Wind Profile Reconstruction Based on Convolutional Neural Network for Incoherent Doppler Wind LiDAR

Round 1

Reviewer 1 Report

This study presents an investigation of how the accuracy of line-of-sight measurements acquired by an incoherent Doppler lidar can be improved using a convolutional neural network. The study focuses on high altitude measurements (>10 km). At those heights wind observations are scarce and thus output data from ERA5 are used to simulate Doppler lidar measurements. Those simulated measurements are used to train a CNN algorithm which subsequently is tested in field observations. The topic of the study is in principle relevant to the scope of the journal of Remote Sensing. However, I think that there is need of improvements in the way that the methods and the results are presented, since in the current version there is lack of clarity in many points. Please find below my suggestions and comments on the current version of the manuscript. 

General comments

1.     The title of the article includes the term “Wind Field Reconstruction”. However, as I understand the material included in the manuscript there is no wind field reconstruction that takes place during the analysis of the data. The authors use a CNN algorithm to produce more reliable line-of-sight measurements which are acquired in different heights. Why is the term “wind field reconstruction” used?

Moreover, in the manuscript it is frequently written that the CCN algorithm is used to improve the wind field reconstruction. This is not accurate, please correct this point in the whole manuscript.

2.     Furthermore, since the authors present measured and simulated vertical profiles of line-of-sight speeds, I suggest replacing the legend of all relevant figures where the term “wind speed” is used with the term “line-of-sight speed”. 

3.     I suggest reducing the Section 2.2 “Theoretical Principle” by removing the presentation of the mathematical equations and instead refer to already published literature like the reference number [17]. My suggestion is because the mathematical formulation to determine the Doppler shift frequency is not focus on this article.  

4.     The article claims that ERA5 data are used to simulate wind lidar measurements that are used as input to simulate wind field. However, there is no information in the article regarding how this is done. This is necessary for understanding the methodology used and for assessing the accuracy of the results. Therefore, it is necessary to describe thoroughly the steps followed to produce the simulating wind fields. 

5.     In section 6, a comparison between the “argmax” and “spectral centroid” methods is presented. It is not explained what the purpose is of comparing those two methods. 

6.     Several of the authors of this manuscript are also authors of the following study:

Song, Y., Han, Y., Su, Z., Chen, C., Sun, D., Chen, T., & Xue, X. (2024). Denoising coherent Doppler lidar data based on a U-Net convolutional neural network. Applied Optics, 63(1), 275-282.

In that study the U-Net CCN is used, why is the CNN is described again here and not just use the Song et al study as a reference?

Specific corrections

Pg1. Lines 15-16: The authors write: “However, this approach has limitations in handling complex wind data, particularly in low Signal-to-Noise Ratio (SNR) regions.”. What is meant with the term “complex wind data”. 

Pg1. Lines 39 – 41: Please add a reference to support this statement. 

Pg2. Lines 50 – 51: First the references [15] and [17] are not presenting a “Spectral Centroid” method. Both studies use the double-edge technique for determining the frequency of the Doppler shift. Please clarify what is meant with the term “spectral centroid” and provide an appropriate reference. Furthermore, the methods used to determine the Doppler shift frequency are used to estimate the line-of-sight speed and not for a wind profile reconstruction.

Pg2. Lines 56 – 58: What are the limitations that the authors have found in the already published literature regarding the use of machine learning algorithms to analyse wind lidar data? And why those cannot be used in the study presented here? 

Pg2. Lines 72-74. How was U-net-based CNN model chosen?

Pg4. Lines 106. The authors write “This ensures that the beam remains focused and tightly collimated over long distances”. Is the beam focused or collimated at longer distances? And how long are the “long distances”?

Pg5. Line 166. How is the frequency of the maximum intensity of the locking filter in Fig. 4 chosen? And are all the curves presented in Fig. 4 scaled in the same way?

Pg8. Line 226: To be able to use Eqns. 12 based on the measuring configuration suggested in this article, it has to be assumed that the wind speed is on average homogenous over horizontal distances of kilometres. Is this a reasonable assumption? 

Pg8. Lines 251 – 253. How close exactly was the location of the two Doppler lidars of Fig. 1 with the coordinates of the ERA5 data products used in this study?

Pg8. Lines 257 – 259. First, if by spectral centroid algorithm is referred to the estimation of the Doppler frequency based on the centroid of a spectrum then this estimates the radial wind speed. I think that this point should be clarified here. Furthermore, I do not understand what is meant with the “average wind field”? Is it meant the average line-of-sight along the probe volume? 

