Rock Layer Classification and Identification in Ground-Penetrating Radar via Machine Learning

Round 1

Reviewer 1 Report

This manuscript is very interesting in terms of novel data interpretation of GPR data. The authors test methods performance both in the synthetics and real data. However, there are some issues to be addressed to make it clear and meet the standard of the Journal.

1.     The structure of the manuscript should be re-arranged. The section 5 (results) and 6 (discussion) seem to be the discussion section, Whereas the section 3 (Synthetic Data Experiment) and 4 (Measured Data Experiment) should be the results section.

2.     In methodology, the authors should explain more details about implementation of the methods, some parameters used, and software or code used.

3.     In discussion, the real data seem to be layer model, but the synthetic data is mainly diffraction pattern. Are there any differences and difficulties in the data interpretation by deep learning? Are there any effect of noisy GPR section on this method? And how to prepare the input data? What happen if the input data was processed by applying gain or migration?

Author Response

Question 1:

This manuscript is very interesting in terms of novel data interpretation of GPR data. The authors test methods performance both in the synthetics and real data. However, there are some issues to be addressed to make it clear and meet the standard of the Journal.

Answer 1:

Thank you for your valuable comments and feedback on our paper. The issues you have raised will be carefully considered, and the necessary revisions and improvements will be made to meet the requirements and standards of the journal.

Question 2：

The structure of the manuscript should be re-arranged. The section 5 (results) and 6 (discussion) seem to be the discussion section, Whereas the section 3 (Synthetic Data Experiment) and 4 (Measured Data Experiment) should be the results section.

Answer 2：

Based on your suggestion, the structure of the manuscript has been rearranged by relocating sections 5 and 6 to appropriate positions. Your guidance is greatly appreciated, and the necessary modifications to the paper have been made accordingly.

Question 3：

In methodology, the authors should explain more details about implementation of the methods, some parameters used, and software or code used.

Answer 3：

Our methodology section provides an overview of the implementation of our methods. We acknowledge the reviewer's suggestion to include more detailed explanations regarding the parameters used and the software or code utilized. All the code used in this study was developed by us. We would like to clarify that we did not use any pre-existing software packages for our analysis. The numerical simulation code for GPR was written in MATLAB, and the deep learning component was implemented using the PyTorch package in Python. Additionally, the neural network model was custom-built by us. In the revised manuscript, a more comprehensive explanation of the parameters used in our methodology will be provided, along with further elaboration on the implementation details of our code.

Question 4：In discussion, the real data seem to be layer model, but the synthetic data is mainly diffraction pattern. Are there any differences and difficulties in the data interpretation by deep learning? Are there any effect of noisy GPR section on this method? And how to prepare the input data? What happen if the input data was processed by applying gain or migration?

Answer 4：

Your questions have been excellent, and indeed, experimentation with deep learning is an exploratory process. Noise can indeed have a detrimental impact on prediction results. To elaborate, the deep neural network can be viewed as a black box, seeking to extract signals from an image. However, noise tends to obscure the signal energy to some extent, resulting in relatively poorer predictions. Nevertheless, predictions can still be made, albeit slightly compromised.

Regarding the input data, it has been artificially synthesized by us to address real-world data challenges. Therefore, the diversity of our synthesized samples is crucial. The more knowledge incorporated into the synthetic data, the better the network can handle real-world data.

If the radar data undergoes gain processing, the results may vary slightly, but not significantly. This is because, in the process of extracting signals from the image, the black box mechanism of the neural network can detect both small and large signals, unlike traditional mathematical algorithms, which rely solely on computed data for decision-making. This illustrates the difference between deep learning and traditional algorithms. Despite the complexity in interpreting the workings of deep neural networks, they excel in handling intricate data relationships.

Reviewer 2 Report

This paper focuses on reconstructing subsurface permittivity maps of rock layer from GPR data using Unet network. FDTD method was used to generate GPR data for training the network, and a field test was proposed to verify their method. The idea is very innovative. But from the abstract to the results, the content is not clear for reader. For example, the rock layer’s specific physics feature for FDTD model, the specific FDTD simulation parameters and filed test parameters. All are ambiguous. Besides, the field test contains lots of noise signal, which is not including in the training dataset, how did you denoise them and how did the ideal model classify it? All the important items are not clear, so, I think this paper is not ready for publication at this state.

Some suggestions for further publication.

(1)    Detail the rock layer physics feature, state how you simulate them and detail your FDTD parameters.

(2)    Please compare the performance of model proposed with other networks like DMRF-UNet or methods like FWI algorithms?

