The paper describes a software package to retrieve wind profile in the atmosphere from Doppler wind lidar. The paper is of good quality and describes well what the basic concept of wind retrieval from Doppler wind lidar are, and how this software implements them. The software uses a clearly defined data structure and is of modular design which allows to flexible set up retrieval chains. The presented examples contain interesting approaches to cope with noisy data, and to either identify reliable data points or reject them. The description is for the most parts clearly written and well understandable. 

Nevertheless I have some comments:

Sampling in height bins:

The retrieval takes all Doppler speeds measured within one height bin, If I understood it right they are used as if all of them were measured at the same height. This neglects that the wind speeds especially in the boundary layer show a strong non-linear height dependence. Basic assumption of the wind retrieval is that the radial Doppler wind speeds measured in the different tilted beams are all result of the same wind vector. By bringing together different heights in one retrieval this assumption is violated. In my opinion it would be necessary to first interpolate the radial velocities to one single height within every height bin. The authors should discuss this in the text.

Combination of different scans patterns:

The common way for wind profile retrieval is to perform a certain scan pattern which should evenly cover the measuring volume (see Paeschke et al 2015, mentioned in the text). Different scan patterns can be used to address different questions (fast scans like DBS with only few beams to achieve high temporal resolution, PPI (or VAD) scans with many beams to reduce the uncertainty within one scan, RHI scans to eventually resolve spatial variability at least in one vertical plane, etc.). Several scans of the same kind can be combined by averaging. 

Here is a different strategy proposed which combines all in a certain time interval available scans (hence the term 'heterogeneous measurements'). This introduces asymmetry in the spatial coverage and may lead to larger errors as is shortly  mentioned in section 4.1 when describing the standard module chain and the removal of data from low elevation beams at large distances (line 375). I recommend to emphasise stronger this strategy but also discuss the effects and dangers of an uneven coverage of the measuring volume in section 2.

Extrapolation in regions with low CNR:

An important part of the advanced retrieval is the attempt to use even data at low CNR. To achieve this a first guess wind profile is derived and inter- and extrapolated to regions with low CNR. The kind of this extrapolation is not documented. The resulting fields are used in an iterative process as 'confidence background'. Calculated wind speeds deviating not too far from this confidence background are accepted even for low CNR values. I am wondering whether this extrapolation is robust, especially in upper heights, where CNR is low over large regions. The systematically large differences in the upper part of the profiles in fig's 13, 14 and 15 might be result of this extrapolation. The authors should discuss this in the text. It would be nice to have an analysis how the retrieval performs for low CNR and whether the difference to the radiosonde data for these low CNR regions is acceptable.

Description of basic scan patterns:

As the paper also addresses researchers not familiar with all the details of doppler Lidar retrieval it would be beneficial to give a compact, and clear overview of possible scan patterns, sources of errors etc. The word scanning is used in an ambiguous way, the step-stare scan mode is mentioned but continuous scan mode not.

To retrieve the 3D wind vector it is necessary to tilt several beams in different directions - this is called a scan. Common scan patterns are PPI (or VAD), RHI, etc. Scans can be performed in two different modes: step-stare and continuous. There is a trade-off between these modes (varying orientation, dead time during movement, time for one scan, etc.). 

There are two main sources of error: the doppler retrieval has always some uncertainty, which increase with decreasing CNR (or SNR) (hence the CNR thresholding as filter strategy) and the violation of the homogenous flow assumption due to the separation of the tilted beams by e.g. turbulence. The along beam doppler uncertainty maps, depending on the elevation angle, to the retrieved wind components. With increasing elevation its contribution to the error of the horizontal wind component increases. With decreasing elevation the separation between the beams becomes larger and the error due to turbulence becomes larger. 

Range of biases

Figures 13, 14 and 15 show the bias between wind speed profiles retrieved with this method and radiosondes. While it is nice that on average the bias is close to zero the percentiles show that still 10% of the wind speeds deviate by more than about 4m/s from the radiosonde data. This could be compared with the results of Rahlves et al. (2022) and Robey and Lundquist(2022) and discussed.

- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - 

Detailed comments:

Line 64: period is missing

line 94: ambiguous use of the word 'scanning' , especially with the list of scans to follow. Are here meant general measurements in different directions or especially measurements during continuous movement? An explanation and discussion of step-stare and continuous mode is missing.

line 97: alternative name for PPI is VAD from 'Velocity Azimuth Display' (doppler velocities are 'displayed' as a function of azimuth. Wind components are retrieved by fitting a sine curve, |v| is proportional to the amplitude etc ...) already cited in Browning and Wexler (1968), stems from radar. Both terms are not perfect: PPI in radar technique means a scan at very low elevation and its presentation on a screen, VAD can also refer to the display of wind vectors as a function of height and time sometimes used in aviation.

line 99: With one RHI at one azimuth it is not possible to retrieve a wind vector, it should be mentioned that this is only possible with a combination of RHIs in different azimuth directions.

line 100: "arrangements of fixed-direction stares, e.g. in step-and-stare"

This gives the impression that the above described scan patterns are not done in step-stare mode but instead in a different way (which is never named in the paper). I would recommend to mention first the different modes (step stare and continuous), discuss advantages and disadvantages and then list the principal scan patterns.

line 104: "RHI scans allow for measurements closer to the ground,"

 A PPI can also be performed at lower elevations. I do not see why it must be a RHI,

line 111: "The wind vector retrieval error is determined by ..."

  I do not understand, is this doubled in the next sentence ?

  

line 117: "Steep measurements" 

  I guess you mean "...  at high elevation angles"

line 118: "shallow scans" 

  I guess you mean "...  at low elevation angles"

  

line 118: "... retrieval errors due to turbulence typically average out within a 10 min to 30 min span,"  

  But the uncertainty never vanishes - you cite Rahfles et al (2022) who showed that RMSD becomes never zero and is surprisingly large.

line 125: "... some instruments provide a signal-to-noise ratio (SNR),"

  I would appreciate some words what the difference between CNR and SNR in Windlidardata is, and whether there is a difference in the use of it.

line 123: "Filtering with the carrier-to-noise ratio (CNR) is ..."

  Please mention that with decreasing CNR the uncertainty/the error increases, Also mention parameterizations: There are estimates for the doppler retrieval uncertainty - see eg Pearson and Collier (1999, doi:10.1002/qj.49712555918), Manninnen et al 2016 (doi:10.5194/amt-9-817-2016) who they refer to Rye and Hardesty (1993)  and O'Connor et al. (2010), 

Also worth mentioning is that if CNR becomes too low the doppler speed becomes evenly distributed in the bandwidth (as is visible in Fig D3 in this paper). 

line 127: "Potential causes of erroneous measurements despite high CNR..."

  I am missing precipiation ... 

  

line 138: "if the beam directions of the radial velocity measurements are not sufficiently dispersed ..."

  I would prefer the word 'distributed [in space]' instead of dispersed.

  (the term dispersion has a different meaning in optics, and another in Chemistry, ...)

line 140: "The condition number (CN) is a measure for the robustness of the beam dispersion"

  I would write: "… CN is a measure for the robustness of the equation system with respect to errors in the input variables. "

line 143: "Maximizing the number of considered measurements allows for the smallest assumption violation, due to the beneficial effect of averaging."

  This is the basic argument for the combination of several different scan patterns in the retrieval. In this universal formulation I doubt it, especially because it could be a violation of the argument about the spatial beam distribution of the previous paragraph. Fig.1 is a nice example: the RHI has a much higher spatial coverage than the VAD, but in only one direction. In a VAD scan a higher number of measurements in one direction leads to a lower condition number (Paeschkae et al 2015).

line 163: "Changing absolute range gate heights for different laser beam positions can be overcome by binning the measurements with respect to the height above ground."

  If understand right this means that radial velocity measurements from different heights within one height bin are assumed to represent the same wind vector. This neglects the height dependence of the wind, and accordingly violates the homogenous wind field assumptions. 

line 180: "timestamps of the measurements,"

  Any requirement whether this should be the begin or end of the averaging interval?

  

line 181: "The time coordinate can differ between instruments at level 0 and will then also differ at level 1."

  Unclear, Please clarify.

  

line 194: "... azimuth (time), elevation (time) ..."

