Band-Optimized Bidirectional LSTM Deep Learning Model for Bathymetry Inversion
Round 1
Reviewer 1 Report
1. One of the more frequently addressed scientific problems is the calculation of shallow water bathymetry. Photogrammetric data is relatively often used for this task. This paper presents another modified method of detecting shallow water bathymetry on satellite images. Very important is results compare to IHO S-44 standard.
2. General remarks
a. Too many abbreviations make it difficult to follow the content of the article. Each abbreviation should be expanded the first time it appears. Not all readers need to know all abbreviations. Especially in the title of the article, in the abstract and conclusions.
b. Please use the language of a scientific research report without personal references: like “we”, “our”, several times used in the article.
c. Please explain more detailed the source of the data used in the experiment also about weather conditions.
d. Some information about photogrammetric data used in the process of shallow water bathymetry calculation could be found in doi.org/10.3390/en14175370, doi.org/10.3390/s22051844 and doi.org/10.3390/rs14164075.
e. However, the article is very well written should be carefully edited. Some remarks included below.
3. Specific remarks
a. Line 378 unneeded space and coma.
b. IHO's S-44 standard should be referenced and cited.
Author Response
Reviewer 1
General remarks
Comment 1: a. Too many abbreviations make it difficult to follow the content of the article. Each abbreviation should be expanded the first time it appears. Not all readers need to know all abbreviations. Especially in the title of the article, in the abstract and conclusions.
Response: Thanks for your kind suggestion. It has been revised. The full name of any uncommonly known abbreviation is given when it appears in the first time.
Comment 2: b. Please use the language of a scientific research report without personal references: like “we”, “our”, several times used in the article.
Response: Thanks for your kind suggestion. It has been revised to the best of our efforts.
Comment 3: c. Please explain more detailed the source of the data used in the experiment also about weather conditions.
Response: Thanks for your kind suggestion. It has been revised. The data sources of PRISMA, Sentinel-2 and ICESat-2 in the new version have provided the link addresses to their documentations, the time when the images were captured. All data images chosen were cloud free in the study area.
Comment 4: d. Some information about photogrammetric data used in the process of shallow water bathymetry calculation could be found in doi.org/10.3390/en14175370, doi.org/10.3390/s22051844 and doi.org/10.3390/rs14164075.
Response: Thanks for your kind suggestion. The useful references were cited in the new edition, lines 42 to 43, references 17 to 19.
Comment 5: e. However, the article is very well written should be carefully edited. Some remarks included below.
Response: Thanks for your kind suggestion, which are very important for us to improve our writing ability.
Specific remarks
Comment 6: a. Line 378 unneeded space and coma.
Response: Thanks for your kind suggestion. It has been revised to line 350 in the new version.
Comment 7: b. IHO's S-44 standard should be referenced and cited.
Response: Thanks for your kind suggestion. This study has referred to the S-57 standard and compared it against the S-44 standard, as shown in Table 1 below. This study is slightly lower than the S-44 standard. Although the LiDAR points of ICESat-2 are not less than 5% (the points are 4142, the prediction points are 42510), their distributions are strip-like and not evenly distributed throughout the survey area, so the sounding accuracy will be reduced. Although the multibeam data are uniformly distributed throughout the survey area, but there are only 750 measured points, which cannot meet the standard. Therefore, it is a pity that this study does not meet the S-44 standard and will not be cited in the article for the time being, we will try to improve in future.
Table 1 Comparison of S-57 standard and S-44 standard
|
IHO standard |
Depth Range (m) |
Required Accuracy (±m) |
RMSE (m) |
|
S-44 |
5（Order 1b） |
0.5 |
0.60 |
|
20（Order 2） |
1.1 |
1.28 |
|
|
S-57 |
0-10 |
0.6 |
0.66 |
|
10-30 |
1.6 |
1.41 |
Please let us know if there are any further issues regarding the new version, we will try our best to enhance.
Best Regards
Xiaotao Xi, Ming Chen, Yingxi Wang, Hua Yang
Author Response File: 
Author Response.docx
Reviewer 2 Report
The authors in this study introduces the Bi-directional Long Short-Term Memory (BiLSTM) for bathymetry inversion along with a band optimization strategy, named to be the Band-optimized BiLSTM (BoBiLSTM) model. By applying this improved model and other existing models to both hyperspectral satellite images (PRISMA) and multispectral satellite images (Sentinel-2) data. The experimental results demonstrate the robustness of the BoBiLSTM model over other compared models, thus showing certain potentials in bathymetry for operational use. The manuscript provides a necessary analysis and validation with detailed descriptions, although the contents could be better organized (see the following comments). Therefore, the manuscript is recommended for publication after the following revisions are made.
