Structure-Preserving Random Noise Attenuation Method for Seismic Data Based on a Flexible Attention CNN
Round 1
Reviewer 1 Report
The paper proposed a CNN architecture for denoising of seismic data, with a noise level map as an additional channel to enhance generalizability and introduced an attention layer with a mix loss function to protect the structure of seismic signals. The idea sounds interesting. However, the methodology, experimentation, and results could be detailed and well-structured. Results could be presented in a better way and validated quantitatively. Language and grammar can be improved. Redundancy in the text could be avoided. The quality of the paper could be significantly improved in order to make it publishable.
Here are some detailed comments which could be useful.
· Many statements (e.g., even in the very first paragraph) should have supporting citations or proofs.
· Could have introduced U-Net at least briefly when referring to it for the first time.
· Better to change the terms ‘in this letter’ to ‘in this article’.
· The proposed FACNN architecture needs a detailed description and an architecture diagram.
· Tables and figures could be placed near to the text where they are referred at first.
· The theory or basis behind the mathematical representation of attention gate (Eq. 1-4), denoising problem (Eq 5-7), and MS-SSIM (Eq. 9-11) should be given. If these are used from others’ work, they should be well cited. To avoid confusion with the ‘minus’ operator, better to use MS_SSIM instead of MS-SSIM.
· Dataset split of 90% training and rest for validation testing doesn’t sound suitable for generalization. Better to try with validation and a test set of at least 10% data.
· Instead of labeling the subfigures sequentially as a,b,c…,g,h, the figures could be better explained in the captions, maybe as top, bottom, and columns or otherwise could label as a, a’; b, b’; …
· It is not clear which real data was used and from where the data was collected.
· Also, it is not clear how the transfer learning method was used.
· The authors claimed the results from the proposed FACNN are better, possibly, looking at the figures. It’d be better to validate using appropriate quantitative metrics.
Author Response
Dear reviewer
Thank you for your helpful suggestions. We have uploaded a point-by-point response. Please see the attachment. For the main manuscript, we have uploaded an improved manuscript(PDF) with a highlighted manuscript(PDF) with the improved locations. If you cannot see the figures, please find the PDF version of the manuscript.
Wenda
Author Response File: 
Author Response.pdf
Reviewer 2 Report
This manuscript proposed a novel method based on flexible attention CNN for structure-preserving random noise attenuation of seismic data. The numerical verification was conducted to validate the performance of the proposed method, with satisfactory results. Overall, the topic of this research is interesting, and the manuscript was well organised and written. I suggest that it can be considered to be published in Remote Sensing, if the authors can well address the following comments.
1. The contribution and innovation of the manuscript should be clarified clearly in abstract and introduction.
2. Broaden and update the literature review in engineering applications of CNN or deep learning, such as image processing, noise attenuation, structural defect recognition, etc. E.g. Crack detection of concrete structures using deep convolutional neural networks optimized by enhanced chicken swarm algorithm. A novel deep learning-based method for damage identification of smart building structures.
3. How did the authors set the hyperparameters of the proposed model for optimal classification performance?
4. A parametric study is necessary to analyse the model architecture.
5. What are the evaluation metrics of the proposed method?
6. All the figures are not well presented. Please revise them
7. More future research should be included in the conclusion part. Since the authors just conducted the numerical validation in this research, I suggest some ideas about the experimental verification can be mentioned here as one of main future work.
Author Response
Dear reviewer
Thank you for your helpful suggestions. We have uploaded a point-by-point response. Please see the attachment. For the main manuscript, we have uploaded an improved manuscript(PDF) with a highlighted manuscript(PDF) with the improved locations. If you cannot see the figures, please find the PDF version of the manuscript.
Wenda
Author Response File: 
Author Response.docx
Reviewer 3 Report
In this paper, the authors proposed a noise depression method based on a flexible attention CNN.
1. The novelty of this paper is not clear.
2. The figures are invisible so that I can not give more comments.
-----------------
My comments can be listed down below:
1. How did the authors train the model? Did the author train the model only on the synthetic data?
2 . For the real seismic data without the clean version, how did the author evaluate the performance?
3. What noise model can the proposed method process ?
4. How did the author get the clean image for the evaluation and training?
Author Response
Dear reviewer
Thank you for your helpful suggestions. We have uploaded a point-by-point response. Please see the attachment. For the main manuscript, we have uploaded an improved manuscript(PDF) with a highlighted manuscript(PDF) with the improved locations.
If you cannot see the figures, please find the PDF version of the manuscript.
Wenda
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
The revised version tried to address most of the detailed comments given in my previous review. However, one of the major concerns regarding the details and structuring of methodology and experimentation still remains.
Before giving test results, the paper could first describe the methodology used in the experimentation, testing, and evaluation, next how experiments/tests were conducted, and then present the results. Quantitative results could be provided for the other denoising methods used in the paper. Discussion is important which is missing. The term ‘blind tests’ used in the conclusion section is nowhere used before to clarify what the authors meant by this. And, the conclusion section should focus on what conclusions could be drawn from the work rather than redundant texts describing it.
Author Response
Dear reviewer
Thank you for your helpful suggestions. We have uploaded a point-by-point response. Please see the attachment. For the main manuscript, we have uploaded an improved manuscript(PDF) with a highlighted manuscript(PDF) with the improved locations. If you cannot see the figures, please find the PDF version of the manuscript.
Wenda
Author Response File: Author Response.docx
Reviewer 2 Report
All the technical issues have been well addressed by the authors. I suggest that this revised version can be accepted for publication in Remote Sensing
Author Response
Dear reviewer
Thank you again for your helpful suggestions and I wish you all the best.
Wenda