(3)     What are the experiment parameters, such as GPR parameters, field site, hardware, devices and so on?

(4)    How did you transfer your model trained on simulated data to field data?Add denoise method?

Author Response

Question 1:

This paper focuses on reconstructing subsurface permittivity maps of rock layer from GPR data using Unet network. FDTD method was used to generate GPR data for training the network, and a field test was proposed to verify their method. The idea is very innovative. But from the abstract to the results, the content is not clear for reader. For example, the rock layer’s specific physics feature for FDTD model, the specific FDTD simulation parameters and filed test parameters. All are ambiguous. Besides, the field test contains lots of noise signal, which is not including in the training dataset, how did you denoise them and how did the ideal model classify it? All the important items are not clear, so, I think this paper is not ready for publication at this state.

Answer 1:

Thank you for reviewing our paper and providing valuable feedback. Your concerns have been addressed by providing further clarification and explanations, and necessary revisions have been made to ensure the content is clearer and more readable. Specifically, detailed explanations have been added to the paper regarding the physical characteristics of rock layers in the FDTD model, as well as the specific parameters used for simulation and field testing. Regarding the noise signals present in the field test, our denoising methods and how the ideal model classifies them have been explained. Efforts have been made to make these key points more transparent and understandable.

Question 2:

Detail the rock layer physics feature, state how you simulate them and detail your FDTD parameters.

Answer 2:

Thank you for the reviewer's inquiry. Our study focuses on achieving precise identification of underground rock formations and stratigraphic structures through the integration of Ground Penetrating Radar (GPR) technology and deep learning models. In our research, the physical characteristics of rock layers involve their dielectric constants, thickness, shape, as well as the reflection and refraction of electromagnetic waves. We employ the Finite-Difference Time-Domain (FDTD) method for numerical simulation to simulate the electromagnetic response of underground rock layers. Additionally, we provide detailed descriptions of the FDTD parameters used in the simulation process, including grid size, time step, frequency range, etc., to ensure the accuracy and reliability of the simulation results.

Question 3:

Please compare the performance of model proposed with other networks like DMRF-UNet or methods like FWI algorithms?

Answer 3:

Our technology is novel, with no similar work done previously. The working principle of deep learning varies depending on the samples. Our method involves using ground-penetrating radar data to locate rock bodies and stratigraphic information, rather than inversely calculating the dielectric constant distribution from radar data. Since our objectives differ, our methods cannot be directly compared. While the DMRF-UNet and our UNet network structures are nearly identical, the former employs two UNet networks for denoising and inversion. Despite the similar network structures, our prediction goals differ, making direct comparison impractical. However, FWI and DMRF-UNet can be compared as they both involve computational parameter models. We have conducted some research on FWI and are prepared to present inversion model results of FWI and our proposed deep learning network (such as CL-UNet, which outperforms DMRF-UNet) in an article currently under submission. Although the tasks are similar, our research does not focus on inversion, hence not suitable for direct comparison. We hope this clarifies the matter.

Question 4:

What are the experiment parameters, such as GPR parameters, field site, hardware, devices and so on?

Answer 4:

Thank you for your reminder. We will incorporate these important parameters into the paper. The detailed descriptions of the FDTD parameters used in the simulation process, including grid size, time step, frequency range, etc., have been provided to ensure the accuracy and reliability of the simulation results. These parameters can be seen between lines 308 and 320, and this task has already been completed.

Question 5:

How did you transfer your model trained on simulated data to field data?Add denoise method?

Answer 5:

Your question is highly significant. Our strategy involves utilizing traditional deep learning and network prediction methods. We train the network using artificially synthesized samples, which have a mapping relationship with real data and can be obtained through technical means. This allows the network to memorize and understand the information in the radar data. Subsequently, we utilize the trained network to infer information from actual measurement data. While artificial synthetic data does not contain noise, it is inevitable in actual measurement data. Therefore, the measured data needs to undergo denoising processing, for which we employ common traditional denoising algorithms such as median filtering, though other algorithms can also be attempted. Although denoised radar data may still contain some noise, the signal becomes clearer. Deep learning networks can predict based on signal characteristics, thus completing tasks. As deep learning networks have not been trained on samples containing noise, noise in actual measurement data is unfamiliar to the network, and there are no corresponding neurons to identify it. Therefore, the key issue lies in the significance of signal components in actual measured data, which directly impacts the network's performance.

Author Response File: Author Response.pdf

Reviewer 3 Report

Rock Layer Classification in Ground-penetrating radar via Feature Extraction and Machine Learning

The English must be improved, some sentences are not clear.

