Improved Cycle-Consistency Generative Adversarial Network-Based Clutter Suppression Methods for Ground-Penetrating Radar Pipeline Data
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsDear Authors,
"Improved CycleGAN-Based Clutter Suppression Methods for GPR Pipeline Data" presents a novel approach to enhancing clutter suppression in Ground Penetrating Radar (GPR) images using an improved CycleGAN network. The study introduces modifications to both the generator and discriminator components of the CycleGAN to better handle the unique challenges of clutter and target signal distinction in GPR data. By incorporating multiple residual blocks and a channel attention mechanism (SE module), the proposed method aims to improve the deep learning model's ability to differentiate between clutter and significant underground pipeline features. The effectiveness of this approach is validated through comparative analysis with traditional clutter suppression methods and other deep learning-based approaches, using both simulated and measured data. The results demonstrate the potential of the improved CycleGAN network in effectively reducing clutter in GPR images, thereby enhancing the detection accuracy of underground pipelines.
However, it's recommended that the article undergoes a minor revision. Although, the core methodology, results, and contributions are solid and innovative, addressing the following points could significantly strengthen the paper:
- Enhanced Validation: Include additional validation on diverse real-world datasets to better demonstrate the model's generalization ability and its performance in practical scenarios.
- Mitigation Strategies for Overfitting: Elaborate on the techniques used to prevent overfitting, especially considering the reliance on simulated data for training.
- Detailed Discussion on Limitations: A more comprehensive discussion on the limitations of the proposed method, including potential challenges in real-world applications and suggestions for future research directions, would provide a more balanced view of the work.
Implementing these revisions would not only address the current cons but also enhance the overall impact and applicability of the research findings.
Comments on the Quality of English LanguageIt seems that the document would only require minor revisions for English language quality. This includes proofreading for typographical errors, ensuring consistency in terminology and formatting, and possibly refining some sentences for better clarity and impact. The scientific content appears to be well-communicated, suggesting that the authors have effectively conveyed their research findings and the significance of their work in the context of existing literature.
Author Response
Reply to reviewer 1：
Dear reviewer, thank you for your comments. The revised parts of the paper are highlighted in
yellow. The following is a detailed reply to your comments.
However, it's recommended that the article undergoes a minor revision. Although, the
core methodology, results, and contributions are solid and innovative, addressing the
following points could significantly strengthen the paper:
1. Enhanced Validation: Include additional validation on diverse real-world datasets to better
demonstrate the model's generalization ability and its performance in practical scenarios.
Reply: Thank you for your comment. We have added comparative validation of measurement data
in different scenarios to better demonstrate the model's generalization ability.The verification of the
added measured data is reflected in section 5.3.2 of the article.
2. Mitigation Strategies for Overfitting: Elaborate on the techniques used to prevent
overfitting, especially considering the reliance on simulated data for training.
Reply: Thank you for your comment. Regarding the prevention of overfitting, the model in this
article has been designed with consideration. In the loss function section, L1 norm is introduced to
regularize the network, thereby reducing the complexity of the model when updating network
parameters and preventing overfitting during network training. The relevant parts have been
highlighted in yellow in section 3.1, line 280. The relevant introduction to the loss function is
highlighted in yellow in section 3.1, line 189 of the text.
The reference is: [28] J. -Y. Zhu, T. Park, P. Isola and A. A. Efros, "Unpaired Image-to-Image
Translation Using Cycle-Consistent Adversarial Networks," 2017 IEEE International Conference on
Computer Vision (ICCV), Venice, Italy, 2017, pp. 2242-2251, doi:10.1109/ICCV.2017.244.[CrossRef]
3. Detailed Discussion on Limitations: A more comprehensive discussion on the limitations of
the proposed method, including potential challenges in real-world applications and
suggestions for future research directions, would provide a more balanced view of the work.
Reply: Thank you for your comment.We provide a more comprehensive discussion on the
limitations of the proposed method in Chapter 6 of the article, including potential challenges in
real-world applications and recommendations for future research directions. The relevant positions in
the text have been highlighted in yellow.
It seems that the document would only require minor revisions for English language quality.
Reply: Thank you for your comments. We have checked and corrected the English grammar
throughout the paper.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThis paper discussed an improved clutter suppression network based on CycleGAN. Some parts of the paper need strengthening. These following issues must be addressed and require major revisions.
1. Abstract: The description of the background is too long, please delete it appropriately. Some abbreviations are not introduced at the first time, and some abbreviations are not introduced at the first time. In addition, Some description of the quantitative results should be given.
2. Some key papers about GPR detection and data augmentation should be discussed in the introduction:
1) GPR-based detection of internal cracks in asphalt pavement: A combination method of DeepAugment data and object detection, DOI: 10.1016/j.measurement.2022.111281.
