the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
An Improved Method Based on VGGNet for Refined Bathymetry from Satellite Altimetry: Reducing the Errors Effectively
Abstract. At present, only approximately 10 % of the global seafloor topography has been finely modeled, and the rest are either lacking in data or not accurate enough to meet practical requirements. On the one hand, satellite altimeter has the advantages of large-scale and real-time observation, thus is widely used in the measurement of bathymetry, the core of seafloor topography. However, there is often room for improvement in its precision. On the other hand, multibeam echosounder bathymetric data is highly precise but normally limited to a smaller coverage, which forms a complementary relationship with the bathymetry derived from satellite altimetry. To combine the advantages of satellite altimetry-derived and multibeam sonar-derived bathymetry, we apply deep learning, which is powerful in the field of digital image automation, to perform multibeam sonar-based bathymetry correction for satellite altimetry bathymetry data. Specifically, we modify and improve a pretrained VGGNet neural network model with a depth of 19 layers to train on three sets of bathymetry data from the West Pacific, Southern Ocean, and East Pacific, respectively. Experiments show that the correlation of bathymetry data before and after correction can reach a high level, with the performance of R2 being as high as 0.81 and the RMSE improved over 19 % compared with previous research. We then explore the relationship between R2 and water depth and conclude that it varies at different depths and thus the terrain specificity was a factor that affects the precision of correction. Finally, we use the difference of water depth before and after the correction to evaluate the correction results, and find that our method can improve by more than 17 % compared with previous research. The results show that using the deep learning VGGNet model can better perform the correction of the bathymetry derived from satellite altimetry, thus providing a method for accurate modeling of the seafloor topography.
- Preprint
(906 KB) - Metadata XML
-
Supplement
(153 KB) - BibTeX
- EndNote
Status: closed
-
AC1: 'Comment on gmd-2022-140', Xiaowen Luo, 10 Jul 2022
In Line 327, 331 and 347, we incorrectly left "0.5% of water depth" in the manuscript without actually showing the result in Table 3. This is because at the beginning of the experiment, we did calculate the result of 0.5% of water depth, but in later discussions, we decided to remove this criterion and leave only the 2% and 1% of water depths. We apologize for any misunderstandings that may have been caused by this minor error.
Citation: https://doi.org/10.5194/gmd-2022-140-AC1 -
RC1: 'Comment on gmd-2022-140', Anonymous Referee #1, 27 Aug 2022
Comments:
The authors refined the bathymetry from satellite altimetry based on an improved VGGNet. While this work is motivated, the manuscript suffers some technical problems.
Major comments:
As a start, it is unclear what the motivation for the choice of the machine learning approach used by VGGNet is. Why did the authors not choose another method for comparison in performance, given that they got poor results from this approach? The author's less innovative approach to VGGNet modification is not in line with the "An Improved Method" mentioned in the article.
The overall structure of the manuscript is unclear. The authors presented some details of the deep learning methods in both the introduction and methods. The multibeam-satellite data and satellite altimetry data are not well described, for example, the data coverage both spatially and temporally, and why they are used. Geographic data from different sources should be preprocessed to eliminate the effect of coordinate errors. What is the specific method of interpolation preprocessing in the article? Why was the method selected? The input and output data of the model are not reasonable. The satellite altimetry data should be used as an input parameter, with the true multibeam satellite bathymetry data being the expected result.The model's experimental testing component should be assigned to one of the three datasets, or to the non-training and validation part of the three regions.
The format of the reference is incorrect. In line 130, the authors cite (Charette et al., 2010), but I did not find any correlation between the authors' method and Charette's theory of the volume of the earth. In section 3.1, the original data source references and links are required. In line 278, the [26][29] method of marking the literature does not fulfill the criteria of uniform reference labeling. There were no similar findings between reference [26] and the approach described in this article. In line 333, what are the previous studies mentioned here? The literature and comparative data need to be annotated.
The manuscript is in need of editing by either a technical writer or, at least, someone with a technical background and conversant with native English. There are many instances of using English words that would better describe the situation than those used.
Â
Specific comment:
L29 and L278:Â Is it RMSE metric improvement or NRMSE?
L182: What is the meaning of the upper mark 40?
Fig.4, Fig.5 and Fig.6: These figures are too tiny to meet GMD's figure content guidelines requirements.
- AC2: 'Reply on RC1', Xiaowen Luo, 13 Sep 2022
-
RC2: 'Comment on gmd-2022-140', Anonymous Referee #2, 03 Sep 2022
General impression:
The paper uses the deep neural network model VGGNet-19 to perform multibeam sonar-based bathymetry correction for satellite altimetry bathymetry data. Some seemingly good experimental results are obtained on three sets of bathymetry data from the West Pacific, Southern Ocean, and East Pacific. However, the paper is not really clear in terms of motivation and method, and is not innovative enough. Furthermore, the evaluation is small and insufficient.