  When performed in continuous mode azimuth and elevation vary during the integration time. It should be mentioned which angles must be provided here (begin, middle, end of integration time). 

line 212: "Level 2 represents wind vectors ..."

  I guess you mean 'Data level 2   repesents ..."

line 239: "flagging of precipitation."

  Is there a module to flag precipitation?   If yes: how is this done ? 

line 250: "The validity_probability variable contains an acceptance (1.0) or rejection (0.0) value for each bin according to a user-defined threshold."

That seems rather to be a flag, or is it possible to implement fuzzy logic based on values between 0 and 1. ?

Table 1: "(8) Bin_statistics_l1_to_l2 ... provide CNR and Doppler spectrum width information to l2"

Interesting! How is this done ?

Doppler spectral width can be regarded as a standard deviation  - not in time but in the measuring volume of the respective range gate. In Sathe et al (2015) equation 4 can be seen that the along beam variance depends on the variances and covariances of the principal components of the wind vectors. Thus a retrieval of doppler spectrum width in the principial spatial components could be done with this. But does it work ?

line 376: "Such an imbalance negatively impacts the vertical wind vector calculation and is, therefore, undesirable."

  This is the argument against the general statement from line 143 that the number of measurements should be used. Line 143 should be adapted. I am wondering how this affects only / especially(?) the vertical wind. Maybe that is result of the strong wind shear in the height bins close to the surface?

line 384: "... a strict CNR filter (the default threshold for Leosphere WLS200s is -25 dB) ... a weak CNR validity flag (default threshold -30 dB for Leosphere WLS200s)"

A guideline on how can these thresholds could be determined for instruments from other manufacturers would be desirable. I would be also of interest to mention at some point how these kind of thresholds or parameters are provided to the software and how they can be changed.

line 392: "... the volume enclosed in the convex hull spanned by the unit vectors"

This is unclear: If I imagine a VAD scan with constant elevation the volume spanned by the endpoints is a polygon (lets say a disk) with vertical extension equal to  zero. Its volume would be zero. If it is the whole cone spanned by the vectors it is easily much larger than the threshold of 0.2.

Please clarify.

line 406: "... even at low CNR reliable measurements may be available in some circumstances,"

  Radial velocity uncertainty increases with decreasing CNR. Values will lie in a large interval around the true value. The value can lie just by chance in the trustworthy range around the 'confidence background'. It does not contain information. It becomes trustworthy just by coincidence. I see this is addressed by limiting the share of accepted values to a minimum.

 

line 412: "...the confidence background is extrapolated in module (7.2)."

How does this extrapolation work ?

In fig 9 it is obvious that wind speeds are extrapolated into large regions where the first guess gave no data. This bears the danger to create unrealistic values. Application of this confidence background to low CNR values may let pass through otherwise unplausible values.

Please discusss this.

line 463: "Whether a higher availability from more iterations justifies the increased calculation effort can be decided by the user in the individual application case."

It would be great to see whether the additionally retrieved wind vectors are of good quality.

line 475: "Heterogeneous scans of type DBS, step-and-stare, and sector PPIs are used for wind profile retrieval (Fig. E2 (a))."

I see the sector PPI at an elevation of only a few degrees in a sector of only 45deg width. Does this data contribute to the retrieval or is it excluded as it is too low and too far away?

line 478: "The lidar conducted PPI scans ... (Fig. E2 (b))."

This looks as if the lidar performed its scan in continuous mode. Please clarify.

line 486: "... also termed VAD pattern for Halo systems ..."

This term is not Halo specific. You can find it for wind lidar scans already in Browning and Wexler (1968). The synonyms PPI and VAD could be introduced with the scan patterns.

line 492: "At weak SNR, very frequent vertical winds with 1ms-1"

You could refer to fig.D3(c) - it is clearly visible

line 507: "Fig. 12 shows the comparison of the radiosonde measurements with the corresponding retrieved horizontal wind speed".

How are radiosonde and wind profile data matched: The radiosonde needs about 16minutes for  1km or more than 1h15m for the 5km displayed. Is it compared to the wind profile at start time, one profile in between or to several lidar profiles according to closest time ?

