Damage Importance Analysis for Pavement Condition Index Using Machine-Learning Sensitivity Analysis

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

First of all, I believe this is an insightful paper with considerable innovation. However, regarding the research background and significance of this article, the author should further summarize it. Here are my specific comments:

1. in line 16, lines 76-79, lines 766-771, the author's analysis on the composition indicators of PCI is sufficient or not, currently all composition indicators are automated analysis, and basically there are no cost and timeliness issues. Instead, the focus of this study should be on finding which indicators are more effective. What the author may want to do more is to screen the indicators that are more adaptable to machine vision technology, for promoting the development of automatic recognition technology to higher accuracy. Therefore, the author should more clearly and in detail demonstrate the process of selecting indicators adaptable to machine vision technology, which may require necessary adjustments to parts of the article's structure and the direction of discussion.

2. in lines 231-234, the author may need to conduct more extensive research and gain a better understanding of the relevant literature. Such evaluations may be too arbitrary in explaining the background of the literature, as we know that any published literature will involve statistical and data analysis, not entirely subjective evaluations.

3. in section 3.2, and lines 340-345, there is no need to overly elaborate on certain universal statistical concepts or well-known knowledge.

4. in lines 446-447, the authors could provide more details on the source of this data or offer a calculation method.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Dear authors, I have read your paper three times to understand the whole purpose of your paper because I need clarification on some things in your paper. I am from the area and am involved in the design of pavement structures. I work around pavements, and some things are not clear to me. First, let us start with the introduction, where you said you did rigid pavement structures. Does that mean you only did concrete pavement, or did that include stabilized layers? Also, in the introduction, you should have emphasized how many kilometers you considered, using purely informative data, so they would know how much data you have available for your network. After that, in the second chapter where, you stated that it is a literature review, which is quite ok; everyone has a different writing style, and I respect that; it is not a mistake that you made a new chapter, but I have a few questions, first why did you put subchapters model sensitivity, then PCI prediction through indirect variables. You do not need it because there you give an overview of works that deal similar to what you are dealing with; you went too much into the width about model sensitivity and PCI prediction; you could have mentioned that in one paragraph; it does not need to be written so much, because you are departing from the essence of your work and doubts arise during reading. I understand your problem, but you must be more specific about it. In the third chapter, give more specific information. As I said in the introduction, repeat how many km of roads with concrete pavement you have worked on; what are the dimensions of the slabs from there, and what are their relationships? Are they all concrete, or are there some reinforced or overstressed concrete? I do not know as a reader of this paper. Then, you can use Google Maps to show the location of the place and then zoom in and show certain places on the analyzed roads. To give some introduction about that, then to describe a little PCI, you gave what types of damage exist; you could show damage from the field and from those analyzed roads with pictures if you understand me, and with variable intensity L, M, H. Second, explain why PCI, why didn't you consider the HMD4 damage catalog developed by the World Bank or perhaps the SHRP damage catalog developed by The Strategic Highway Research Program. After that, you start talking about what methods you used and what models for neural networks you can use.

Why didn't you use Neural Networks, your ANN method, and Super vectors for this analysis to see how far they deviate from the others? Could you schematically show all your inputs and what your outputs are? I wonder how this can be used in practice. This is unlike deflections, so machine learning is needed to speed it up to shorten and see what the modules are for asphalt and other layers. Did you have all 17 PCI-prescribed damages on each analyzed route? After that, when you do the PCI index, what should we do with that data? It is complex, and many questions arise about this work. As a reviewer, I think the paper is not bad, but in this form, it is unsuitable for publication in this journal. I would love to help the authors with more comments to make the work work, but these are my initial comments for now.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

Overall Gap and Contribution: Since the PCI is calculated using equations based on these damages, their significance might already be embedded in the index for each type of damage. The research gap and main contribution in this area need to be clearly articulated. It would be helpful to explain the PCI calculation process in the manuscript to prevent any confusion.

Dataset for Model Training: Are the PCI output values calculated based on the input of 19 distress types? If so, the intrinsic correlation between input and output could naturally lead to high R² values in machine learning models. This point may need further clarification.

Line 261: Table 1: Please add a description of the correlation column. How was the correlation calculated?

Line 338: Figure 2: The collected PCI values are primarily in the fair to good range. Is this range representative enough to serve as the main output for the prediction model?

Line 339: Figure 3: What is the difference between the Pearson correlation coefficient shown here and the correlation coefficient presented in Table 1?

Line 545: Table 3: What is the difference between the feature importance and the correlation coefficient in Table 1 and Figure 3? Is there any connection between these results?

Emphasizing the Contribution: The manuscript should emphasize its contribution more explicitly. While deep learning methods can identify important factors and predict PCI, they are often perceived as black boxes compared to equation-based calculations. How can the identified important factors be used in calculating PCI in other contexts? Should other researchers train similar models to predict PCI? Clarifying how the findings of this manuscript can be implemented in others' work would be beneficial.

Comments on the Quality of English Language

Language is ok. Some descriptions can be changed --

Line 676 and line 754 - "we" to "authors". In academic journal writing, it is typically recommended to use third-person language rather than first-person

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

Dear authors, Thank you very much for explaining all my questions. I have some comments to improve and upgrade your work.

Kind regards and good luck with your subsequent research.

Reviewer 3 Report

Comments and Suggestions for Authors

The manuscript looks good now.