Overall comments:
The paper does not explain clearly why the VGGNet-19 model is used for bathymetry correction. There are many existing deep models, all of which have achieved promising results in the field of image processing. Why is the VGGNet-19 model effective in this work and outperforms other models? The authors should summarize the relevant deep models and combine the characteristics of the experimental data to further illustrate the advantages of VGGNet-19 in this work. Except for synthesizing the data from multibeam sonar-derived bathymetry and satellite altimetry-derived bathymetry to obtain a corrected version of the latter, there is not any innovation seen in the current manuscript. The authors claim to propose a novel optimization algorithm, but the adopted VGGNet-19 model does not have any improvements in the network structure except for different input and output parameters. The design of the loss function only defines a common distance function, lacking some valuable constraints that can reveal the hidden physical mechanism. The authors only simply apply the conventional VGGNet-19 architecture to bathymetry correction, making the contribution of this research to related fields insufficient.
The original shipborne multibeam sonar bathymetry data and satellite altimetry data used in the experiment need to be explained in detail. What are the spatial and temporal resolutions of these data? Which interpolation method is used for preprocessing? How to ensure data consistency after interpolation? Another weakness of the paper is the small experimental evaluation. As the authors mention that the proposed algorithm can achieve higher accuracy than the two similar studies in the literatures of [26] and [29], they should supplement the differences between the two studies and the proposed algorithm to further explain why the proposed algorithm can obtain better results. The simulation experiment lacks the hyperparametric analysis of the model. The authors also need to add more comparison methods for evaluation and to conduct a more detailed analysis and discussion of the experimental results.
The authors should supplement how to modify and improve a pretrained VGGNet neural network model with a depth of 19 layers in the abstract and reorganize the logical structure of the introduction. In addition, the authors should give a broader discussion on the advantages and disadvantages of the proposed solution and provide future research directions in the conclusion.
Figures and tables are not well presented. Some references are incorrectly formatted and cited, such as references [26] and [29]. English writing is poor. Sentences are somewhere complex and the proper meaning cannot be extracted. Please improve the readability with the help of natives.
Specific comments:
L30, L35, L144, L204, L332, and L361: What does the previous research that appears in these places refer specifically to? The authors should give a clear description or cite the source literature.
L282-284: Table 2 should not be displayed across pages.
L312-313: The title and picture of Fig. 5 should be displayed on the same page.
L482-483: This reference is not in alphabetical order.
There are many format problems in the text, which need to be carefully checked. Pages 4, 7, 8, and 11 have large blank areas at the bottom. It is suggested to re-typeset the manuscript.
Citation: https://doi.org/10.5194/gmd-2022-140-RC2 - AC3: 'Reply on RC2', Xiaowen Luo, 13 Sep 2022
Status: closed
-
AC1: 'Comment on gmd-2022-140', Xiaowen Luo, 10 Jul 2022
In Line 327, 331 and 347, we incorrectly left "0.5% of water depth" in the manuscript without actually showing the result in Table 3. This is because at the beginning of the experiment, we did calculate the result of 0.5% of water depth, but in later discussions, we decided to remove this criterion and leave only the 2% and 1% of water depths. We apologize for any misunderstandings that may have been caused by this minor error.
Citation: https://doi.org/10.5194/gmd-2022-140-AC1 -
RC1: 'Comment on gmd-2022-140', Anonymous Referee #1, 27 Aug 2022
Comments:
The authors refined the bathymetry from satellite altimetry based on an improved VGGNet. While this work is motivated, the manuscript suffers some technical problems.
Major comments:
As a start, it is unclear what the motivation for the choice of the machine learning approach used by VGGNet is. Why did the authors not choose another method for comparison in performance, given that they got poor results from this approach? The author's less innovative approach to VGGNet modification is not in line with the "An Improved Method" mentioned in the article.
The overall structure of the manuscript is unclear. The authors presented some details of the deep learning methods in both the introduction and methods. The multibeam-satellite data and satellite altimetry data are not well described, for example, the data coverage both spatially and temporally, and why they are used. Geographic data from different sources should be preprocessed to eliminate the effect of coordinate errors. What is the specific method of interpolation preprocessing in the article? Why was the method selected? The input and output data of the model are not reasonable. The satellite altimetry data should be used as an input parameter, with the true multibeam satellite bathymetry data being the expected result.The model's experimental testing component should be assigned to one of the three datasets, or to the non-training and validation part of the three regions.
