Model for Estimating the Modulus of Elasticity of Asphalt Layers Using Machine Learning

Round 1

Reviewer 1 Report

The quality of English in the paper should be improved in order to have a more clear and easier understanding of the paper for the readers. In some parts it seems that the text is a simple translation on a single word by word way, which is not adequate.

Row 127 - words "training set" repeated.

Row 158 - how was the best ANN architecture chosen?

Row 256 - there is no explanation for removing the 2 points

Row  328 - in Table 6 - Serbian word "Radovi" was not translated into English

Row 361 - why only for Serbia?

Author Response

We would like to thank the Reviewer 1 for submitted comments and advice, and would like to provide further clarifications and answers to improve the quality of the paper and resolve ambiguities.

The quality of English in the paper should be improved in order to have a more clear and easier understanding of the paper for the readers. In some parts it seems that the text is a simple translation on a single word by word way, which is not adequate.

Thank you for this comment. The text has been revised and certain changes have been made.

1. Row 127 - words "training set" repeated.

We do not understand this question. In row 127, the sentence reads as follows: “In this paper, the available dataset is divided into training data, validation data and testing data.” In that sentence we did not use words “training set” and we did not repeat those words.

2. Row 158 - how was the best ANN architecture chosen?

We created a loop with different numbers of neurons, and a training model for different numbers of neurons. Then we saw that the RMSE for the validation model was set as the function of the number of neurons, and based on that we chose the optimal number of neurons which gives the lowest RMSE of the validation set.

3. Row 256 - there is no explanation for removing the 2 points

Thank you for this comment. We fixed that in the text: ‘’Data analysis and processing mean that when deflections are measured on the primary and secondary roads and highways, recurring points which represent extremes and deviate significantly are removed.  In this way, they do not affect the data analysis.’’

4. Row 328 - in Table 6 - Serbian word "Radovi" was not translated into English

Thank you for this comment. It has been corrected

5. Row 361 - why only for Serbia?

This is a very good question and thank you Reviewer 1 for this. The reason why we say “only for Serbia” is because we only had the data from Serbia. Once the data from other countries are available, it will be possible to apply this model to other countries as well. This serves as a good idea to create a new paper using this model, but with data from another country/countries.

Reviewer 2 Report

1. The third author does not have a link with information about the employment institution and email.

2. Equations 1-9 and 11-14 should be listed in the text as equation 10 is listed.

3. Figure 4 should be listed in the text above.

4. The text in which figures 7-9 are listed should be above those figures.

5. There is nothing in the paper about the software ELMOD 6 for backcalculation of FWD measurements?

6. For the completeness of the topic of the article, it is necessary to show and describe the procedure of measuring the layers of the pavement structure with the FWD device.

Author Response

We would like to thank the Reviewer 2 for submitted comments and advice, and would like to provide further clarifications and answers to improve the quality of the paper and resolve ambiguities.

1. The third author does not have a link with information about the employment institution and email.

Thank you for this comment. It has been corrected.

2. Equations 1-9 and 11-14 should be listed in the text as equation 10 is listed.

Thank you for this comment. It has been corrected.

3. Figure 4 should be listed in the text above.

Thank you for this comment. It has been corrected.

4. The text in which figures 7-9 are listed should be above those figures.

Thank you for this comment. It has been corrected.

5. There is nothing in the paper about the software ELMOD 6 for backcalculation of FWD measurements?

Thank you for this comment. We added some sentences in the paper. “ELMOD 6 software was used for backcalculation. Three methods for calculating the asphalt modulus can be selected in this software. They include Finite Element Method, Linear Elastic Theory and Method of Equivalent Thicknesses. In this paper, Linear Elastic Theory was used, since in the Republic of Serbia there is no recommendation for using these three methods and British standards are used [40].”

6. For the completeness of the topic of the article, it is necessary to show and describe the procedure of measuring the layers of the pavement structure with the FWD device.

Thank you for this comment. The authors described in the paper something about the FWD, in the section from line 265 to line 269. We wrote that the input variables are deflections d₀, d₃₀₀, d₆₀₀, d₉₀₀, d₁₂₀₀, d₁₅₀₀ and d₁₈₀₀, i.e. measured deformations at the vertical distance of 0, 300, 600, 900, 1200, 1500, 1800 mm from the centre of the loading plate and the temperature of the upper surface of the asphalt layer. The output variable is the modulus of elasticity of asphalt layers (E_AC) expressed in MPa. Our opinion is that the focus is on machine learning and that is the reason why we did not provide a more detailed description of the FWD measuring procedure.