Fig 12: 

Instead of repeating the scatter plot with different, difficult to depict colours, it would be helpful to see differences lidar - radiosonde as a function of the parameter of interest i.e. radiosonde distance, vertical velocity and eventually wind speed.

Fig 12 - caption:

"For Villingen-Schwenningen, one radiosonde profile is omitted"

  I guess the crosses are the data from this profile.

Fig 13: 

Systematic negative Delta v above 4km:

I recognize that the interquartile range does not deviate as much as (arithmetic?) mean, i.e. only few extreme datapoints with Delta << -4m/s (light grey shading) are responsible for this negative Bias. N is also low, so this maybe only one single Profile ?

I am wondering whether this extreme bias is result of the extrapolation of the confidence background. Similar deviation are visible in fig 14 and 15.

line 534: "For Payerne and Villingen-Schwenningen, the comparison shows good agreement"

What about the negative bias at Villingen-Schwenningen between 0 and 3km ? Ah, i see this is discussed later.

line 539: "The slight variation in the number of retrieved wind vectors ..."

This is an aliasing effect. Heights bins are every 100m. Lets assume your range bins map to heights every 30m. As 100m and 30m do not have a common denominator you get alternating 3 and 4 measurements in one height bin. As several measurements are collected in every height bin, the variation becomes a multiple of 3 and 4, respectively.

line 547: "(bias) in the u component of about 0.4ms-1"

I see negative values in fig 14.

line 550: "The u-component shows *a* systematically higher mean values,"

(Typo with the 'a' )

The bias in the plot is negative and I thought it is 'lidar - radiosonde'.

I have the impression that at least the total range of the deltas is larger for the u than for the v component. 

This paper describes a wind profile retrieval software called AtmoProKIT to retrieve wind profiles from Doppler Lidar measurements. This open-source retrieval code is modular with standardized data formats and configurable module chains to handle complexity related to different types of Doppler Lidars, and different types of scans (PPI, RHIs, step-stare) in order to ensure quality controlled, standardized wind profiles with traceable uncertainties.  While availability of an open-source, standardized wind profile retrieval code for Doppler lidar would benefit the scientific community, I am not sure the paper fits the scope of Geoscientific Model Development journal. It is primarily focused on describing the publicly available code to retrieve wind profiles from measurements, and nothing to do with model development or validation. One could argue that adaptation of this code could result in uniform wind profiles for data assimilation and model validation. The following comments need to be addressed before the paper is accepted for publication.

 Major Comments:

Wind profile retrievals from heterogeneous Doppler lidar data: Heterogeneous Doppler lidar data from different scans have been used to retrieve wind profiles on a different vertical retrieval grid before (e.g. Tucker et al., 2009, Bonin et al., 2017, Pichugina et al. 2019). Pichugina et al. (2019) used wind profiles retrieved at high vertical resolution near the surface from multiple PPIs, including very shallow elevation angles, to evaluate NWP for wind energy applications. Thus, the use of heterogeneous Doppler lidar data and the concept of retrieval volume is not novel.  

Choice of CNR Threshold and Confidence Interval: The choice of upper CNR thresholds of -25 dB for WLS200s seem arbitrary, and very conservative. Based on Figure E1, the CNR threshold should be closer to the lower CNR threshold i.e. -30 dB. By arbitrarily setting the upper threshold at -25 dB, a lot of data is categorized as potentially bad, and exaggerates the benefit of the confidence background method presented in the paper. I suggest the authors characterize the noise floor for one of the WLS200s presented in the paper and then assess the benefit of the confidence background method.  If a proper noise floor (e.g. -29 dB) had been considered, I think all the gains made during the first iteration shown in Figurer 11a would have been included in the retrieval in the first place. In such a case, the gains from using the confidence interval method would be minimal and could even argued as not worth the effort.

Note that depending upon the definition of the CNR used by manufacturers, and measurement bandwidth, usable CNR threshold for a Doppler lidar will be different. Thus, noise floor for each lidar needs to be characterized independently so that all good data are used in the retrieval.

What CNR thresholds were used from the WLS200s at Payerne, WTX at VS and StreamLine XR+ at Neumayer Station? Based on figure D3, the CNR threshold for WLS200s, WTX and StreamLine XR+ should be very different.