Pg. Line 263. What is meant with the word “label” here?

Pg9 Line 278. What is meant with that the two components (“mean” and “perturbation”) need to be “furtherly discriminated”? Please elaborate. 

Pg10 Line 294 -296. Clarify the terms in Eq. (13).

Pg11. Lines 301 - 302. What is the meaning of this sentence?

Pg11. Lines 309 – 310. Figure 6b. According to the figure the lidar measuring error between 40 km and 50 km increases from 100 m/s to 400 m/s. I am wondering what is the purpose of presenting measurements over those heights if those are characterized by very large errors?

Pg12. Lines 366 – 368. Please explain what are the ReLu activation layers and what is the maxpooling layer?

Pg12. Line 374. What is a 3-challened image?

Pg13. Line 387. What is the ADAM optimizer? 

Pg13. Line 405. How is the “near field” defined? The authors write “where the signal intensity is relatively strong”. What is meant with that?

Pg14. Line 419. How are the sections specified?

Plg15. Figure 11. Each of the 4 plots included in the figure corresponds to a different “ground truth”. Therefore, it is not possible to visually compare neither the differences of the two Doppler shift estimation nor the output of the CNN. Please use the same reference in all four plots. Also, what is the averaging time of the profilers presented here?

Pg16. Line 448-449. Figure10 basically presents simulated Doppler spectra per height. You should also add profiles of the SNR, since you refer to “high SNR and intensity region”.

Pg16. Line 457. Please describe what is the Gaussian smoothed spectral centroid method. 

Pg16. Lines 474 -475. Which heights are used for the estimation of the r-squared and what is meant with the Max. and Min. in the table?

Pg17. Figure 12. Do the outliers in these two plots correspond to high altitude measurements?

Pg17. Lines 493 – 502. How low was the SNR of these cases? And how the authors can explain the negative correlation?

Pg17. Sections 6.2.1 and. 6.2.2. How are the “low” and “high” SNR conditions defined and how are the corresponding profiles simulated? What I mean is the SNR of the measurements the same at all heights?

Pg19. Lines 546 –553. Which two quantities where they used to estimate the R2? If it was the GSSC and the SCCNN this does not say something about the accuracy of the CNN output since there are not reference wind speeds to be compared to. Have the authors considered to compare the lidar measurements with ERA5?

Pg19. Lines 570 – 571. The improvement of the results of the analysis of the CNN output rather than height dependent is SNR dependent. As mentioned in one of my comments it is not clear what it the improvement at high altitudes from the results presented in this version of the manuscript. The authors could consider including the average error between the reconstructed wind profiles and the ERA5 profiles to support their argument.

Pg19. Lines 579-580. Why was the random error assumed to be 1m/s and what is the implication of this assumption?

Minor corrections

Pg1. Line 12: Please replace “acquired from remote sensors” with “acquired by remote sensing sensors”.

Pg1. Line 35: What is meant with the sentence “This has greatly enhanced the compatibility of data analysis in all dimensions …”? Please clarify.

Pg3. Line 88: Replace “should consist” with “consists”.

Pg3. Line 93: Please write what does the accronym FPI denotes. 

Pg5. Lines 157-158. Please speficy that this statement applied to incoherent Doppler lidars.

Pg6. Line171. Please delete the word “ideal” .

Pg6. Line177. Please delete the word “accurate”. What is described here is the principle of a frequency estimation using an incoherent wind lidar. Whether this happens accurate or not is dependent on the technical specifications of the instrument.  

P8. Line 287. Replace “donates” with “denotes”.

Pg18. Line 545. Correct the 4000 km. 

The manuscript includes grammatical and syntax errors. Please do a thorough check of the text and improve it. Most of my comments regarding minor corrections concern grammatical corrections. However, the list should not be treated as a complete list of the minor corrections needed, but rather the errors that I have spotted.  

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

This manuscript introduced a U-net deep learning algorithm to reconstruct the wind field observed by the Doppler LiDAR systems. This approach improved the precision of the re-constructed wind filed, especially for those with low signal-to-noise ratio. However, the methods and results should be described detailed and further examination are required.

Major comments:

1.     The introduction of the LiDAR system from line 84 to 241 is unnecessary for the topic of this research. A few sentences are enough or a reference about the LiDAR system may be cited.

