A Method for Inverting Shallow Sea Acoustic Parameters Based on the Backward Feedback Neural Network Model
, Zhiqiang Cui 3,4,†, Jia Liu 5,*, Shenghui Jiang 6,*, Xu Liu 1 and Jiahui Wang 3
Round 1
Reviewer 1 Report
1. In line 60, please clarify what is meant by "the results of the anti-discrimination results."
2. In line 66, please clarify what is meant by "The network has been improved to improve..."
3. In line 76, please clarify what is meant by "the quality of a model" in this context.
4. In the introduction, please emphasize the application of neural networks specifically to the problem under consideration.
5. In line 81, please clarify what is meant by "the anti-research work..."
6. The text as a whole is difficult to read, and contains many typos and capitalization errors.
7. In line 89, please provide an overview of the following content.
8. In line 136, please remove the undefined character.
9. It seems that subsection 2.1 does not refer to Section 2. Please clarify this.
10. Please explain the selection of your particular activation function.
11. Please specify the termination conditions mentioned in Figure 2.
12. Please provide a more detailed explanation of the application of formula (6).
13. The description of a network with a single layer and the selection of its parameters are too complex and difficult to understand. Please simplify this section.
14. Please clarify the purpose of Figure 3b.
15. In line 276, please clarify what is meant by "the coincidence degree of each group of data" and provide a more intuitive explanation.
16. Have you checked for the possibility of network overfitting?
17. It is not always clear whether the error is given and evaluated on the training sample or on the test sample. Please clarify this.
18. Please explain the nature of the areas with a large error in Figure 7b.
The text as a whole is difficult to read, and contains many typos and capitalization errors.
Author Response
In response to your review comments, I have made some revisions, and the following are my replies to these comments one by one:
1.There are some mistakes in the statement here, which has now been revised to “In recent years, due to the superior performance of neural network algorithms in data processing, scholars have been exploring the use of neural network models as an alternative to iterative optimization methods for inversion purposes. ”
2.Now it has been modified to “improved to modify”
3.Now it has been modified to “the predictive quality of the model.”
4.Changes have been made in the introduction.
5. Now it has been modified to “However, little work has been reported on the application of BP neural networks to inverse research work on shallow sea acoustic parameters, and it is worthwhile to try”
6.Changes have been made to the text.
7.Now it has been modified to“
The main content of this paper consists of the following four parts:: the first part is an overview of the sound parameters based on the acoustic pressure field; part 2 introduces the BP neural network-based ground sound parameters countermeasure method studied; In the third part, the application effect and performance of the BP neural network model are analyzed through simulation and scaling experimental data; The last is the conclusion. This paper will explore inversion methods based on this logical structure‘.
8.Undefined symbols have been removed.
9.The content of Section 2.1 is the forward modeling model of shallow sea sound field. The fast field method is used to simulate the submarine sound field, and the physical relationship model between ground sound parameters and sound pressure is established, which is the source of training data of BP neural network model.
10.The article has been added about activation function selection.
11.The picture has been modified to achieve the target accuracy.
Formula (6) is used to calculate the number of nodes in each layer, which has been detailed in the paper
13. Made some simplifications.
14. Figure 3 (b) shows the linear regression analysis of BP neural network inversion results.
15. The coincidence degree of each set of data refers to the degree of coincidence between the predicted value and the actual value, indicating the accuracy of the predicted result. This section has been added to the article
16. In the process of the experiment, because only <> parameters were retrieved, the inversion efficiency was high, so there was basically no overfitting.
17. Figure 4 (a) error decline refers to the evaluation results given by the training sample during the training process, and Figure 5 (a) refers to the error of the test sample.
18. The measured values are shown in the figure. 7 (b) is the experimental data of pool shrinkage. Because it is difficult for pools to fully mimic the conditions of the semi-infinite seafloor, there are some large errors.
As for the comments you gave, I have made a reply and revised the article accordingly. I will upload the revised article as an attachment for your review.
Author Response File: 
Author Response.docx
Reviewer 2 Report
Authors employed BP and SA algorithm for aping the Inverting Shallow Sea Acoustic Parameters. Here are my comments for authors:
1. In literature there are many advanced networks available like CNN, RNN, GA-NN, TLBO-ANN and So on. The above methods are proven efficient in performing better performance than BP ANN, as BP algorithm possess greater probability of getting trapped at local minima. Why the author used BP ANN need to be strongly justified.