Some peak detection algorithms used by Doppler Lidars to determine line of sight velocity (LOSV) follow the strongest peak from the high CNR region to lower CNR region. This could potentially result in wrong LOSV determination in presence of wind shear. This could be the case with StreamLine XR+ at Neumayer Station. The confidence background method would potentially double down on determining these erroneous LOSV data points as good data points. So, one needs to be cautious when applying methods such as confidence interval to determine reliable data points in the low CNR regions.  

Uncertainty Analysis: There is no description of the uncertainty in the paper. How is it calculated? How is it impacted by having different types of scans? Depending upon the scans and retrieval grid used, some grids will have more data points than others, how does this impact the uncertainties? Please add how uncertainty is calculated in Appendix C.

Comparison with radiosondes do not provide retrieval error for individual wind profiles. Depending upon the atmospheric conditions, some wind profiles will have larger uncertainty than others. How is this calculated and is the calculated uncertainty representative? This is especially important for low CNR regimes. Bulk comparison as shown in the paper usually hides the larger bias and uncertainty likely present in the low CNR regimes.  For example, all three stations show a bias in v at higher altitudes (Figure 13, 14 and 15). How does this comparison look if only the low CNR regime data are included? I suggest the authors characterize performance/uncertainty as function of CNR.

Minor Comments:

First Guess: It is not appropriate to call the results of the initial VAD fit as first guess. It is a result of the full VAD retrieval, and not a guess. I suggest the authors replace it with initial result or first result.

Line 126: “CNR is not a completely reliable indicator”. This is a misleading statement. One could argue no variable is a completely reliable indicator. One could always filter out very high CNR data to remove contamination from hard targets or moving objects. I suggest the authors revise this statement to “CNR for each lidar needs to be carefully characterized to determine the appropriate thresholds for reliable data”.

Line 182: How and why does time coordinate differ between level 0 and level 1 data? It seems level 1 data is just standardizing the data format for different types of lidars?

Line 188: It is not clear why anyone would want to change the range gate settings for different scans. It seems less practical and adds to complexity to lidar operation, data analysis, and in case of this paper, complexity of description. For simplicity, I suggest the authors use same settings for all scans so that the number of range gates is limited by the pulse repetition frequency (PRF), and all scans use same number of range gates. In that case, Figure 2 will look same as Figure 3 with all the block filled in.  

Line 188: I would assume level 0 data is in range gate index, and you add range information in level 1? For example, the StreamLine level 0 data used to include range gate index and not range. Maybe this has changed with newer versions?

Line 446: The lower number of retrieved bins in simple -30 dB CNR threshold retrieval due to +/-3 m/s filtering?

Figure 7 & 8: It would be better to include Figure 8 as the 4^th panel in Figure 7. Figure 8 caption: The data gap is at 21:40 UTC.

Figure 12: Is the omitted radiosonde profile shown as x?

References:

Tucker, S. C., W.A. Brewer, R.M. Banta, C.J. Senff, S.P. Sandberg, D.C. Law, A.M. Weickmann, and R.M. Hardesty, Doppler lidar estimation of mixing height using turbulence, shear, and aerosol profiles, Journal of Atmospheric and Oceanic Technology, 26, 673–688, 2009

Bonin, T.A., B. Carrol, R.M. Hardesty, W.A. Brewer, K. Hajny, O. Salmon, and P. Shepson, Doppler lidar observations of the mixing height in Indianapolis using an automated composite fuzzy logic approach, Journal of Atmospheric and Oceanic Technology, doi:10.1175/JTECH-D-17-0159.1, 2018.

Pichugina, Y., R. Banta, T. Bonin, A. Brewer, A. Choukulkar, B. McCarty, H.S. Fernando, J. Kenyon, R. Krishnamurthy, M. Marquis, J. Olson, J. Sharp, and M. Stoelinga, Spatial variability of winds and HRRR-NCEP model error statistics at three Doppler-lidar sites in the wind-energy generation region of the Columbia River basin, Journal of Applied Meteorology and Climatology, doi:10.1175/JAMC-D-18-0244.1, 2019.