1) Page 1: The abstract is too long. Suggest to cut the length of the abstract to half of its current length by focusing what this study has done and contributed.
2) ‘SDB’ are explained twice (see line 51 in page 2 and line 141 in page 3). Delete the 2nd one.
3) Delete a duplicated ‘.’ in line 129 in page 3.
4) The contents in section 1 ‘Introduction’ are not concisely organized. Here are a few suggestions for the authors to consider.
a. Cut the length of the first three paragraphs in this section;
b. Move ‘Therefore, this paper attempts… the performance of bathymetry inversion’ (see line 113 to 118 in page 3) to the last paragraph in this section to address what this study is going to do (a more concise version is recommended).
c. Add a short paragraph to structure rest of this manuscript
5) Section 2 is not well organized since the existing methods are mixed with the new method. In order to highlight the new method, this section is suggested to re-organized to focus on the existing method. In the meanwhile, a new section is added to introduce the new method. Those sections are suggested to be re-organized in the following way.
2. Analysis Area, Existing Methods and Datasets
2.1. Analysis Area (i.e., 2.1 in the current version)
2.2. Existing Methods
A briefing introduction to the Stumpf Model, BiLSTM and other used methods in this sub-section, i.e., 2.2.1, 2.2.2. and 2.2.4 in the current version. Their details are given in an appendix.
2.2. Datasets (only limit to 2.2.1 and 2.2.2 in the current version)
3. BoBiLSTM Model, Training Data, and Evaluation Method
3.1. BoBiLSTM Model
3.2. Training Data (i.e., 2.2.3 and 2.2.4 in the current version)
3.3. Evaluation Method (i.e., 2.2.5 in the current version)
After the manuscript is re-organized according to the above suggestions, please revise the section number properly.
6) Some of the contents in Section3 and Section 4 are duplicate. It would be better to have a concise version for sub-sections from 4.1 to 4.3 by removing duplicated contents. Then, they can be moved to Section 3. In fact, many of the paragraphs in this manuscript are very long, so there is a space to improve.
7) For Fig. 5 in page 15 and Fig. 7 in page 19, it is better to plot images in (b) to (k) as their differences to the reference, which can provide a visual picture how the retrieved bathymetry values in the map deviate from those in the reference. The explanations should be much easier and shorter.
Author Response
Reviewer 2
Comment 1: 1) Page 1: The abstract is too long. Suggest to cut the length of the abstract to half of its current length by focusing what this study has done and contributed.
Response: Thanks for your kind suggestion. It has been revised accordingly. The journal requires an abstract of about 300 words, the previous abstract has 323 words, and the revised abstract has been shortened to 280 words. The first three sentences of the abstract lead to insufficient research on the spectral information of multispectral and hyperspectral satellite images. This is followed by a succinct statement about the data used by the new method BoBiLSTM, the methodology employed and the study area, and a description of the results achieved by this study. Finally, the contribution of the new method studied to water depth inversion is illustrated.
Comment 2: 2) ‘SDB’ are explained twice (see line 51 in page 2 and line 141 in page 3). Delete the 2nd one.
Response: Thanks for your kind suggestion. It has been revised.
Comment 3: 3) Delete a duplicated ‘.’ in line 129 in page 3.
Response: Thanks for your kind suggestion. It has been revised.
Comment 4: 4) The contents in section 1 ‘Introduction’ are not concisely organized. Here are a few suggestions for the authors to consider.
a. Cut the length of the first three paragraphs in this section;
Response: Thanks for your kind suggestion. It has been revised. The 998 words were reduced to 696 words, and the content of the cited references was further refined and compressed.
b. Move ‘Therefore, this paper attempts… the performance of bathymetry inversion’ (see line 113 to 118 in page 3) to the last paragraph in this section to address what this study is going to do (a more concise version is recommended).
Response: Thanks for your kind suggestion. It has been revised. In line 92 on page 2 of the new version, the method of this study and the main contributions of this article are highlighted.
c. Add a short paragraph to structure rest of this manuscript
Response: Thanks for your kind suggestion. It has been revised. On page 3, line 116 of the new edition.