2. In addition, SA, GA, PSO, TLBO, JAYA, and So on will be used to train the neural network to avoid local minima solutions. Authors need to strongly justify, why SA performed comparatively inferior performance. Further, why authors not used the aforementioned algorithms for better results.
3. The number of training and testing data points, whether batch mode or incremental mode of training, bias value, justification for selection of activation values, number of hidden layers and hidden neurons, and they are optimized is not clearly defined in their work.
4. Network stopping criterion is not properly defined.
5. Justification for the reason for selection of only five acoustic parameters for your study.
6. Conclusion seen to be lengthy, authors need to revise align to the objectives.
7. Authors need to define the novelty of the work.
Authors need to revise manuscript with native english speaker. As there are many sentences seen to be confusing and often found difficult to readers.
Author Response
In response to your review comments, I have made some revisions, and the following are my replies to these comments one by one:
- Although BP neural network has the probability of falling into the local minimum, in the process of inversion of acoustic parameters of single-layer seabed in shallow sea using BP neural network algorithm in this paper, BP neural network performs well, the maximum error of inversion is 0.065, and no situation of falling into the local minimum occurs. Therefore, BP neural network is adopted in this paper to inversion acoustic parameters in shallow sea. In the process of multi-layer submarine acoustic parameter inversion in the future, we also consider using GA-NN and other methods to test, in order to avoid the possible risk of falling into the local minimum.
- For inversion results, the SA algorithm does not show poor performance. The relative poor performance of the SA algorithm is due to the fact that the SA algorithm has more iterations than the BP neural network algorithm, and the real-time performance of the algorithm is poor. In addition, more iterations will increase the risk of the algorithm falling into the local minimum value.
- 2000 sets of training data and 200 sets of test data were obtained in batches of training and test data. The reasons for selecting the activation function are shown in lines 236-250 of the article and the number of hidden layers and hidden neurons in lines 223-232 of the article. All these problems have been modified in the article.
- The criteria for stopping the network to achieve the set accuracy has been modified in Flowchart 2 (b).
- In high-frequency submarine acoustics, the geophysical properties of sediments are divided into two categories, namely earth acoustic parameters and physical properties, among which earth acoustic properties mainly include the velocity and attenuation of P-wave and shear wave, and acoustic impedance, where acoustic impedance is the product of the density of sedimentary objects and sound velocity. Combined with previous studies, the acoustic velocity and attenuation of p-wave are discussed. The five parameters of shear wave velocity, attenuation and density can fully reflect the properties of submarine sound, so these five parameters are selected as the inversion targets.
- The modification of the conclusion has been completed.
- There are few literatures on the application of neural network in acoustic parameter inversion in shallow sea. This paper aims to apply backward feedback neural network in this aspect of parameter inversion, and also makes comparison and verification with traditional optimization algorithms, proving the feasibility and convenience of the method.
As for the review opinions you gave, I have made a reply and modified the article accordingly. I will upload the modified article as the attachment for your review.
Reviewer 3 Report
A shallow water acoustic parameters inversion method is suggested for the application. For this aim, the measured sound pressure field data is used in the study. Some suggestions and comments to the authors are presented below:
1. The flowchart of the suggested methodology should be given by more branches and in detail in Figure 2b. Thus, the readers can easily follow the application procedures.
2. Confidence intervals of the regressional line in Figure 3b can be presented to evaluate the application results.
3. Literature part is looking weak. Give new and last updated examples from literature about “Neural Network” as
(2023). A Neural-Network Based MPAS—Shallow Water Model and Its 4D-Var Data Assimilation System. Atmosphere, 14(1), 157.
(2022). Comparison of different ANN (FFBP, GRNN, RBF) algorithms and Multiple Linear Regression for daily streamflow prediction in Kocasu River, Turkey. Fresenius Environmental Bulletin, 31(5), 4699–4708.
4. The performance metrics part is weak in the paper. More metrics can be calculated to evaluate the application results additionally R-squared as NSE, RSR (Ratio of RMSE to the standard deviation of the observations), MAE, MSE, RMSE, etc. …
5. Some statistical properties as coefficient of variation, confidence intervals, distribution characteristics, min and median, etc. of used data should be given in a table.
6. Conclusions part can be improved in the paper. Here is presented in a general concept.
7. Is the used methodology in the paper valid for all areas or is there any limitation or classification for the application?
8. The resolution of the figures are very low.
9. The variables in Tables 1 & 4 should be explained in the paragraphs.
10. Discussions part is missing in the paper.
Check the tenses of the sentences. There are present and present perfect tenses in a paragraph. See the lines 42-50
There are some crucial errors.
