ROI-Binarized Hyperbolic Region Segmentation and Characterization by Using Deep Residual Convolutional Neural Network with Skip Connection for GPR Imaging

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have developed a deep residual convolutional neural network for the automatic segmentation of ROI in GPR images and the detection of hyperbolic sections. The paper is well written, the algorithm is successfully implemented and the results support the conclusions.

Minor language correction:

Change “convolution neural network” to “convolutional neural network” throughout the paper

Author Response

Thank you very much for your valuable suggestion. We have noted your proposed modification and have made the change from "convolution neural network" to "convolutional neural network" throughout the paper. We will continue to strive for accuracy and quality in our work, and we appreciate your support and feedback. Once again, thank you for your review and valuable input.

Reviewer 2 Report

Comments and Suggestions for Authors

The article introduces advanced deep learning techniques to the processing of Ground Penetrating Radar (GPR) images, significantly enhancing the field. By employing a Residual Convolutional Neural Network (Res-CNN) with skip connections for segmenting binarized regions of interest (ROI) in GPR images, it opens up new possibilities for practical applications in geophysics, civil engineering, and land surveying. The authors have conducted comprehensive testing of their model against other image segmentation methods, demonstrating its superiority in various test scenarios. This is supported by detailed experimental results and discussions that illuminate the effectiveness and limitations of the proposed method, helping to understand its efficiency.

However, the article has some shortcomings. The authors acknowledge that the training dataset primarily consists of simulations, which may limit the model's ability to generalize to real-world conditions. They also highlight the need for further research to enhance model generalization and adapt the method for use in more diverse media, such as anisotropic media. Additionally, while comparisons are made with existing segmentation techniques, the article lacks a direct comparison with the latest deep learning algorithms used in similar applications, which could provide a more robust validation of the proposed method's capabilities.

Is the article suitable for publication? The article demonstrates significant advancements in the field of GPR image processing using deep neural networks and could be a valuable addition to the scientific literature. However, before the article is published, it would be beneficial to conduct additional tests on actual field data to better assess the model's generalization and practical utility in more varied conditions. Furthermore, including comparisons with the latest techniques in the field could strengthen the article's standing as a significant and contemporary scientific contribution.

Detailed remarks: Discussion chapter: Overall, the discussion section is well-written but would greatly benefit from a deeper comparative analysis with cutting-edge methods and a more detailed outline of future research directions. Such enhancements would improve the section's effectiveness in situating the study within the broader context of advancements in GPR image processing.

Conclusion chapter: Overall, the conclusion section effectively summarizes the study's contributions but could be enhanced by offering more detailed insights into future research directions and broader applications.

The references are optimal. I have no comments about the figures."

Comments for author File: Comments.pdf

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

In the manuscript “ROI Binarized Hyperbolic Region Segmentation and Characterization by Using Deep Residual Convolution Neural Network with Skip Connection for GPR Imaging” submitted by Hua Zhang, Qianwei Dai, Deshan Feng, Xun Wang, and Bin Zhang, the authors investigate the deep residual convolution neural network (Res-CNN) with skip connection in the segmentation of region of interest (ROI) in GPR images. The algorithm provides a more efficient and accurate way than the old segmentation methods.

The authors first introduce the background of both GPR image segmentation and the application of deep learning algorithms. The overall data processing procedure including the structure of Res-CNN is proposed and explained along with the evaluation methods. After that, the authors presented the training methodology as well as the testing model. The final testing results are compared with the Otsu thresholding method and K-means clustering segmentation method. The authors also emphasize the importance of skip connection. Finally, the experiment results are presented and discussed.

Overall, this is a good manuscript with a clear structure and a decent scientific result. I am glad that the authors included the experiment parts and a discussion about the possible limitations and future directions. I would suggest that the authors provide some more discussions on the results part which I will mention below. The authors could also improve their figure quality.

This manuscript can be published with the consideration of the following remarks:

(1) In Figure 3, the authors introduce the random medium results. Please include all the labels of the figure and introduce them in the caption.

(2) In Figure 4, the authors present the loss value changing with epochs. There are many small spikes on the curve. What could be the reason for that? What are the criteria of convergence in this case? Would a better preprocessing help the convergence of the curve? Could the authors also provide a reasoning for choosing the proper η and γ?