Comment 5: 5) Section 2 is not well organized since the existing methods are mixed with the new method. In order to highlight the new method, this section is suggested to re-organized to focus on the existing method. In the meanwhile, a new section is added to introduce the new method. Those sections are suggested to be re-organized in the following way.
2. Analysis Area, Existing Methods and Datasets
2.1. Analysis Area (i.e., 2.1 in the current version)
2.2. Existing Methods
Response: Thanks for your kind suggestion. The main structure of the journal format is required to be "Introduction-Materials and Methods-Results-Discussion-Conclusions", so only the internal sub-sections have been modified on the premise to keep the main structure as required.
A briefing introduction to the Stumpf Model, BiLSTM and other used methods in this sub-section, i.e., 2.2.1, 2.2.2. and 2.2.4 in the current version. Their details are given in an appendix.
Response: Thanks for your kind suggestion. It has been revised. In section 2.2 of the new version, the contents of Stumpf model, DLSTM model, CNN-LSTM model and BiLSTM model have been compressed and refined, and details of those models such as formulas and algorithms can be found in the citations, so they will not be placed in the appendix for the time being. The BiLSTM model is the base model of the proposed model BoBiLSTM, so it would be better to put specific algorithms and formulas in the introduction of the new model.
2.2. Datasets (only limit to 2.2.1 and 2.2.2 in the current version)
Response: Thanks for your kind suggestion. This secondary heading has been revised to 2.3. Datasets. Every model used in this article will use all the data. For example, the satellite images used by the Stumpf model are PRISMA and Sentinel-2, and the sources of bathymetry data are ICESat-2 and Multibeam Scan. The finally generated bathymetry inversion map is compared with the Independent Reference Bathymetry Map. Therefore put all the data in this section. The specific second-level headings and third-level headings are as follows:
2.3. Datasets
2.3.1. PRISMA – Hyperspectral Satellite Images
2.3.2. Sentinel-2 – Multispectral Satellite Images
2.3.3. ICESat-2 Data – Training data
2.3.4. Multibeam Scan Data – Training data
2.3.5. Independent Reference Bathymetry Map – Validation data
3. BoBiLSTM Model, Training Data, and Evaluation Method
3.1. BoBiLSTM Model
Response: Thanks for your kind suggestion. This title is changed to the second-level title and included in the main frame Materials and Methods, so the specific modification is as follows:
2.4. BoBiLSTM Model
2.4.1. BiLSTM
2.4.2. Band optimization method
2.4.3. Bathymetry inversion framework
The third sub-level titles in this chapter can highlight the neural network algorithm, band optimization algorithm and bathymetry inversion framework used by the new method BoBiLSTM model. Since BoBiLSTM uses the BiLSTM model, the description of this neural network algorithm is expanded in 2.4.1.
3.2. Training Data (i.e., 2.2.3 and 2.2.4 in the current version)
Response: Thanks for your kind suggestion. Because each model uses all the data, not only the BoBiLSTM model uses the data in the training set. Therefore, the data set is all placed in Section 2.3. of the new version.
3.3. Evaluation Method (i.e., 2.2.5 in the current version)
Response: Thanks for your kind suggestion. The performance of all models is evaluated using R2 and RMSE. Not only the BoBiLSTM model uses this evaluation method, so the title is changed to 2.5. Evaluation Method
After the manuscript is re-organized according to the above suggestions, please revise the section number properly.
Response: Thanks for your kind suggestion and the section number has been revised. The last chapter directory is changed to:
1. Introduction
2. Materials and Methods
2.1. Analysis Area
2.2. Existing Methods
2.3. Datasets
2.3.1. PRISMA – Hyperspectral Satellite Images
2.3.2. Sentinel-2 – Multispectral Satellite Images
2.3.3. ICESat-2 Data – Training data
2.3.4. Multibeam Scan Data – Training data
2.3.5. Independent Reference Bathymetry Map – Validation data
2.4. BoBiLSTM Model
2.4.1. BiLSTM
2.4.2. Band optimization method
2.4.3. Bathymetry inversion framework
2.5. Evaluation Method
3. Results
3.1. Bathymetry Inversion Using ICESat-2 Data
3.1.1. Bathymetry of ICESat-2 Data
3.1.2. Bathymetry Inversion Using ICESat-2 Data
3.2. Bathymetry inversion using Multibeam scan data
4. Discussions
5. Conclusions
Comment 6: 6) Some of the contents in Section3 and Section 4 are duplicate. It would be better to have a concise version for sub-sections from 4.1 to 4.3 by removing duplicated contents. Then, they can be moved to Section 3. In fact, many of the paragraphs in this manuscript are very long, so there is a space to improve.