2) S-CycleGAN: A Novel Target Signature Segmentation Method for GPR Image Interpretation, DOI: 10.1109/LGRS.2024.3365470.
3) Automatic recognition of pavement cracks from combined GPR B-scan and C-scan images using multiscale feature fusion deep neural networks, DOI: 10.1016/j.autcon.2022.104698.
3. The grammar in the essay should be checked thoroughly.
4. Paper layout should be beautiful and comfortable, and a lot of white space is not appropriate. There is a lot of white space before Figure 4 and Figure 10, which can be solved by adjusting the layout of the text and figures.
5. Equations should be given the references. In addition, Equations (X) should be mentioned in the main text.
6. Does the abnormal feature of a hyperbola in a GPR image necessarily represent a pipeline? Please give some model or field test verification results to increase the convincing.
7. Is your experiment strictly controlled for variables? How is it reflected? What variables play an important role in your experiment? Why?
8. The overall quality of the pictures in this article is poor. It is recommended to use vector images.
9. Section 4.2, The parameters of many models are given. Please briefly explain the basis for selecting these parameters.
10. Limitations of the study should be appropriately mentioned in the conclusion.
11. Authors should strictly follow the paper template on the official website for formatting.
12. The model proposed in this paper lacks comparative experiments with other mainstream models.
13. Academic writing tends to use the passive voice and the third person. However, the first person appears several times in this paper which needs to be checked for the full text and corrected. Again, the passive voice is recommended.
Comments on the Quality of English LanguageModerate editing of English language required
Author Response
Reply to reviewer 2：
Dear reviewer, thank you for your comments. We have revised the paper based on your feedback. The revised parts of the paper are highlighted in yellow. The following is a detailed reply to your
comments. Comments and Suggestions for Authors
The manuscript can be major revision for many reasons but these can generally be divided into
technical reasons. 1. Abstract: The description of the background is too long, please delete it appropriately. Some abbreviations are not introduced at the first time, and some abbreviations are not
introduced at the first time. In addition, Some description of the quantitative results should
be given. Reply: Thank you for your comments. We have made appropriate deletions to the abstract and
supplemented some explanations of the quantitative results. The modified abstract is as follows:
Ground-penetrating radar (GPR) is a widely used technology for pipeline detection due to its fast
detection speed and high resolution. However, the presence of complex underground media often
results in strong ground clutter interference in the collected B-scan echoes, significantly impacting
detection performance. To address this issue, this paper proposes an improved clutter suppression
network based on CycleGAN (cycle-consistency generative adversarial network). By employing the
concept of style transfer, the network aims to convert clutter images into clutter-free images. This paper
introduces multiple residual blocks into the generator and discriminator respectively to improve the
feature expression ability of the deep learning model. Additionally, the discriminator incorporates the
SE (Squeeze and Excitation) module, a channel attention mechanism, to further enhance the model's
ability to extract features from clutter-free images. To evaluate the effectiveness of the proposed
network in clutter suppression, both simulation and measurement data are utilized to compare and
analyze its performance against traditional clutter suppression methods and deep learning-based
methods, respectively. From the result of the measured data, it can be found that the improvement
factor ( Im ) of the proposed method has reached 40.68dB, which is a significant improvement
compared to the previous network. 2. Introduction Some key papers about GPR detection and data augmentation should be
discussed in the introduction：
1) GPR-based detection of internal cracks in asphalt pavement: A combination method of
DeepAugment data and object detection, DOI: 10.1016/j.measurement.2022.111281. 2) S-CycleGAN: A Novel Target Signature Segmentation Method for GPR Image
Interpretation, DOI: 10.1109/LGRS.2024.3365470. 3) Automatic recognition of pavement cracks from combined GPR B-scan and C-scan
images using multiscale feature fusion deep neural networks, DOI: 10.1016/j.autcon.2022.104698.
Reply: Thank you for your comments. We have added a discussion on literature on ground
penetrating radar detection and data augmentation in the introduction section. I cited and discussed
these latest papers in the introduction. The first paper is highlighted in yellow in the first chapter of the
main text, located on line 30, with reference number 2. The second paper is highlighted in yellow on
line 101 of Chapter 1 of the main text, with reference number 21. The third paper is highlighted in
yellow on line 105 of Chapter 1 of the main text, with reference number 22. 3. The grammar in the essay should be checked thoroughly. Reply: Thank you for your comments. We checked the grammar throughout the paper. 4. Paper layout should be beautiful and comfortable, and a lot of white space is not
appropriate. There is a lot of white space before Figure 4 and Figure 10, which can be solved
by adjusting the layout of the text and figures. Reply: Thank you for your comments. We have adjusted the layout of the paper and resolved the
blank space in front of Figures 4 and 10. 5. Equations should be given the references. In addition, Equations (X) should be mentioned in
the main text. Reply: Thank you for your comments. The equation has provided references at the corresponding
positions. The second thing has also been checked, and all equations are mentioned in the text. 6. Does the abnormal feature of a hyperbola in a GPR image necessarily represent a pipeline?