The format of the reference is incorrect. In line 130, the authors cite (Charette et al., 2010), but I did not find any correlation between the authors' method and Charette's theory of the volume of the earth. In section 3.1, the original data source references and links are required. In line 278, the [26][29] method of marking the literature does not fulfill the criteria of uniform reference labeling. There were no similar findings between reference [26] and the approach described in this article. In line 333, what are the previous studies mentioned here? The literature and comparative data need to be annotated.
The manuscript is in need of editing by either a technical writer or, at least, someone with a technical background and conversant with native English. There are many instances of using English words that would better describe the situation than those used.
Â
Specific comment:
L29 and L278:Â Is it RMSE metric improvement or NRMSE?
L182: What is the meaning of the upper mark 40?
Fig.4, Fig.5 and Fig.6: These figures are too tiny to meet GMD's figure content guidelines requirements.
- AC2: 'Reply on RC1', Xiaowen Luo, 13 Sep 2022
-
RC2: 'Comment on gmd-2022-140', Anonymous Referee #2, 03 Sep 2022
General impression:
The paper uses the deep neural network model VGGNet-19 to perform multibeam sonar-based bathymetry correction for satellite altimetry bathymetry data. Some seemingly good experimental results are obtained on three sets of bathymetry data from the West Pacific, Southern Ocean, and East Pacific. However, the paper is not really clear in terms of motivation and method, and is not innovative enough. Furthermore, the evaluation is small and insufficient.
Overall comments:
The paper does not explain clearly why the VGGNet-19 model is used for bathymetry correction. There are many existing deep models, all of which have achieved promising results in the field of image processing. Why is the VGGNet-19 model effective in this work and outperforms other models? The authors should summarize the relevant deep models and combine the characteristics of the experimental data to further illustrate the advantages of VGGNet-19 in this work. Except for synthesizing the data from multibeam sonar-derived bathymetry and satellite altimetry-derived bathymetry to obtain a corrected version of the latter, there is not any innovation seen in the current manuscript. The authors claim to propose a novel optimization algorithm, but the adopted VGGNet-19 model does not have any improvements in the network structure except for different input and output parameters. The design of the loss function only defines a common distance function, lacking some valuable constraints that can reveal the hidden physical mechanism. The authors only simply apply the conventional VGGNet-19 architecture to bathymetry correction, making the contribution of this research to related fields insufficient.
The original shipborne multibeam sonar bathymetry data and satellite altimetry data used in the experiment need to be explained in detail. What are the spatial and temporal resolutions of these data? Which interpolation method is used for preprocessing? How to ensure data consistency after interpolation? Another weakness of the paper is the small experimental evaluation. As the authors mention that the proposed algorithm can achieve higher accuracy than the two similar studies in the literatures of [26] and [29], they should supplement the differences between the two studies and the proposed algorithm to further explain why the proposed algorithm can obtain better results. The simulation experiment lacks the hyperparametric analysis of the model. The authors also need to add more comparison methods for evaluation and to conduct a more detailed analysis and discussion of the experimental results.
The authors should supplement how to modify and improve a pretrained VGGNet neural network model with a depth of 19 layers in the abstract and reorganize the logical structure of the introduction. In addition, the authors should give a broader discussion on the advantages and disadvantages of the proposed solution and provide future research directions in the conclusion.
Figures and tables are not well presented. Some references are incorrectly formatted and cited, such as references [26] and [29]. English writing is poor. Sentences are somewhere complex and the proper meaning cannot be extracted. Please improve the readability with the help of natives.
Specific comments:
L30, L35, L144, L204, L332, and L361: What does the previous research that appears in these places refer specifically to? The authors should give a clear description or cite the source literature.
L282-284: Table 2 should not be displayed across pages.
L312-313: The title and picture of Fig. 5 should be displayed on the same page.
L482-483: This reference is not in alphabetical order.
There are many format problems in the text, which need to be carefully checked. Pages 4, 7, 8, and 11 have large blank areas at the bottom. It is suggested to re-typeset the manuscript.
Citation: https://doi.org/10.5194/gmd-2022-140-RC2 - AC3: 'Reply on RC2', Xiaowen Luo, 13 Sep 2022
Viewed
HTML | XML | Total | Supplement | BibTeX | EndNote | |
---|---|---|---|---|---|---|
634 | 203 | 56 | 893 | 98 | 39 | 41 |
- HTML: 634
- PDF: 203
- XML: 56
- Total: 893
- Supplement: 98
- BibTeX: 39
- EndNote: 41
Viewed (geographical distribution)
Country | # | Views | % |
---|
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1