Response: Thanks for your kind suggestion. It has been revised. The contents of Section 4 regarding the figures and tables describing Section 3 have been revised into Section 3. The content of the fourth section of the new version only retains three small paragraphs, mainly for the performance comparison of different models, the comparison of the bathymetry inversion performance of hyperspectral and multispectral images, and whether the generated bathymetric map meets the IHO standard. Other longer paragraphs have been broken down into several smaller paragraphs.
Comment 7: 7) For Fig. 5 in page 15 and Fig. 7 in page 19, it is better to plot images in (b) to (k) as their differences to the reference, which can provide a visual picture how the retrieved bathymetry values in the map deviate from those in the reference. The explanations should be much easier and shorter.
Response: Thanks for your kind suggestion, hope the following explanations could meet you suggestion.
(1) The water depth range of the independent reference bathymetric map is basically in the range of 0 to 23 meters underwater. In order to compare with the multibeam bathymetric data, when actually using different models for water depth inversion, it will refer to the range of multi-beam scanning for drawing. This allows the study of the differences and accuracy of ICESat-2 bathymetry in deep water areas in PRISMA and Sentinel-2 imagery. Figure 1 is the water depth inversion map (left) and residual map (right) of different models. It can be seen that the inversion advantages of different deep learning models in the water depth range of 0 to 20 meters are not very different. From Table 7 of the article (line 571 of the new version), it can be seen that the RMSE difference of ICESat-2 depth sounding in different images is not very large.
Figure 1. Bathymetry inversion map and residual map generated by different models using ICESat-2. The left side is the bathymetry inversion map of different models, and the right side is the residual map of the bathymetry inversion map of different models and the reference bathymetry map
(2) The data of multibeam scanning is theoretically in the range of 0 to several thousand meters underwater. However, the reference bathymetry map of multibeam scanning has no value in the extremely shallow water area (as shown in Figure 2 below, indicated by the red rectangle). If only the difference between the sounding value and the reference value is drawn, the generated residual map will have missing data (as shown in the residual map on the right side of Figure 3). Therefore, in order to reflect the bathymetric advantages of the model in extremely shallow waters, retaining the bathymetric maps of different models can more intuitively allow readers to understand the different deep learning models compared in this article. On the other hand, the deep learning models DLSTM, CNN-LSTM, and BiLSTM have overfitting in both PRISMA image and Sentinel-2 image in multi-beam scanning bathymetry, and the degree of overfitting cannot be seen from the residual map. The over-fitting situation can be clearly observed with the bathymetric map.
Figure 2. Multibeam scanning reference bathymetry
Figure 3. Bathymetry inversion map and residual map generated by different models using multibeam scans. The left side is the bathymetry inversion map of different models, and the right side is the residual map of the bathymetry inversion map of different models and the reference bathymetry map
(3) Taking into account the revised opinions of other reviewers, the range of water depth can be clearly seen by using isobaths (as shown in Figure 4). Therefore, both of these two figures use the method of drawing isobaths.
Figure 4. Bathymetry inversion maps of different models. The bathymetry inversion map of ICESat-2 is on the left, and the bathymetry inversion map of multibeam scanning is on the right.
Please let us know if there are any further issues regarding the new version, we will try our best to enhance.
Best Regards
Xiaotao Xi, Ming Chen, Yingxi Wang, Hua Yang
Author Response File: Author Response.docx
Reviewer 3 Report
The authors submitted a well written and an interesting manuscript dealing with shallow water bathymetry mapping using deep learning algorithms and hyperspectral and multispectral imageries. The manuscript contains sufficient scientific information, and the results are well presented. However, before the manuscript could be considered for publication, the authors should revise it, and provide sufficient explanation on some of the raised comments and suggestions presented below.
Line 184: Please provide the references of recent studies conducted using Sentinel 2 and PRIMA satellites images and containing information on the numbers provided in the Table 1. This would help emphasize the argument presented on lines 71-72.
Lines 247-250: Please check and rephrase this section *Additionally, this empirical model takes into account the change in bottom albedo due to 248 bottom type heterogeneity, which affects both bands in a similar manner, while directly 249 observing changes in depth* because *The log-band ratio method of Stumpf et al. for SDB mapping assumes that the area has a uniform bottom and a log-band ratio of water-leaving reflectance that decreases linearly with water depth*
Line 514: it could be very helpful to include the iso-bathy contour lines to the Figure 7, which could indicate the variation of water depth.
