Suggestions for revision or reasons for rejection  I'm not sure that the authors correctly obtain estimates for SNR (ratio of peak spectral amplitude to rootmeansquare deviation of the noise spectrum) less than 5 dB. The fact is that at low SNR (for example, at SNR = 2 dB), the high probability that the spectral maximum is associated with noise, rather than with a signal. For the bandwidth of B = 100 MHz and the width of the time window T = 200 ns, the number of independent spectral channels L = B * T = 20. Accordingly, the number of peaks in the measured spectrum is not less than L / 2 = 10. When the amplitude of the signal peak in the spectrum is comparable with noise peaks, the probability that the signal peak is the maximum of the spectrum is about 1/10 = 0.1. Because of this, if the mean and the standard deviation of the noise spectrum are correctly obtained, the SNR estimate will be at least 2.5 (4 dB), even when there is no signal at all. The authors can verify this by using numerical simulation. The algorithm described by Eqs. (8)  (10) here does not help.
The authors have done a good job: an algorithm for processing of lidar measurements has been developed, which makes it possible to compensate for the measurement errors associated with the motion of the ship; joint measurements of the wind by a lidar and a radiosonde were carried out and a comparative analysis of the results of these measurements was made. The efficiency of the developed algorithm is shown. I think that the manuscript could be accepted for publication in AMT, if the authors will remove the results and analysis of the data received for the SNR less than 5 dB from the manuscript.  

Suggestions for revision or reasons for rejection  The manuscript was significantly enhanced and revised. All my major comments and most of the minor comments were addressed in an extensive reply and also revised in the manuscript. I thank the authors for this very good reply. I have only a few technical comments
1) Explain acronyms MEMS, RTK, CEP, CFD
2) Mention distance hard target to lidar
3) "2min" should be "2 minutes"
4) Fig. 8: replace "RMSM" by "RMSE"  

Dear Authors
Thankyou for your revised version of "Shipborne Wind Measurement and Motioninduced Error Correction by Coherent Doppler Lidar over Yellow Sea in 2014", which you will see is much improved in the eyes of the reviewers. There remains a small number of issues to address, so could you please respond to the points made by the reviewers?
regards
Murray Hamilton
(Assoc. Ed.) 
Dear Authors,
re
AMT2017206 Shipborne Wind Measurement and Motioninduced Error Correction of a Coherent Doppler Lidar over the Yellow Sea in 2014, Zhai et al.
This paper is nearly of an acceptable standard for publication but first there is some clarification needed.
The sticking point is in the response to reviewer #1 where they question the validity of SNR determinations below 5dB. Unfortunately I think that the text that has been added to the manuscript serves to confuse the issue as much as explain.
The referee is assuming that, for a particular range bin, the FFT is calculated with a data time series of duration 200 ns, and that the sampling of that time series was done with a sampling interval of 5 ns. The assumed sampling interval is based on the bandwidth that you give, of 100 MHz. (As an aside, what is the significance of the subscript "100" on the B?) Presumably the reviewer has assumed that you then chose the sample rate to be 200 MS/s (twice the bandwidth according to the sampling theorem or Nyquist criterion). It is seems to me however that you are oversampling at 2 GS/s, based on your statement of the frequency resolution being approximately 1 MHz. It would be helpful if you were explicit about the sampling rate.
However the main part of the reviewer’s objection centres around the fact that the uncertainty (variance) associated with the estimate of spectral power (or amplitude  you aren’t very specific on this) in a single bin is equal to that estimate, because it must be assumed to be a random variable with Poisson statistics. Thus if you have the power in one bin sticking up above the average power in the bins, so that it is a factor of two larger than the average (3dB), there is a significant probability that this is merely a background fluctuation, even if it is the highest bin. This assumes that you have just the one FFT for that particular range. Now if you average spectra the uncertainty (variance again) of the power in each bin (in the average) is reduced by a factor of the number of spectra used in the averaging. Then the reviewer’s objection would not hold.
Earlier in the paper you state that you do averaging in fact, but you don’t really say how. This is important because if you average the signals before calculating the FFT, rather than averaging the spectra, the reviewer’s objection holds.
There are other issues with the text that has been added. The word “spectral” is an adjective for one thing, which is somewhat distracting. The quantity S that it refers to is not well introduced, and it is not obvious or explained, what the reader is seeing in the new figure 3.
regards
Murray Hamilton
(Assoc. Ed.) 