Keywords should be ordered A to Z. Two more keywords as “Shallow Sea” & “Sound Field” can be added to keywords.
One sentence can’t be a paragraph. See the lines 225-227 …
Author Response
In response to your review comments, I have made some revisions, and the following are my replies to these comments one by one:
1.Figure 2 (b) has added more branches to help the reader understand the process.
2.The content related to confidence intervals has been added in the paper, including confidence bands with 95% confidence and related explanations.
3.new and last updated examples from literature about "Neural networks" have been given.
4.MAE has been added as an indicator to evaluate application results, as shown in Table 2.
5.Related properties such as confidence intervals have been added.
6.The conclusion has been changed.
7.The experimental background of this paper is mainly aimed at the inversion of ground sound parameters of the semi-infinite seabed in the shallow sea. In the shallow sea environment, the water depth is relatively shallow and the horizontal detection distance is 1.5 kilometers, so the seabed is approximately regarded as the horizontal seabed. On this basis, the case of the seabed being multi-layered will be discussed later.
8.As for the resolution of the picture, we have checked that the resolution of the picture is above 300dpi. Considering the possible image loss caused by word, we have uploaded all the pictures in the paper as compressed packages separately.
9.The variables in Table 1 and Table 4 have been explained in the paragraphs.
10.The results have been discussed in the paper.
As for the review opinions you gave, I have made a reply and modified the article accordingly.
Round 2
Reviewer 1 Report
3. Please specify how you define the predictive quality in your study.
12. Please explain why you use this formula and provide information on what alfa is and how it is selected.
14. Please provide an explanation for why you performed this analysis in your study.
Author Response
As for the questions you raised in the paper, I have made corresponding modifications, and the following is my reply to these questions:
3.As for the definition of prediction quality, this paper mainly evaluates three indicators. First, EMSE in Figure 4 (a) is the root-mean-square error. In the process of neural network training, the training will be terminated only when the error reaches the preset accuracy, so that the prediction error can be guaranteed to be within the preset accuracy. The fluctuation of all parameter errors is not obvious and all are below 0.1. Finally, MAE in Table 2 is the average absolute error, and the average absolute error of five parameters is within a reasonable range. These three indexes all reflect the prediction accuracy of the model. In addition, the accuracy of prediction can also be demonstrated through simulation and measurement, as well as comparison between the results of BP neural network algorithm and SA algorithm.
12.For the selection of formulas, formula (6) is an empirical formula commonly used in BP neural networks to determine the number of nodes in each layer. Alpha values are usually determined through experimentation and tuning. In this paper, different alpha values are selected for model testing, and the performance of BP neural network under different alpha values is evaluated in validation sets or cross-validation. According to the change of performance index, the optimal performance alfa of the model is determined to be -15.
14.There are two analyses in the paper, so the two analyses mentioned in the paper are provided here to put forward corresponding explanations. First, the regression analysis shown in fig 4 shows a high linear correlation between the predicted value and the actual value. It shows that the BP neural network model used in this paper has a good accuracy in the estimation of semi-infinite seabed acoustic parameters in shallow sea. The MAE value of each parameter in Table 2 is the average absolute error of each parameter. By calculating the average absolute error, the prediction accuracy of the BP neural network model used in this paper can be evaluated. A smaller MAE value indicates a higher prediction accuracy of the model, and it can be seen that the MAE values of the five parameters are all smaller. The BP neural network algorithm used in this paper has a good performance in the prediction and inversion of semi-infinite seabed acoustic parameters in shallow sea.
Reviewer 2 Report
There are two comments need to be addressed by authors:
1. Separate nomenclatures section could help any novice reader for better understanding.
2. Scope for future work can be high lightened at the end of conclusion.
Author Response
As for the questions you raised in the paper, I have made corresponding modifications, and the following is my reply to these questions:
1.A Separate nomenclatures section has been added to the paper
2.A highly specific scope of future work has been added at the end of the conclusion
Reviewer 3 Report
I suggest accepting the manuscript. The authors carefully revised the paper by answering each comment from the last round. There is only a small correction in the paper. The references and citations must be checked in the paper. Some of them are improper. See [16]: Halil, I. B. -> Burgan, H. I.; [6, 8, 10 & 31]: Year should be in bold character, [32]: Journal name should be in italic form …
Author Response
As for the errors you raised about the format of references, I have completed the modification