Lines 41-43: Please add, among other disadvantages, that it has been shown that shallow water bathymetry mapping using Airborne Lidar Bathymetry is so expensive. Please refer to (1) MUZIRAFUTI, A.; Crupi, A.; Lanza, S.; Barreca, G.; Randazzo, G. Shallow water bathymetry by satellite image: A case study on the coast of San Vito Lo Capo Peninsula, Northwestern Sicily, Italy. In Proceedings of the IMEKO TC-19 International Workshop on Metrology for the Sea, Genoa, Italy, 3–5 October 2019; (2) Zhang, X.; Chen, Y.; Le, Y.; Zhang, D.; Yan, Q.; Dong, Y.; Han, W.; Wang, L. Nearshore Bathymetry Based on ICESat-2 and Multispectral Images: Comparison between Sentinel-2, Landsat-8, and Testing Gaofen-2. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2022, 15, 2449–2462.
Lines 378-379: The authors should provide sufficient information on how the Bathymetry inversion framework presented on the Figure 3 was implemented. For each step, they should provide the software or the code they used to perform it. This would help readers get a sense of how such methodology could be replicated because equations and internal structures presented are not sufficient.
Minor editing of English language required
Author Response
Reviewer 3
Comment 1: Line 184: Please provide the references of recent studies conducted using Sentinel 2 and PRIMA satellites images and containing information on the numbers provided in the Table 1. This would help emphasize the argument presented on lines 71-72.
Response: Thanks for your kind suggestion. It has been revised. The references have been cited in the new version lines 169-171 and 186-188. Table 1 is based on the official user documentation provided by Sentinel 2 and PRISMA, which provide specific band information for images. The provided numerical data can then be obtained using the band range formula: Wavelength = Central wavelength ± Bandwidth/2.
Comment 2: Lines 247-250: Please check and rephrase this section *Additionally, this empirical model takes into account the change in bottom albedo due to 248 bottom type heterogeneity, which affects both bands in a similar manner, while directly 249 observing changes in depth* because *The log-band ratio method of Stumpf et al. for SDB mapping assumes that the area has a uniform bottom and a log-band ratio of water-leaving reflectance that decreases linearly with water depth*
Response: Thanks for your kind suggestion. It has been revised in modified lines 135 to 137 in the new version.
Comment 3: Line 514: it could be very helpful to include the iso-bathy contour lines to the Figure 7, which could indicate the variation of water depth.
Response: Thanks for your kind suggestion. It has been revised.
Comment 4: Lines 41-43: Please add, among other disadvantages, that it has been shown that shallow water bathymetry mapping using Airborne Lidar Bathymetry is so expensive. Please refer to (1) MUZIRAFUTI, A.; Crupi, A.; Lanza, S.; Barreca, G.; Randazzo, G. Shallow water bathymetry by satellite image: A case study on the coast of San Vito Lo Capo Peninsula, Northwestern Sicily, Italy. In Proceedings of the IMEKO TC-19 International Workshop on Metrology for the Sea, Genoa, Italy, 3–5 October 2019; (2) Zhang, X.; Chen, Y.; Le, Y.; Zhang, D.; Yan, Q.; Dong, Y.; Han, W.; Wang, L. Nearshore Bathymetry Based on ICESat-2 and Multispectral Images: Comparison between Sentinel-2, Landsat-8, and Testing Gaofen-2. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2022, 15, 2449–2462.
Response: Thanks for your kind suggestion indeed they are very useful references. It has been revised. References have been cited in the new edition, line 39, references 11 to 12.
Comment 5: Lines 378-379: The authors should provide sufficient information on how the Bathymetry inversion framework presented on the Figure 3 was implemented. For each step, they should provide the software or the code they used to perform it. This would help readers get a sense of how such methodology could be replicated because equations and internal structures presented are not sufficient.
Response: Thanks for your kind suggestion. It has been revised. Figure 3 presents the key components of the new model structure; the corresponding code has been made available and uploaded to GitHub at: https://github.com/miaomiao-tech/Bathymetry_Inversion
Please let us know if there are any further issues regarding the new version, we will try our best to enhance.
Best Regards
Xiaotao Xi, Ming Chen, Yingxi Wang, Hua Yang
Author Response File: Author Response.docx
Round 2
Reviewer 3 Report
No further comments
Minor editing of English language required
Author Response
Thank you.