The structure of the paper must be revised. Paragraph 5 – results: I think that is the “discussion”. The paragraph 6 “discussion”, I think is "conclusion". In paragraphs 4 and 5 there are several comments on “future works”. I suggest collecting all these future works statements and moving into the conclusions. The paper emphasizes the goodness of the work with no scientific words, like immense, judiciously, meticulously, adroitly etc. that must be removed. The paper must be a clear description of the work done, without the subjective comments of the authors.

The introduction must “introduce” the paper. From lines 124-137, too many details are given about the presented work. I suggest rephrasing it, and giving an overview of the presented method, validation procedure and case study.

References in some cases are missing (see my detailed comments)

References have to be improved. For example, you can add references, at line 45 after [9,10] to GPR applications related to mining exploitation and fracture detection like:

Arosio, D. (2016). Rock fracture characterization with GPR by means of deterministic deconvolution. Journal of Applied Geophysics,

Elkarmoty, M., Colla, C., Gabrielli, E., Bonduà, S., & Bruno, R. (2017). Deterministic three-dimensional rock mass fracture modeling from geo-radar survey: A case study in a sandstone quarry in Italy. Environmental and Engineering Geoscience, 23(4), 313–330. https://doi.org/10.2113/gseegeosci.23.4.314

There is no information about the coding of the algorithm, source code, computational time, etc. Also computational time for training and validation must be given.

I also recommend to share all data by using open science data repository. I suggest that making accessible the training images, the syntethic radargrams and the true radargram and interpretation will be useful for future references to be computed with other algorithm or methods

Detailed comments:

L7: multidisciplinary knowledge: too vague, please remove

Figure 1: it is not clear what the arrows represent. Also the 2 words “Rock” under the circle are not appropriate. What the circles represents? I suggest to give a scheme and the corresponding radargram (the radargram of figure 1 is not from the presented scheme. It is recommended to add (a), (b) (c) to the 2 images and add a detailed description in the caption of the figure. Figure 1 is not cited in the text.

L195: add reference to UNet

L201: please remove “immense”

L206: judiciously? Please rephrase it and explain more in detail

L211: adroitly? Please rephrase it and explain more in detail

Figure 3:Convolution, RElu, deonvolution, in the images, are not explained in the text. Please give a fully explanation of filters/algorithm

L216: the authors stated that they tune a set of parameters. I suggest adding a table with all parameters and tuned values.

L217: explain and give references to learning rate parameter

L219: please specify 300 synthetic images (“epoch” is not appropriate here)

L215-230: I cannot understand the number of training images: firstly, I found 300, then 800. What is the difference? Please revise and be clearer.

L282. “Circular rock mass” is no so common. Maybe you can refer to rounded shape body or similar

L221: Adam algorithm: explain and give references

L333: PML : please define the acronym

Figure 5 is not clear: what the different colours represents? There are images on the left and one in the right. Why? I suggest labelling all images in (a), (b) etc. For each item give a description and comments about the goodness of the results.

Line 376-385 ->in the left there is a green arrow that is not explained.

Figure 7: I suggest having same Y max on both graph

L460-476 – the paragraph is not relevant in this context. May be must moved in the conclusion. It is vague. Results are not proved and it can be a reasonable interpretation of the radar image.

Figure 10 is a zoom of the figure 9. Please put a rectangle in Figure 9 to understand which part represents.

English is comprehensible but needs of improvements

Author Response

Question 1:

Rock Layer Classification in Ground-penetrating radar via Feature Extraction and Machine Learning. The English must be improved, some sentences are not clear.

Answer 1:

Thank you for your suggestion. I have already changed the title to: 'Rock Layer Classification and Identification in Ground-Penetrating Radar via Machine Learning'.

Question 2:

The structure of the paper must be revised. Paragraph 5 – results: I think that is the “discussion”. The paragraph 6 “discussion”, I think is "conclusion". In paragraphs 4 and 5 there are several comments on “future works”. I suggest collecting all these future works statements and moving into the conclusions. The paper emphasizes the goodness of the work with no scientific words, like immense, judiciously, meticulously, adroitly etc. that must be removed. The paper must be a clear description of the work done, without the subjective comments of the authors.

Answer 2:

Thank you for your suggestion. I have made the requested revisions. Please review lines 486-534 where I have rewritten this section.

Question 3:

The introduction must “introduce” the paper. From lines 124-137, too many details are given about the presented work. I suggest rephrasing it, and giving an overview of the presented method, validation procedure and case study.