(3) The authors implement two values of pipeline outer diameter in the training and testing process. Could the authors discuss the case of pipelines of larger or smaller outer diameters? Will the algorithm still have a good performance? What about some special cases, a large pipe above the small pipe, for example?

(4) In Figure 11, the authors claim that Figure 11b leaves many background noises. However, it seems to me that it has less noise than Figure 11d. My observation is that the right three pipes are easier for me to distinguish with the Otsu thresholding method. Could the authors explain more about the observations here?

(5) Here is a minor point: In Figures 8 and 12, the authors compare the results with and without the skip connections. It is good to see the comparison but one of the figures is the same as what we have seen in Figures 7 and 11. I would recommend avoiding repeating figures.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

1.

First, there is a large amount of similarity (>24%). This similarity should be reduced to less than 10%.

2.

Define GPR in the abstract.

3.

‘Need’ should be ‘needs’ in line 13.

4.

The abstract should be improved.

5. Throughout the paper, paragraphs are excessively long, making it difficult to read. Consider using shorter paragraphs

6.

Use ‘GPR image’ instead of ‘ground penetrating radar image’ in line 101.

7.

The text should come first, followed by figures and tables in Sections 2 and 3.

8.

The quality of Figure 2 should be improved.

9.

In Section 3, define titles( like (a) and (b) ) for Figure 3.

10.

Did you consider different frequencies?

11.

How many transmitting and receiving antennas were considered?

12.

Did you consider the presence of noise on data?

13.

It is suggested that the authors improve the structure of the article.

14.

English in this paper needs improvement, which can make this paper more like a journal paper.

15.

Please highlight the advancement of the method in Introduction, and improve the Conclusion section.

Comments on the Quality of English Language

English in this paper needs improvement.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 5 Report

Comments and Suggestions for Authors

The paper presents the residual convolutional neural network with a skip connection for GPR image binarization. The authors have yet to explicitly highlight the specific research gap and provide innovative means to present the results of the proposed segmentation model. Below are the reviewer's point-by-point comments and suggestions.

Point 1: Do not use the first-person pronoun in the paper.

Point 2: The abstract must explicitly provide the research gap.

Point 3: The abstract should provide 2–3 quantitative vital results, giving the reader a clear idea of the research findings.

Point 4: Define all acronyms and (or) abbreviations during their first appearance in the abstract and main text and use them consistently throughout the paper.

Point 5: Besides stating that there is limited literature on applying deep neural networks to GPR image segmentation, the paper must also explicitly present the existing issues of similar studies. Including existing methods' detection capability and accuracy in the introduction would also be helpful.

Point 6: Provide the paper's organization and (or) structure in the last paragraph of the introduction.

Point 7: Ensure that all figures and (or) tables appear after they are mentioned in the paper.

Point 8: The labels in Figure 2 are not legible. Label the left and right-most ends of Figure 2 (e.g., GPR raw data and binarized ROI).

Point 9: The skip connection part of the algorithm is not discussed thoroughly. The authors must provide concise details and aid about the skip connection.

Point 10: All parameter symbols used in each equation must be entirely defined. Ensure none of these symbols are used more than once with a different meaning.

Point 11: The authors are requested to include subfigure letters (i.e., Figure 3).

Point 12: How many datasets were used for the model training?

Point 13: In the model test setup (Case 1), what medium was used for the ground where the pipes were buried?

Point 14: The authors are requested to mark differences in the forward simulation and preprocessing results from Figure 6 onwards.

Point 15: The numerical results summarized in the tables must be presented textually. What do a higher PSNR, SSIM, and FSIM mean?

Point 16: Provide legends for Figure 9.

Point 17: Where is the summary of the comparison of the quantitative evaluation for the third case (i.e., using field data)?

Point 18: The results presentation could be more innovative and provide new insights or advantages over other conventional image segmentation models.

Point 19: The discussion section needs to be stronger. The discussion must focus on the key findings of the study and not on a general perspective. How does the proposed model perform under different conditions?

Point 20: The study's limitations (i.e., intrinsic and extrinsic) must also be elucidated.

Point 21: The conclusion section must provide the study's key findings and implications.

Comments on the Quality of English Language

Moderate editing of the English language is required.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

It can be concluded that all major comments from the reviewers have been addressed in the revised version of the article. Additional tests on real-world data have been included, the proposed method has been compared with other techniques, the comparative analysis has been deepened, and the conclusions have been expanded to include future research directions and broader applications. I have not father suggestions.