Improving Shallow Water Bathymetry Inversion through Nonlinear Transformation and Deep Convolutional Neural Networks
Round 1
Reviewer 1 Report
This article discusses the application of deep learning to shallow-sea bathymetry, providing a great deal of model detail, detailed experimental results, and largely reliable conclusions. Limited by my expertise, I am unable to evaluate all the details, and I would like to ask the authors to refer to the following questions for further revision and improvement.
1. P1, L 27, ‘The first is…’, where is the second one?
2. P2, L33, Lasers are not suitable for measuring turbid waters, and many state-of-the-art lasers (e.g. czmil) do not give good results in turbid situations, due to natural optical properties. If you want to describe it that way, please add the relevant literature.
3. P2, L43-46, The sentence is too long, you can rewrite it as 2 sentences.
4. P2, L46, Was the ratio model the first model proposed? Be very sure and add references.
5. P6, L201-202, The author should give a more detailed explanation as to why this is the case.
6. P6, L216, The text refers to depths up to -2m as deep water and shallows up to -2m as shallow water, is this agreed upon based on the distribution of water depths from the experimental data in this paper, or is it an industry consensus? The author should explain.
7. P7, L257, Please add time information of worldview-2 images.
8. P10, Fig8, As shown in Figure 8, there are large areas of very deep water on the image, and these deep water areas are not very meaningful to invert with remote sensing and can be considered as background. The authors give inversion values for these background areas, so Figure 8 loses contrast and the bathymetric map is poorly schematized. It is recommended that the authors, with enough time, mask the background with a template and do not participate in the calculations to get a better presentation. It is up to the authors whether to modify it or not.
9. P11, Fig9, There are small patches of land in the islands of Figure 9, and the results are labeled as negative water depths. The authors should have masked the land with a template to avoid such erroneous results.
10. P12, table3, Is the table only a result of the South China Sea islands? This should be made clear in the title of the table.
11. P12, Fig 10 (b), Why the grid shape appears in fig10(b)?
12. P31, L356, How do you arrive at that conclusion? Give an example in the precision analysis table.
13. P14, Fig 12. The picture is too small to see clearly. Add units to the horizontal and vertical axes.
Comments for author File: 
Comments.pdf
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
This manuscript has compared various deep learning models and the SVR method for bathymetry prediction using optical remote sensing imagery to be compelling. The work of this study presents a certain degree of innovation. However, there are some points I would like you to consider for improving your manuscript. These points are aimed to further enhance the depth, clarity, and impact of your study. I look forward to seeing the revisions of your manuscript.
1 You've mentioned that the structure and depth of the convolutional network don't significantly affect the bathymetry prediction results. What might be the underlying reasons?
2 How do the deep learning models handle data with missing or incomplete information?
3 How does the size of the training dataset impact the performance of these deep learning models?
4 How does the computational cost and time of these deep learning models compare with traditional methods?
5 In your introduction, you mention the use of bathymetry for multiple applications. Could you please cite some key papers that have used bathymetry in these contexts to strengthen this argument?
6 Your methodology section refers to several deep learning models, including UNet++ and KFBNet. Could you provide more references to the original papers where these models were proposed and other significant works using these models?
7 In the method section, you explained how you preprocessed the data. Are there any references that you can cite which have used similar preprocessing steps for this type of data? This could help to give your methodology more context and credibility.
8 In Figure 10, there is a noticeable checkerboard artifacts in the results of the RefineNet. What might be the cause of this phenomenon?
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
The manuscript was much improved. This revised version has taken into account of my earlier comments and given a point to point reply in detail. I have no more comments and suggest publishing this paper.
