Denoising Method for Seismic Co-Band Noise Based on a U-Net Network Combined with a Residual Dense Block
Round 1
Reviewer 1 Report
Dear Editor,
I consider my review only a "partial" review, because I am from the old school, modern techniques of machine learning, deep learning, artificial intelligence, I am only superficially familiar with them, nor am I interested in delving into these topics. So, I mainly checked the experimental part.
The article is worthy, but it needs revision, especially of the English language. Some sentences need to be revised because they are not clear, such as:
Abstract:
rows 13-14 Please explain better the phrase.
Introduction:
Rows 43-48 Not clear, phrase too long. Please break up with interline points
In the text:
formula (2) in row 141 is not clear or write:
or in compact form
Rows 200-201 The phrase isn’t clear, rewrite
I want also to report that the authors at rows 84-85 write: “In 2019, Zhu et al.[24] used the symmetric encoder-decoder structure of the U-Net model to denoise seismic signals”. I found that the two papers follow the same approach, and I don’t know if Zhu et al. used the approach descripted by the present authors. I suggest to check.
Other observations:
I suggest extending the title “2.2. Middle Layer of ARDU Model – the role of the Atrous Convolution”
rows 156-157 Have you some paper to show this result?
Formulas (4), (5), rows 223 – 226, I cannot understand if the variables are only in frequency domain or only in time domain, because in formula (5) if they are in frequency domain (5) is a convolutional product, different of a normal product in time domain.
Row 364 Please, better specify what kind of data: synthetic, real??? If synthetic, I also suggest a test with a real dataset.
In Fig. 13 The ARDU signal is a bit distorted, in fact it lacks high freq. Same thing in Fig. 14. As a U-net it is a bit low-pass, but this is obvious, generally high frequencies are missed.
I suggest to shortening a bit the paper, maybe erase one or two Figs among 9-12 and insert a real data example.
Given my unfamiliarity with deep learning and co., I do not plan to review the article again; I cannot have the proper critical sense for the topic part.
Major revision, anonymous.
Author Response
Applied Science - Reply on Manuscript ID applsci-2069342
4-Jan-2023
Dear Editor and Reviewers
Thank you very much for your insightful comments and suggestions. I have made corresponding modifications according to their suggestions. The red text and the Yellow highlighted text are my changes.
The following are my answers and modifications to the referee's questions and suggestions one by one.
Reviewer#1: Response to the questions one by one
Comments to the Author
I consider my review only a "partial" review, because I am from the old school, modern techniques of machine learning, deep learning, artificial intelligence, I am only superficially familiar with them, nor am I interested in delving into these topics. So, I mainly checked the experimental part.
1) The article is worthy, but it needs revision, especially of the English language. Some sentences need to be revised because they are not clear, such as:
rows 13-14 Please explain better the phrase.
Response 1:
Thank you for your guidance. I have revised it according to your comments. Please refer to the Yellow highlighted part on page 1 for details.
2) Rows 43-48 Not clear, phrase too long. Please break up with interline points.
Response 2:
Thank you for your guidance. I have revised it according to your comments. Please refer to the Yellow highlighted part on page 2 for details.
3) formula (2) in row 141 is not clear or write:
Response 3:
Thank you for your guidance. I have revised it according to your comments. Please refer to the Yellow highlighted part on page 3 for details.
4) Rows 200-201 The phrase isn’t clear, rewrite
Response 4:
Thank you for your guidance. I have revised it according to your comments. Please refer to the Yellow highlighted part on page 5 for details.
5) I want also to report that the authors at rows 84-85 write: “In 2019, Zhu et al.[24] used the symmetric encoder-decoder structure of the U-Net model to denoise seismic signals”. I found that the two papers follow the same approach, and I don’t know if Zhu et al. used the approach descripted by the present authors. I suggest to check.
Response 5:
Thank you for your guidance. The method used in this paper is to introduce the hollow convolution and residual modules on the basis of Zhu's 2019 paper. The atrous convolution and residual modules replace the step convolution and ordinary convolution in the model encoder in Zhu 2019 respectively, so as to enhance the feature extraction capability of the network, reduce waveform distortion and protect effective signals.
6) I suggest extending the title “2.2. Middle Layer of ARDU Model – the role of the Atrous Convolution”
Response 6:
Thank you for your guidance. I have revised it according to your comments. Please refer to the Yellow highlighted part on page 4 for details.
7) rows 156-157 Have you some paper to show this result?.
Response 7:
Thank you for your guidance. This parameter setting I refer to the literature "Ziye Yu, Risheng Chu, Minhan Sheng, Haichao Ma. Seismic phase picking by deep neural networks with both speed and accuracy [J]. Acta Seismologica Sinica, 2020, 42(3):15."
8) Formulas (4), (5), rows 223 – 226, I cannot understand if the variables are only in frequency domain or only in time domain, because in formula (5) if they are in frequency domain (5) is a convolutional product, different of a normal product in time domain.
Response 8:
Thank you for your guidance. The variables are in the time-frequency domain, and the model inputs seismic signals from the time domain into the data in the time-frequency domain through short-time Fourier transform. Time-frequency analysis includes both the time domain and the information in the frequency domain.