2.     More detailed introduction of the methods used to reconstruct the wind field, such as SC algorithm, Spline Transformer, etc. is necessary.

3.     The results should be examined more carefully and clearly. For example, the reconstructed wind field should be compared with observations, such as radio sounding. At least, the LiDAR raw observations and reconstructed winds should be compared with ERA5 data.

4.     Generally, the observations with very low SNR are not useful. In this study, the SNR of the observations above 20km are very low. Are those data suitable for reconstructing wind field? In other words, are these LiDAR systems suitable for observing winds above 20km? In addition, as shown by the Figure 6, the errors are much bigger than the wind perturbation above ~35km, are those observation useful?

5.     The captions of the figures are too simple. The contents of the figures are not clearly described. 

6.     Are the contents of the Figure 7, 9, and 10 observations of one time or a period? The wind profile reconstructed by this research is for one time or the average of a period?

Minor comments:

1.     The description of ERA5 data is not precise (line 243-248). ERA5 dataset is not a model. Are LiDAR observations employed by EAR5? The spatial resolution of ERA5 are 0.25 degree * 0.25 degree.

2.     Did the Butterworth filter (line 274) applied to a wind profile or series wind profiles to get the background wind in figure? In other word, is the background wind profile in figure 5 vertical averaged of one profile or time averaged of series profiles?

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

In the paper，the authors propose the use of Convolutional Neural Network（CNN） to reconstruct atmospheric wind fields. The paper is interesting and relevant in content. However, there are some provisions that require additions or clarification.

1. The introduction highlights the importance of the CNN in detecting wind fields.

2. The last part of the introduction should introduce the work content and structure of this paper instead of emphasizing the defects of previous research.

3. The second chapter hardware system part does not need too much language description.

4. The wind direction appears in the annotation of Figure 13. How does the CNN perform in wind direction inversion of wind field?

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

I would like to thank the authors for providing an answer to my comments and questions. They have covered almost everything, but I think that there are still four points that need to be improved. In specific:

 1.        There are still sentences in the text where the term ”wind field” is used instead of “wind profile” (e.g Section 2.2.3)

2.       The authors have used the East-West wind speed from ERA5 to estimate the background wind profile. How was the East-West wind speed estimated? And how close was the location of the ERA5 data to the location of the line-of-sight lidar measurements?

3. Figure 8. The authors state the “ground truth” is the same for all plots. However, the profile of the “ground truth” is different in each plot. Why is that?

4. The authors in their reply to my question regarding the accuracy of the lidar measurements at high altitudes they answer “The GSSC measurement is accurate with sufficient SNR but varies a large uncertainty when the SNR is low”.  I think that they need to support this statement either through the comparison to a reference data set or include a reference of a previous study that has demonstrated the accuracy of the wind lidar measurements at high altitudes. Otherwise, what the authors show is that the GSSC method provides line-of-speed wind speed estimations that have the same precision regardless the SNR level, but this does not say something about the accuracy of the measurements. 

No comments

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

This revised version was improved obviously. This manuscript introduced a CNN network to reconstruct the wind profile observed by a LiDAR system. It's a new approach. However, the precision of the observed and reconstructed wind profile are not clear. In the stratosphere, the wind speed could be 100m/s, but the vertical shear could not be so big. In this manuscript, the facts are lack to demonstrate the precision of the observations. For example, in figure 11, the difference between the observed wind profile and ERA5 are very large. Meanwhile, there are some questions.

1) In this manuscript, are the perturbations of the wind speed the real perturbation of atmosphere motion or just the errors of observation?

2) How to demonstrate the accuracy of the perturbations of the wind speed observed by the LiDAR system?

3)Line 97-98, LiDAR point at a zenith angle of 30 degree with the azimuth angles in the East-West direction. Is that a line or a PPI scan?

4)What is the background wind trend of ERA5? (line 139)

5)Line 146-147, what does "horizontal wind speed at height of 195 meters and 390 meters for the east-west direction" mean?

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Round 3

Reviewer 2 Report

This revised version was further improved. There are some comments.

 As I read in this manuscript, the results of SCCNN and GSSC were compared to demonstrate the performance of SCCNN. Meanwhile, some results were compared with ERA5 data. As my opinion, the results of GSSC and ERA5 are not in-situ observations. In other words, the results of this paper were not verified using observations of other instruments. If so, authors should mention it in the section of Conclusion and Discussion.

Author Response

Please see the attachment.

Author Response File: Author Response.docx