Please give some model or field test verification results to increase the convincing. Reply: Thank you for your comments.The paper provides relevant descriptions and additional
references in the background description section.In the first chapter of this paper, the explanation that
pipelines are displayed as hyperbolas during GPR pipeline detection is highlighted in yellow. The reference is: [7] Ozkaya, U.; Melgani, F.; Belete Bejiga, M.; Seyfi, L.; Donelli, M. GPR B
scan image analysis with deep learning methods.Measurement 2020, 165, 107770.[CrossRef]. Moreover, our experimental scenario is the pipeline burial point in residential areas, with the
main underground burial material being pipelines. 7. Is your experiment strictly controlled for variables? How is it reflected? What variables
play an important role in your experiment? Why?
Reply: Thank you for your comments. The experiment in this paper strictly controls variables, with the main variables being the radar band, acquisition time window, different pipeline materials, and
pipeline positions. The radar uses the actual experimental band of 1600MHz, and the acquisition time
window is also set at 15ns according to the actual experiment. The pipeline material is conventional, and the pipeline position is randomly set. By changing the relevant variables, different types of data
can be obtained, thereby improving the model's generalization ability for clutter suppression. The most
important variable in this paper is the radar band, which will determine the hyperbolic characteristics of
the pipeline. The selection of variables is described in [27]. The reference is: [27] H. -H. Sun, W. Cheng and Z. Fan, "Learning to Remove Clutter in Real- World GPR Images Using Hybrid Data," in IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1-14, 2022, Art no. 5113714, doi: 10.1109/TGRS.2022.3176029.[CrossRef]
8. The overall quality of the pictures in this paper is poor. It is recommended to use vector
images. Reply: Thank you for your comments. We have replaced the image with a clearer one. 9. Section 4.2, The parameters of many models are given. Please briefly explain the basis for
selecting these parameters. Reply: Thank you for your comments. In order to make the simulation data closer to the actual
measurement data, we used the parameters set by the radar during the simulation using gprMax. The
specific parameter selection was set with reference to paper [27]. The reference is: [27] H. -H. Sun, W. Cheng and Z. Fan, "Learning to Remove Clutter in Real- World GPR Images Using Hybrid Data," in IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1-14, 2022, Art no. 5113714, doi: 10.1109/TGRS.2022.3176029.[CrossRef]
10. Limitations of the study should be appropriately mentioned in the conclusion. Reply: Thank you for your comment.We provide a more comprehensive discussion on the
limitations of the proposed method in Chapter 6 of the paper, including potential challenges in real- world applications and recommendations for future research directions. The relevant positions in the
text have been highlighted in yellow. 11. Authors should strictly follow the paper template on the official website for formatting. Reply: Thank you for your comment.We have rechecked the format of the paper and made
corrections. 12. The model proposed in this paper lacks comparative experiments with other mainstream
models. Reply: Thank you for your comment.We conducted comparative experiments in Chapter 5 of the
paper, where SVD and RPCA are the mainstream methods in traditional clutter suppression, while the
original cycleGAN and U-Net are the mainstream deep learning based clutter suppression methods. Through comparison, we found that SVD and RPCA cannot remove some background clutter. The
deep learning methods U-Net and CylceGAN have shown good capabilities in eliminating GPR clutter, but they still have the problem of weakening or distorting some target responses. The improved
CycleGAN network in this paper has significant improvements compared with traditional methods and
deep learning methods. On the basis of removing background clutter, it can well retain most of the
characteristic information of the target response. The improvement factor Im has increased to 40.68dB
in the measured data of a single pipeline compared with before, and has also increased to 47.02dB in
the measured data of multiple pipelines. 13. Academic writing tends to use the passive voice and the third person. However, the first
person appears several times in this paper which needs to be checked for the full text and
corrected. Again, the passive voice is recommended. Reply: Thank you for your comment. We have converted the person of the paper according to
your suggestion, and the passive voice issue has also been changed. Moderate editing of English language required. Reply: Thank you for your comments. We have checked and corrected the English grammar
throughout the paper.
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
Comments and Suggestions for AuthorsMy comments on the initial version of the manuscript have been sufficiently addressed by the authors in this revised version. I have no further comments on the technical aspects. The manuscript may be considered for publication after a proofreading.