9) Row 364 Please, better specify what kind of data: synthetic, real??? If synthetic, I also suggest a test with a real dataset.
Response9:
Thank you for your guidance. The data is synthesized and the experiment has actual data added at the end of the experiment in accordance with your opinion.
10) I suggest to shortening a bit the paper, maybe erase one or two Figs among 9-12 and insert a real data example.
Response 10:
Thank you for your guidance. I have revised it according to your comments. Please refer to the Yellow highlighted part on page 10-15 for details.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Dear authors,
I find this work interesting from a technical point of view since the topic addresseed is of great importance for the seismic community and the construction of early warning systems. However, after a deep revision, there are some issues that are still not entirely clear to me and that I think they should try to explain better.
First, the improvement over the U-NET architecture is not clear to me. Following the text I haven't been able to find severe differences between your approach and U-NET, just the use of dilated convolutions. It seems that even the number of intermediate layers and the design of the residual block itself are the same. If not, please try to describe in more detail along the text. Secondly, and related to the first point, it would be interesting to go deeper into the description of the architecture. For me, it is not clear the use of cross entropy and the softmax layer during training, as well as the selection of 5 groups of features extractors (instead of 4 or 6, what is the difference in performance between them). It would be good to include at least a comparison between some models to give the reader an idea of the performances and training times. Third, it seems that the system converges very quickly after a couple of iterations. Why this rapid convergence? It might be thought that the authors have used Transfer Learning and have adapted their architecture by speeding up the training, but this is not the case. At least it is not cited in the text. So why is this rapid convergence? It should be widely described and discussed in the manuscript.
Finally, I would like to point out that some sections of the text have been difficult for me to follow. Although the technical part is correct, the description of the ideas in English should be revised and improved.
Author Response
Applied Science - Reply on Manuscript ID applsci-2069342
4-Jan-2023
Dear Editor and Reviewers
Thank you very much for your insightful comments and suggestions. I have made corresponding modifications according to their suggestions. The red text and the Yellow highlighted text are my changes.
The following are my answers and modifications to the referee's questions and suggestions one by one.
Reviewer#2: Response to the questions one by one
Comments to the Author
I find this work interesting from a technical point of view since the topic addresseed is of great importance for the seismic community and the construction of early warning systems.
Response:
Thank the reviewers for giving such a high evaluation of the author's work.
Major comments:
1) First, the improvement over the U-NET architecture is not clear to me. Following the text I haven't been able to find severe differences between your approach and U-NET, just the use of dilated convolutions. It seems that even the number of intermediate layers and the design of the residual block itself are the same. If not, please try to describe in more detail along the text.
Response1:
Thank you very much for your comments. The method of this paper is that on the basis of U-Net, atrous convolution and residual modules are introduced, which respectively replace maximum pooled subsampling and ordinary convolution in U-Net encoders, so as to enhance the feature extraction capability of the network, reduce waveform distortion and protect effective signals.
2)Secondly, and related to the first point, it would be interesting to go deeper into the description of the architecture. For me, it is not clear the use of cross entropy and the softmax layer during training, as well as the selection of 5 groups of features extractors (instead of 4 or 6, what is the difference in performance between them). It would be good to include at least a comparison between some models to give the reader an idea of the performances and training times.
Response2:
Thank you very much for your comments. Your comments are very good. The function of the cross entropy loss function is that the loss function adopts the cross entropy loss function to determine the closeness between the predicted value and the expected value. The softmax layer is used for the last layer of the model to generate normalized masks. The different number of convolutional layers does have different influences on the noise reduction effect. Generally speaking, the more layers, the better the noise reduction effect. As for the network degradation caused by the increase of layers, the residual module is adopted in this paper to solve this problem. However, the opinions of experts are very good. Due to time, it is a little hasty to supplement this part of experimental data, and the author will focus on this problem in the following. This will be the next research focus of the author.
3) Third, it seems that the system converges very quickly after a couple of iterations. Why this rapid convergence? It might be thought that the authors have used Transfer Learning and have adapted their architecture by speeding up the training, but this is not the case. At least it is not cited in the text. So why is this rapid convergence? It should be widely described and discussed in the manuscript.
Response3:
Thank you very much for your comments. The accelerated convergence of the model is due to the use of Adam accelerator in training. The function of Adam algorithm is to optimize the calculation process of loss function to accelerate the convergence of the model. I have revised it according to your comments. Please see the Yellow highlighted part on page 8 for details.
Author Response File: Author Response.pdf
Reviewer 3 Report
Dear Authors, it is almost acceptable for the publication at the first glance however, while looking through the references, the high ranked paper deserves to have more wider scope of referencing recognizable geophysical journals, here mostly just one is presented and other conference papers. Please take advantage to improve and add some from "Geophysical Prospecting" journal on denoising efforts from theory viewpoint as well as as independent reviewer I still recommend to add some references from the journal history where the publication is submitted (Applied Sciences). The highly specialized papers as it is this one still need some broader scope and I am sure Applied Sciences have devoted some special issues to geophysical problems solutions for the benefit of society
Author Response
Applied Science - Reply on Manuscript ID applsci-2069342
4-Jan-2023
Dear Editor and Reviewers
Thank you very much for your insightful comments and suggestions. I have made corresponding modifications according to their suggestions. The red text and the Yellow highlighted text are my changes.