Answer 3:

Thank you for your suggestion. I have revised the content between lines 122 and 137, providing an overview of the presented method, validation procedure, and case study.

Question 4:

References in some cases are missing (see my detailed comments)

References have to be improved. For example, you can add references, at line 45 after [9,10] to GPR applications related to mining exploitation and fracture detection like:

Arosio, D. (2016). Rock fracture characterization with GPR by means of deterministic deconvolution. Journal of Applied Geophysics,

Elkarmoty, M., Colla, C., Gabrielli, E., Bonduà, S., & Bruno, R. (2017). Deterministic three-dimensional rock mass fracture modeling from geo-radar survey: A case study in a sandstone quarry in Italy. Environmental and Engineering Geoscience, 23(4), 313–330. https://doi.org/10.2113/gseegeosci.23.4.314

Answer 4:

Thank you very much for your valuable feedback. I have carefully reviewed the detailed comments you provided and will improve the references accordingly based on your suggestions. I have already completed adding references related to ground-penetrating radar applications in mining exploitation and fracture detection after the references [9,10] in line 45 to enhance the completeness and credibility of the paper.

Question 5:

There is no information about the coding of the algorithm, source code, computational time, etc. Also computational time for training and validation must be given.

Answer 5:

We have developed our own code and derived the algorithm formulas in the original manuscript. Computational time details for training and validation will be included in the revised paper. Additionally, information about the source code availability and any relevant computational details will be provided to ensure transparency and reproducibility.

Question 6：

I also recommend to share all data by using open science data repository. I suggest that making accessible the training images, the syntethic radargrams and the true radargram and interpretation will be useful for future references to be computed with other algorithm or methods

Answer 6:

Your suggestion is very valid, but the openness of our data and code may be restricted by our research institution or partners, potentially due to copyrighted content or involvement of commercial secrets. The research data could be the accumulation of years of study, consuming significant time and resources for analysis and processing. While sharing data openly may benefit other researchers, it could also diminish the original research's value. We have provided the emails of the authors and corresponding authors in the paper, allowing future researchers to contact us for data sharing based on their needs and reasonableness. I have revised the content between lines 542 and 543.

Question 7：

L7: multidisciplinary knowledge: too vague, please remove

Answer 7：

I have made the modifications as per your suggestions.

Question 8：

Figure 1: it is not clear what the arrows represent. Also the 2 words “Rock” under the circle are not appropriate. What the circles represents? I suggest to give a scheme and the corresponding radargram (the radargram of figure 1 is not from the presented scheme. It is recommended to add (a), (b) (c) to the 2 images and add a detailed description in the caption of the figure. Figure 1 is not cited in the text.

Answer 8：

I have made the modifications as per your suggestions. Please refer to Figure 1.

Question 9：

L195: add reference to UNet

Answer 9：

I have made the modifications as per your suggestions.

Question 10：

L201: please remove “immense”

Answer 10：

I have made the modifications as per your suggestions.

Question 11：

L206: judiciously? Please rephrase it and explain more in detail

Answer 11：

I have removed such supervisory words. Thank you for your reminder and suggestion.

Question 12：

L211: adroitly? Please rephrase it and explain more in detail

Answer 12：

I have removed such supervisory words. Thank you for your reminder and suggestion.

Question 13：

Figure 3:Convolution, RElu, deonvolution, in the images, are not explained in the text. Please give a fully explanation of filters/algorithm

Answer 13：

Your suggestion is highly valued. I have now provided explanations in the caption of Figure 3, making it easier for readers to understand

Question 14：

L216: the authors stated that they tune a set of parameters. I suggest adding a table with all parameters and tuned values.

Answer 14：

I have made the modifications as per your suggestions.

Question 15：

L217: explain and give references to learning rate parameter

Answer 15：

I have already provided an explanation in the original text, in lines 199-200.

Question 16：

L219: please specify 300 synthetic images (“epoch” is not appropriate here)

Answer 16：

Epochs is a technical term in deep learning, and there is no mistake here; it refers to the number of training iterations, which in our case is indeed 300. However, I realize that my expression may have been unclear, so I have revised and polished the language in that area, hoping it meets your satisfaction. Please refer to line 403.

Question 17：

L215-230: I cannot understand the number of training images: firstly, I found 300, then 800. What is the difference? Please revise and be clearer.

Answer 17：

I have made the modifications as per your suggestions. 800 represents the number of training samples, while 300 represents the number of training iterations; these are not the same concepts. It's possible my expression was unclear, so I have made modifications to the original text. Please refer to lines 198-204.

