach is tested using the previous seven days of Aeolus HLOS O-B statistics as a function of latitude and longitude averaged and weighted (Equation 7). Bin sizes are chosen to be 10 • for both, latitude and longitude. To also consider the seasonal variation, Figure 7 displays the residuals (rose cross markers) 430 averaged for each month for the whole validation period for the ascending and descending orbit. To get an impression of how strong the longitudinal bias variation is, the results are compared to the one-dimensional latitude dependent correction approach from Section 4.1.1. The mean absolute remaining bias for both correction formulations is provided in Table 4. Altogether, the residual has been decreased by almost 50 % when considering the longitude dependence for both satellite orbit passes. Main improvements occur for the bias correction in late winter and early spring 2019, where a one-dimensional correction approach 435 is not that effective. Right after the mission's start, in May 2019 and at the end of the year the remaining bias is increased when taking the longitudinal dimension into account. In these months, the one dimensional latitude dependent correction approach almost removes the systematic differences already.
A discussion on possible reasons for the systematic bias variations and a summary of the findings in this study are presented in the final Section 5. 440 Figure 7. Residual after a latitude dependent bias correction (magenta diamond marker) and a two dimensional latitude-longitude dependent bias correction (rose cross marker) averaged over one month, using the ECMWF model equivalents. On the left for Rayleigh ascending (a) and on the right for descending (b) orbit phase.

Summary and discussion
This study provides an overview of validation activities to determine the Aeolus HLOS wind errors and to understand the biases by investigating possible dependencies. To ensure meaningful validation statistics, collocated radiosondes and two different global NWP models, the ECMWF IFS and the ICON model of DWD, are used as reference data.
Overall, the determined mean wind differences of the comparisons with all three reference data sets show good concordance.

445
This confirms that the detected bias is due to Aeolus L2B systematic wind errors and not the reference data set. A time series demonstrates that the Aeolus wind systematic differences vary considerably during the time period from the satellite's launch until the end of December 2019 (Section 3.1). Further, there are differences in bias between the ascending and descending orbit phase, which mainly occur for the Rayleigh channel in late summer and autumn. Whereas the Rayleigh descending phase winds are positively biased in these months, the ascending phase shows negative bias values. The Mie winds are less biased 450 in total, but more fluctuating. The mean absolute bias is found to be approximately 1.8 -2.3 m s −1 for the Rayleigh winds and 1.3 -1.9 m s −1 for the Mie winds. These values are beyond the mission requirements of Aeolus, which state that the bias should be smaller than 0.7 m s −1 (ESA, 2016). However, it is demonstrated that the bias can be reduced to values lower than the mission requirement through calibration with observations and model fields of the preceding week.
The random differences of the Rayleigh winds show temporal changes, that are mainly related to changes in the laser output The second part (Section 4) of the results of this study further investigates the Rayleigh wind bias and its dependencies. Besides the satellite's flight direction and seasonal differences, also latitude and longitude influence the systematic differences. Again, 470 the good agreement between the different validation data sets raises confidence that the results are not influenced by issues of the reference data sets. The latitude bias dependence and differences between the orbit phases mainly occur in late summer and autumn in the subtropics and temperate climate zone. A one-dimensional latitude dependent correction approach, based on the previous seven days, is able to reduce the bias, but still, a temporal trend of remaining bias values of 0.37 -0.59 m s −1 occur. It turned out that additionally, a longitude dependent bias component is present that should be taken into account. When 475 the satellite moves north, longitudinal variations are especially found in the tropics and between 20 and 60 • N, while for the descending orbit phase systematic differences mainly occur between 20 and 60 • S. These variations suggest correlations with land-sea distribution and tropical convection. A latitude-longitude correction approach using the ECMWF model equivalents is able to reduce the systematic error to 0.23 -0.25 m s −1 . As the bias correction approach is essentially a temporal and spatial smoothing, it is suggested that fast changes in systematic errors are one source of the bias residuals.

480
At ECMWF, as part of the Aeolus Data Innovation and Science Cluster (DISC), the dominant source of the Rayleigh wind bias issues have been explained. It was found that the bias is correlated with the temperature gradients across the ALADIN primary mirror M1 of the telescope (Rennie and Isaksen, 2020). The M1 mirror temperature variation in turn is related to varying short and long wave radiation of the top of the atmosphere and the mirror's on-board thermal control in response to this, which explains the seasonal differences and the connection to features like convection and variations between land and sea. Since 20

485
April 2020 a M1 bias correction scheme has been applied operationally in the L2B processor, using a multiple linear regression method of all M1 telescope thermistors developed by the Aeolus DISC (Rennie and Isaksen, 2020). A re-processed data set including a M1 bias correction will be available in near future. This data set should decrease the Aeolus instrumental error estimate and differences between the model and radiosonde comparisons.
Data availability. Since May 2020, Aeolus data is publicly available at the ESA Aeolus Online Dissemination System.

490
Author contributions. AM performed the data analysis and prepared the main part of the publication. MW supervised the work. MW, OR, MR and AG contributed to the development of methods and analysis of the data. OR and AG communicated important information on the Aeolus data quality and processing. MR provided knowledge about the ECMWF feedback files and the Aeolus wind bias. AC provided ideas for the bias correction approach and information about the DWD monitoring data. All co-authors engaged in discussions and contributed to the interpretation of the results.

495
Competing interests. The authors declare that they have no conflict of interest.
Disclaimer. The presented work includes preliminary data (not fully calibrated/validated and not yet publicly released) of the Aeolus mission that is part of the European Space Agency (ESA) Earth Explorer Programme. This includes wind products from before the public data release in May 2020 and/or aerosol and cloud products, which have not yet been publicly released. The preliminary Aeolus wind products will be reprocessed during 2020 and 2021, which will include in particular a significant L2B product wind bias reduction and improved L2A 500 radiometric calibration. Aerosol and cloud products will become publicly available by spring 2021. The processor development, improvement and product reprocessing preparation are performed by the Aeolus DISC (Data, Innovation and Science Cluster), which involves DLR, DoRIT, ECMWF, KNMI, CNRS, ST, ABB and Serco, in close cooperation with the Aeolus PDGS (Payload Data Ground Segment). The analysis has been performed in the frame of the Aeolus Scientific Calibration and Validation Team (ACVT).