The following are my answers and modifications to the referee's questions and suggestions one by one.
Reviewer#3: Response to the questions one by one
Comments to the Author
It is almost acceptable for the publication at the first glance.
Response:
Thank the reviewers for giving such a high evaluation of the author's work.
Major comments:
1) As independent reviewer, I still recommend to add some references from the journal history where the publication is submitted (Applied Sciences)
Response1:
Thank you very much for your comments. I have revised it according to your comments. Please see the Yellow highlighted part on page 15 for details.
Author Response File: 
Author Response.docx
Round 2
Reviewer 1 Report
Dear Editor,
Notwithstanding I wrote that I didn't want to see the second revision because I am not an expert of Deep Learning nor I'm not interested in delving into the subject now or ever, I took a look at the article. For these reasons I wrote that I checked only the results.
I asked for a revision, albeit not a thorough one, of the English and the authors had only retouched some phrases indicated by me.
The authors didn't improve the formula (2) and in my opinion is not clear in this way.
I advise you and the authors that the manuscript is almost the same in layout as the paper Zhu et al. (2019) even if the authors cite it. I don't speak of plagialism because the two methods are a bit different, but in my opinion is always better to avoid to copy the style of a previous paper.
Anyway the examples showed in the figure 7-14 are exhaustive and the conclusion too. The English in my opinion is to improve a bit. I suggest a revision.
Author Response
Applied Science - Reply on Manuscript ID applsci-2069342
12-Jan-2023
Dear Editor and Reviewer
Thank you very much for your insightful comments and suggestions. We have carefully considered the suggestions of Reviewer and make corresponding modifications. The red text and the Yellow highlighted text are my changes.
The following are my answers and modifications to the referee's questions and suggestions one by one.
Reviewer#1: Response to the questions one by one
Comments to the Author
Notwithstanding I wrote that I didn't want to see the second revision because I am not an expert of Deep Learning nor I'm not interested in delving into the subject now or ever, I took a look at the article. For these reasons I wrote that I checked only the results.
Response:
1) I asked for a revision, albeit not a thorough one, of the English and the authors had only retouched some phrases indicated by me.
Thank you very much for your suggestion. We are sorry for the problem in our old manuscript. We have revised the whole manuscript and carefully proof-read the manuscript to minimize typographical, grammatical, and bibliographical errors. In addition, we have invited a native English speaker to check the language. We hope the revised manuscript could be acceptable for you, please refer to the yellow part on page 1-3 and 7 for details.
2) The authors didn't improve the formula (2) and in my opinion is not clear in this way.
Thank you for your guidance. I have revised it according to your comments. Please refer to the Yellow highlighted part on page 4 for details.
3) I advise you and the authors that the manuscript is almost the same in layout as the paper Zhu et al. (2019) even if the authors cite it. I don't speak of plagialism because the two methods are a bit different, but in my opinion is always better to avoid to copy the style of a previous paper.
Thank you very much for your advice. Your comments are excellent. In terms of form, our form is not much different from zhu's paper. However, zhu's model structure uses step convolution to downsample seismic signal features. In this paper, void convolution and residual network are used to downsample seismic signal features in view of the difficulty in removing the same frequency band noise in seismic signals. Hollow convolution can be used to expand receptive field and residual network, which can enhance the network's advantage in feature extraction ability of seismic signals and enhance the network's ability to capture details such as polarization information in seismic signals. Compared with zhu's method, it can better distinguish seismic information and noise information in the same frequency band, reduce waveform distortion and protect effective signals. In the experiment, the comparison between FIG. 13(i) and (k) shows that the noise reduction effect of ADRU is better than that of U-Net model, with perfect details and low signal-to-noise ratio. As for the paper, we have made further modifications from the general form to the underlying connotation, which is more different from that paper, please refer to the yellow part on page 2 for details.
4) Anyway the examples showed in the figure 7-14 are exhaustive and the conclusion too. The English in my opinion is to improve a bit. I suggest a revision.
Thank you very much for your suggestion. We apologize for the poor language of our manuscript. We have now worked on both language and readability and have also involved native English speaker for language corrections. We really hope that the language level have been substantially improved.
Reviewer 2 Report
All my comments have been addresed
Author Response
Applied Science - Reply on Manuscript ID applsci-2069342
12-Jan-2023
Dear Editor and Reviewers
Thank you very much for your insightful comments and suggestions. I have made corresponding modifications according to their suggestions. The red text are my changes.
The following are my answers and modifications to the referee's questions and suggestions one by one.
Reviewer#1: Response to the questions one by one
Comments to the Author
All my comments have been addresed.
Response:
Thank the reviewers for giving such a high evaluation of the author's work.