Question 18：

L282. “Circular rock mass” is no so common. Maybe you can refer to rounded shape body or similar

Answer 18：

Your expression is more accurate. I have made modifications based on your advice.

Question 19：

L221: Adam algorithm: explain and give references

Answer 19：

Adam is a commonly used optimization algorithm in deep learning. I have added references in the original text. Adam is used to adjust the weights of neural networks to minimize training errors. The Adam algorithm combines the features of momentum methods and adaptive learning rates, allowing the learning rate to be adjusted adaptively during the training process, thereby improving training speed and increasing model stability.

Question 20：

L333: PML : please define the acronym

Answer 20：

I have made the modifications as per your suggestions.

Question 21：

Figure 5 is not clear: what the different colours represents? There are images on the left and one in the right. Why? I suggest labelling all images in (a), (b) etc. For each item give a description and comments about the goodness of the results.

Answer 21：

I have made the modifications as per your suggestions. "a" represents the expected output, also known as the ground truth. "b" represents the prediction results of the trained deep neural network, with the color scale indicating probabilities. It can be observed that in "a", positions with a value of 1 usually represent structural information, and in "b", most structural information can be predicted well, with some values also very close to the ground truth.

Question 22：

Line 376-385 ->in the left there is a green arrow that is not explained.

Answer 22：

Thank you for your reminder. I have made modifications to line 374 in the original text.

Question 23：

Figure 7: I suggest having same Y max on both graph

Answer 23：

This description is accurate. L1 loss represents the error between the output and the ground truth. The left image represents the training set, while the right image represents the validation set. The left image is used for model training, while the right image is used to evaluate the model's performance. It can be likened to studying while simultaneously doing practice questions. Errors may occur during the learning process, and similarly, errors may occur when tackling new questions. Both images represent the concept of errors. Only when both training and validation are satisfactory can the deep network be considered qualified.

Question 24：

L460-476 – the paragraph is not relevant in this context. May be must moved in the conclusion. It is vague. Results are not proved and it can be a reasonable interpretation of the radar image.

Answer 24：

You are absolutely right. I have made the modifications as per your request. Your feedback has improved the quality of our manuscript

Question 25：

Figure 10 is a zoom of the figure 9. Please put a rectangle in Figure 9 to understand which part represents.

Answer 25：

Your feedback is crucial. To enhance clarity, we have added annotations to Figure 9 and provided explanations in the caption of Figure 10.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

The topic is very good. But I am confused by the effort conducted at current revision, due to  most of the previous questions are not addressed, and the authors are not marking any change in the current manuscript.  The field test can not validate your simulation result  due to different situations. So, I don't think this paper is suitable for publication at this state.

Author Response

I apologize for only uploading a new version of the manuscript last time without including a marked version, as there were significant changes made to the manuscript. This time, I have uploaded all versions of the manuscript. Firstly, I will address your questions in order.

Regarding the FDTD parameters, you can find all descriptive information, including grid size and the number of sources, in the marked manuscript on lines 228-239.

Regarding noise, our training set does not contain noise. Deep networks only need to learn the relationship between data and labels. Even without noise, GPR data contains a lot of interference, such as multiple waves, direct waves, and diffracted waves, which pose challenges for prediction and are almost indistinguishable from added noise. Real measurement data undergo some preprocessing, including simple denoising and gain adjustment, while retaining most of the reflected signals. For our network, simple preprocessing is sufficient for regular use. Even with a small amount of noise, it will not significantly affect the results. Convolutional neural networks have excellent computer vision capabilities, allowing them to extract valid signals from blurry images and then predict labels.

Regarding the comparison between DMRF-UNet and FWI, our approach is unrelated to these two methods. They focus on inversion, while our work focuses on signal extraction. Due to the different labels used, direct comparison is not feasible. The network structure used in DMRF-UNet is very similar to ours, and if inversion were to be performed, the results should be very similar to DMRF-UNet, but our work does not involve inversion.

Regarding your request to include the ground penetrating radar parameters for the measured data, we have made changes in the content section of the revised manuscript. You can find this information in lines 432-439 of the revised manuscript.

Lastly, regarding the relationship between noise and data, we use noise-free data during network training. Real measurement data require denoising preprocessing before use. Noise is not considered during the training phase, and efforts are made to remove noise from the measured data to adapt to the trained network.

Author Response File: Author Response.pdf

Reviewer 3 Report

the authors fully replied to the reviewers comments and suggestions

acceptable

Author Response

Thank you for your advice, it has been very helpful. Additionally, I appreciate your recognition and support, and I will continue to strive for improvement.