Reviewer 3 Report

This study evaluated three machine learning models for estimating the elasticity of the asphalt layer. Some comments regarding this study are shown as follows:

Line 256: Could the authors clarify the first and second categories of state roads? Do the authors mean primary and secondary roads?

 

Line 265: Do the authors mean the vertical distance? Please clarify it.

95% data was used for training, and 5% remaining data was used for testing and validating. The validation data set is too small compared to the training data. Since the three models didn't give a significant difference om the accuracy, the increase in the performance data set may result in different findings.

Compared to other similar researches, what kind of research significance can be identified for this study? Any significant advantages or improvements?

Author Response

We would like to thank the Reviewer 3 for submitted comments and advice, and would like to provide further clarifications and answers to improve the quality of the paper and resolve ambiguities.

1. Line 256: Could the authors clarify the first and second categories of state roads? Do the authors mean primary and secondary roads?

Thank you for this comment. Yes, we meant primary and secondary roads, it has been corrected in the text.

2. Line 265: Do the authors mean the vertical distance? Please clarify it.

Thank you for this comment. Yes, we meant the vertical distance, it has been corrected.

 

3. 95% data was used for training, and 5% remaining data was used for testing and validating. The validation data set is too small compared to the training data. Since the three models didn't give a significant difference om the accuracy, the increase in the performance data set may result in different findings.

The database for ANN modeling is divided in the following way: 85% for the training set, 10% for the validation set and 5% for the testing set. The database for SVM and BRT modeling is divided as follows: 95% for the training set and 5% for the testing set. The models were evaluated based on an identical testing set. The percentage values of the number of data in the test set, the modulus of elasticity E_AC, correspond to the percentage values of the training set in the given ranges shown in Figure6. Also, we used only 5% for the testing set because we had a lot of data for analyzing.

 

4. Compared to other similar researches, what kind of research significance can be identified for this study? Any significant advantages or improvements?

Compared to paper [25], the advantages of our work are:

The first advantage of this work is that a large database is defined for analysis. The methods analyzed in the paper have better prediction abilities than in the paper [25], the coefficient of determination R² between backcalculated and predicted E_AC value is higher than 0.93 for all models. The results of the analyzed methods in this paper do not have large deviations.

Other advantages over other papers [21, 24] are:

-The number of input data has been reduced. We did not take layer thicknesses as input data, which greatly simplifies the calculation if layer depths are obtained from exploratory pits,

- Analyses of modeling using artificial neural networks (ANN), support vector machines (SVM) and boosted regression tree (BRT) methods were analyzed and compared,

- We used primary and secondary roads and highways,

- We use Linear Elastic Theory for backcalcualtion in ELMOD 6 for calculated modulus of elasticity of asphalt layer,

- We did not use homogeneous sections for the input.

Round 2

Reviewer 3 Report

Comment 1: the authors mentioned that one of the most significant improvements in the proposed model is fewer inputs than other models (i.e., the thickness of the asphalt layer is not required). For the data set collected in this study, what's the range of thickness? Is this model applicable to the HMA overlay? Especially for thin HMA overlay?

Comment 2: Since this study consider the temperature as an important input in the model. The data measurement time and season should be briefly introduced in the section of Dataset.

 

Author Response

We would like to thank the Reviewer 3 again for submitted comments and advice, and would like to provide further clarifications and answers to improve the quality of the paper and resolve ambiguities.

1. The authors mentioned that one of the most significant improvements in the proposed model is fewer inputs than other models (i.e., the thickness of the asphalt layer is not required). For the data set collected in this study, what's the range of thickness? Is this model applicable to the HMA overlay? Especially for thin HMA overlay?

For the database used in this analysis, the thicknesses of the asphalt layers were measured every 1 km and they ranged from 125 to 210 mm. Measurements were made on flexible road constructions, with different types of binder (BIT and PmB) final layers. Therefore, the model can be applied to the HMA layer. Measurements were not made on pavement structures with a thin HMA layer, therefore the model is not applicable to these pavement structures.

2. Since this study consider the temperature as an important input in the model. The data measurement time and season should be briefly introduced in the section of Dataset.

Thank you for this comment. We added some sentences in the paper. “Deflections were mostly measured in spring and autumn in the early morning hours, except for one section on the highway, which was measured in the summer period in the morning hours.“ The reason why the measurement was made in the summer period is the closure of a single lane on the highway with a total length of 4 km.

Author Response File: Author Response.docx