Research on Intelligent Recognition Method of Ground Penetrating Radar Images Based on SAHI

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

this is a good paper.  The math escapes me, but the profiles are good.  I would like to know more about the reflection features that you are modeling.  They appear to be only dipping reflections and complex hyperbolas.  That needs to be explicit here. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

I carefully read the Chen et al. paper. "Research on intelligent recognition...", and I found it interesting and worthy of publication.

I confess that I chose to review the paper because I was attracted by the topic. Although I am an expert on the GPR method, I am not an expert in deep learning: my goal was also to learn some concepts from reading  the paper (which is also the the goal of many readers).

I found the paper very technical, complex to fully understand for those who are not deeply involved in this type of research. However, the aim are clear and so are the general concepts of the methodological approach. The results obtained in the tests show the increases in GPR image recognition after data augmentation based on deep convolutional generative adversarial networks (DCGAN), and the use of SAHI.

The methodologies are adequately explained, and the results are clearly separated from the interpretations.

The first three lines of the introduction are unhelpful and based on inappropriate references: Wang et al. 2017 is about a specific GPR system used for exploring tsunami deposit, and cannot represent the GPR near-surface investigation as a whole. It is more appropriate to cite a more general paper, or a textbook. The same for Thabit et al. 2018, which looks like a local application (the reference in the reference list is incomplete).

For the rest the paper looks logically organized and the conclusion appears supported by the test results.

Hope this may help

Best wishes

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The main objective of the paper is a method to detect structural loose diseases in ground penetration radar (GPR) images. The method is based on slicing aided hyper inference (SAHI) and uses data augmentation based on deep convolutional adversarial networks (DCGAN) and YOLOv5 (you only look once v5) neural network for classification. Considering the methods employed are well known, the contribution of the paper would be focused on practical results as it contains extensive experiments. However, there are some issues that need to be solved. In general, the literal presentation of the paper is good, but there is still room for improvement in this regard. The discussion on data augmentation should be extended. The evaluation of results should be improved. In summary, I consider the contents of the paper are potentially publishable, but the following issues should be addressed in a revised version of the paper.

- There is room for improvement in the literal presentation of the paper. For instance, (i) please explain the term “structural loose diseases”. Is “disease” the proper term? (ii) Line 212, “Data augmentation of structural loose diseases using DCGAN”. perhaps it was meant: ““Data augmentation of structural loose diseases using DCGAN was applied”.

Therefore, an English proofreading is recommended.

- The evaluation of results should be improved. (i) Besides the precision-recall (PR) curve analysis included, a receiver operating characteristic (ROC) curve analysis should be also included. The area under those curves should be estimated and the performance of the detector be analyzed in the regimen of low or very low false positive rate. A zoom of the curves in that area could be shown. (ii) The paper lacks a comprehensive analysis of the statistical significance of the results. It is important to include measures of statistical significance, such as p-values or confidence intervals, to assess the reliability and significance of the reported findings. Incorporating a rigorous statistical analysis would enhance the scientific rigor and strengthen the conclusions drawn from the experiments. (iii) The variability of the classification results should be estimated and discussed, i.e., the mean and standard deviation of a set of Montecarlo experiments (randomly changing the training and testing datasets).

- A comparison with competitive state of the art methods is required. This allows the real contribution of the paper to be realized.

- In several parts of the paper is stated the need to apply data augmentation given the “insufficient structural loose disease data”. Besides, some experiments were made to verify the effect of data augmentation (lines 245-247). The “sufficient” data are related with the sample size required to obtain a stable classification performance. Thus, scarcity of the data imposes synthetic data generation to make machine learning (ML) feasible. This idea should be more elaborate. The scarcity of the data is considered when data sample is not enough to ML estimate the real distribution of the data. This can lead to bias and overall poor performance of ML methods. Learning curves could be estimated, providing an estimate of the reduction in the excess of the probability of error obtained by increasing the training set size. Please extend this discussion; I suggest the following recent reference: https://doi.org/10.1016/j.patcog.2022.109240.

 

Comments on the Quality of English Language

Please see the "Comments and Suggestions for Authors" section.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 3 Report

Comments and Suggestions for Authors

The quality of the paper has been improved. Some of my concerns have been adequately addressed in the revised version and some issues such as statistical significance analysis and comparison with state of the art methods have been left opened for future work. Concerning data augmentation subject that is one strong point of the paper, a discussion on theoretical estimation of the sample data to obtain a classification performance is essential. It was highlighted in previous review to improve the scientific level of the paper and it should be discussed in the paper. In summary, I consider the contribution of the paper is just experimental since the methods are well known, but it should not avoid the publication of this paper after editing revision and including the changes commented above.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf