Estimation of the Wind Field with a Single High-Frequency Radar

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

In this manuscript, the authors proposed a method for estimating sea surface wind direction and speed from first-order echoes of a single HF radar. Experiment results through validation using in-situ and model data show that the proposed methods are effective and achieves the best performance than existing methods. The topic fits the journal and the manuscript is well written. I suggest the authors consider the following problems in revision:

 

Technical comments:

1.     Lines 91-92, are these numbers for wave height? Clarify it. If yes, in later part of the manuscript, you mentioned the wave heights are 1-3 m, but the directional spreading model is for 5-20 m, some explanation is suggested.

2.     Line 105, whether you are using left or right beam, more details are required.

3.     Explain the reason why the specific numbers of neurons are chosen.

4.     What is the water depth in the observation area?

 

Other comments:

1.     Line 3, correct “…direction and speed is…”.

2.     Line 80, “section” should be “Section”, same problems also exist in other locations.

3.     Line 120, the plus sign is not required before the variable.

4.     Line 189, add a period before “To”.

5.     Line 258, correct “…direction and speed…relies…”.

Comments on the Quality of English Language

Fine.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This is an interesting and generally well-written article on an important topic.  Unfortunately, there are a number of what I would consider to be fatal flaws in the paper.  Most importantly, I think the authors have abandoned an analytic approach too quickly in favor of neural networks, which they concede are difficult to interpret and do not generalize to other radars, regions, or conditions.  Neural networks are used to determine the relationship between the wind speed and wind-driven current speed, which they say is "complex and cannot be precisely expressed through an analytical expression."  That may be true, but surely some work has been done in this area and at minimum the relationships arrived at by the neural network should be examined and compared with previous work.  

The second major shortcoming is illustrated in equation (5). The total surface current observed by HF radar is decomposed into three components, including the tidal current, which can be predicted, at least in principle.  The other components are described as the time-averaged current and the residual current.  A more appropriate division, I would think, would be into wind-driven and non-wind-driven currents (such as riverine and geostrophic currents).  In any case, there does not appear to be any further discussion of the currents (other than tides) that are unrelated to the wind.  Such non-wind-related currents must certainly be present in the experimental region, and I suspect that the effects of these currents are lumped into the neural network processing.  Examining the relationships produced by the neural network might uncover these effects, and explicitly accounting for these effects might produce a more general and widely applicable algorithm.

One minor comment is that Figure 10 is difficult to read, and I don't for example see any brown or yellow arrows. Separate interpolated colormaps of the wind speed and direction might be more effective.  It is also not clear to me what is the difference between the validation and test sets (lines 148-149).

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

This is an excellent and very interesting paper. I have a few main suggestions that I think will make the paper better and more useful to other investigators going forward. 

 

The first main suggestion is to add more information about the data being used and the results obtained. Because the N used is not clearly defined anywhere, it’s not clear how significant the results are.  I suggest adding a table to describe the data used, for the two years and the different radars, including the time ranges and the total number of data points. Figures and the table 1 reporting results should include the RMS difference and the SI (which might be more commonly referred to as a normalized RMS difference), and the N used in the calculation. Also it’s not clear when which correlation coefficient is being used, of the two defined. I’m not familiar with these so perhaps it would be clear if I looked into the calculations. An important detail is that its not entirely clear if the correlation coefficients and other results are computed on the test data or the whole dataset. If the figure 8 represents the 15% of the data that is the test set, I would suggest using longer test/validation sets and or performing cross validation to show how the model is performing overall, and to demonstrate that the excellent results are not due to the choice of test data. The text mentioned randomizing, which suggests to me that the results are computed on the whole dataset, including the training data. This should be stated clearly, and perhaps the statistics of the test data or cross validation results reported separately. 

 

The other main suggestion is to more clearly explain the method for obtaining direction. The abstract says the method is new, then describes a well known method for obtaining the wind direction. Later in the text (about line 100), the new method is described briefly and in my opinion incompletely. I think this could use a bit more explanation to be clear what is being done. What is meant by the ‘left or right solution’? What is the solution matrix, and how are these concatenated? I’m not certain if left/right is referring to the wind directions or the Bragg peaks. This method deserves a more substantial paragraph and perhaps an equation if warranted. 

 

A few suggestions for the abstract: Move line 14 “Wind speed … “ up to about line 7, since this is one of the main objectives of the paper. Include RMS and perhaps SI results in the abstract, along with the correlation results. 

 

In the discussion or conclusion section, it would make the paper much more generally applicable to specify (or speculate) whether or not these results depend on the residual currents being highly related to the wind. Perhaps a figure scattering residual currents and wind speed would make this clear. What do the authors think about how this would work in typical HF radar deployments (e.g. with more complex current/wind environments, swell, etc)?

 

Other suggestions:

 

Line 45 might want to make it clear that the 2nd order has lower SNR and thus decreased range.

 

Suggest a similar figure to figure 6 showing the SI vs SNR if possible.

 

Suggest a new paragraph on line 144 starting with “Training…”

 

Line 153: It would be useful to readers to define the correlation coefficient equations here. Also, it appears that the reference description for [18] is incomplete. 

 

As noted above, Figs 7 and 8 should make it clear if this is the test data being shown in the caption and I would like to see the RMS difference between these and the HFR results along with the SI values. 

 

Line 292 “should be approached with caution” - is there any expectation that these results would apply more generally? If not, could this be done with future work?

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

I would say that the second version of this paper is improved and can be published in substantially its present form, but I would suggest making a few additional modifications as discussed below, particularly those relating to Figures 4 and 5, and equation (7).

I am still skeptical that neural networks are the only way or necessarily the best way of determining the relationship between the wind speed and wind-driven current speed, and I still think that the relationship derived by the neural network should be compared with previous work (e.g. ref [17]). 

Regarding my second comment, I am surprised that riverine currents are always negligible in comparison with wind-driven currents in the Saint Lawrence River, but I don't have any data to back that up, so I will take the authors' word for it.

Regarding the new material, the new method for disambiguating the wind direction (lines 109-117) does not seem to me to be materially different from that of Huang et al. (2004), since the wind directions with the narrower distribution are likely dominate the modal direction. The examples given in Figures 4 and 5 are helpful, but it would be better if those figures were more consistent with each other.  In Figure 4, the left-hand solutions are less variable (have a narrower distribution) than the right-hand solutions and are more consistent with the "true" wind direction indicated in the upper right corner. In Figure 5, the roles of the left and right solutions are reversed.

In equation (7), the complex wind direction should be defined. It is fairly obvious that the real and imaginary parts of the complex wind speed correspond to the east and north components of the wind vector, but this should be made explicit.

Author Response

Please see the attachment

Author Response File: Author Response.pdf