IRI Prediction using Machine Learning Models
ANKIT SHARMA, PRAVEEN AGGARWAL
Department of Civil Engineering,
National Institute of Technology Kurukshetra, Haryana-136119,
INDIA
Abstract: - Road infrastructure is the backbone of the economy of any country. The recent increase in the
length of roads has never been matched in history. The increase in length comes with huge demand for the
maintenance of pavements in an orderly fashion. The pavement management system is used for planning
maintenance based on pavement performance evaluation. The international roughness index (IRI) is considered
a standard parameter for the functional evaluation of flexible pavements. In the present study, IRI is predicted
through machine learning models using the LTPP database. The main objective of the study is to find the
optimal machine learning which can be used for IRI prediction. Three machine learning models, (i) linear
regression, (ii) optimised trees, and (iii) optimised Gaussian process regression (GPR), has been used for
predicting IRI. Different models have been compared based on various statistical parameters. The optimised
GPR model performed best per the R-Squared value (0.89).
Key-Words: - International roughness Index, Pavement performance evaluation, Machine learning
Received: June 22, 2022. Revised: May 12, 2023. Accepted: June 15, 2023. Published: July 10, 2023.
1 Introduction
India is a vast country with a huge road network
with a total road length of 63.86 lakh kilometers in
2019. The road length has increased at a compound
annual growth rate of 4.2% since 1951. Rural roads
constitute a major portion of the road network in
India (approximately 71%). The huge network of
rural roads needs regular maintenance. Hence
pavement evaluation is vital for decision-makers to
preserve the road infrastructure and to maximise its
benefits to the public.
Researchers have conducted multiple studies to
conclude the importance of distress in quantifying
pavement quality, as the road length is increasing
per year at a fast pace. The already available
infrastructure needs to be simultaneously evaluated
for maintenance requirements. The limited budgets
and increasing public expectations stressed the
decision-makers in adopting a temporary solution to
pavement distress. The inter-correlation between the
pavement distresses is often ignored. The
correlation analysis of multiple pavement sections
will help understand the requirement for treating
underlying problems. The decision-makers can then
provide adequate treatment accordingly. An ideal
Pavement management system (PMS) should hence
integrate the evaluation of all distresses
simultaneously.
Visual inspection of various pavement distresses
is adopted by in-field engineers for assessment of
the current condition of the pavement. This ought to
be a very initial part of the pavement evaluation and
needs quantitative measurements of ride quality/
pavement unevenness for judging the maintenance
requirement of the road. The international roughness
index (IRI) is used in evaluating the ride quality of
the road. The IRI has been widely adopted by
transportation agencies worldwide as a key indicator
for evaluating pavement ride quality. However, the
measurement of IRI is costly and unaffordable for
highway agencies limited by their fiscal resources.
The recent advent of machine learning methods
has enabled researchers to explore the applicability
of machine learning field to their respective fields.
Many past studies have been performed for the
evaluation of the applicability of machine learning
in the transportation engineering field, [1], [2], [3],
[4].
In this study, data has been collected for 211
pavements from the LTPP database, and machine
learning methods, i.e., linear regression, Optimised
trees, and Optimised gaussian process regressions
(GPR), have been explored. The results have been
compared based on different model performance
matrices, i.e., Root Mean Square Error (RMSE), R-
squared, Mean Square Error (MSE), and Mean
Average error (MAE). MATLAB computer
program, [5], has been used in this study to achieve
the desired results. The optimised GPR performed
WSEAS TRANSACTIONS on COMPUTER RESEARCH
DOI: 10.37394/232018.2023.11.10
Ankit Sharma, Praveen Aggarwal
E-ISSN: 2415-1521
111
Volume 11, 2023
the best in terms of all the model’s performance
values.
The manuscript is arranged into 5 Chapters in the
order: Chapter 1: Introduction, Chapter 2: Data
Preparation, Chapter 3: Machine Learning Models,
Chapter 4: Methodology, Chapter 5: Results, and
Chapter 6: Conclusions.
2 Data Preparation
The data is extracted from the LTPP infopave
database provided by Federal Highway
Administration (FHWA). The database is accessible
on the internet [6]. Table 1 describes the various
variables retrieved for conducting the study. The
data were extracted for all the pavement sections
with flexible pavement surfaces.
Table 1. Name and description of different variables
under consideration
NAME OF
VARIABLE
DESCRIPTION OF VARIABLE
STRUCTURAL
NUMBER
The structural number of asphalt
concrete pavements.
HPMS16_CRA
CKING_PERC
ENT_AC
Total percentage of sections cracked
by HPMS field guide definitions.
MEPDG_CRA
CKING_PERC
ENT_AC
This variable represents the sum of
alligator cracks in percentage, per
MEPDG, [7].
MEPDG_CRA
CKING_LENG
TH_AC
This variable defines transverse cracks
per meter run of the road via MEPDG
definitions.
MEPDG_LON
G_CRACK_LE
NGTH_AC
The length of longitudinal cracks
conforms to the wheel path per meter
length.
ME_PERCENT
_WHEEL_PAT
H_CRACK
Percent of area cracked in 0.61-meter
wide wheel paths as per MEPDG
crack percent.
kesal_year
It defines the computed ESAL
(Equivalent Single Wheel Load) for
the kth year.
unbound
granular base
thickness
average
The average thickness of the unbound
granular base layer.
asphalt concrete
thickness
average
The average thickness of the asphalt
concrete layer.
MEAN_ANN_
TEMP_AVG
It defines the average mean
temperature for the whole year.
TOTAL_ANN_
PRECIP
The sum of precipitation values for
the whole year.
initial IRI for
pavement
IRI value for the pavement in the year
of construction.
AGE
This variable defines the age of the
pavement in years.
IRI
International Roughness Index
The data extracted has been pre-processed using
Microsoft Excel. The outliers were identified and
removed to maintain the uniformity of the data. The
scatter plots have been prepared for all the
independent variables with dependent variables
(Figure 1).
0,5
1
1,5
2
2,5
3
0 2 4 6 8
IRI
STRUCTURAL NUMBER
0,5
1
1,5
2
2,5
3
0 20 40 60
IRI
HPMS16_CRACKING_PERCENT_AC
WSEAS TRANSACTIONS on COMPUTER RESEARCH
DOI: 10.37394/232018.2023.11.10
Ankit Sharma, Praveen Aggarwal
E-ISSN: 2415-1521
112
Volume 11, 2023
Fig. 1: Variation of IRI with different parameters
under consideration
3 Machine Learning Models
In the present study, three machine models avail in
MATLAB are used.
3.1 Linear Regression
Linear regression models describe the relationship
between the dependent variable y and independent
variables X. The matrix X is usually named as
“Design matrix” of independent variables. The
dependent variable is also named the response
variable. A general equation for any line regression
model has been given below (equation 1).
yi=β0+β1Xi1+β2Xi2++βpXip (1)
, i = 1 , … , n
where
n, represent the number of observations
yi, represents the ith response
β0 represents the constant term in the
linear regression model
β1, β2, … βn represents the coefficients
for all the independent variables
Xi1, Xi2, … , Xip represent different
independent variables for the ith
response.
The method of least squares is used in this study
to develop regression models. The method of least
squares is used to determine the best-fit line for the
independent variables.
3.2 Optimised Trees
In this study, optimised trees were used for the
modelling. Decision trees are generally used for
classification tasks in the machine-learning field.
However, for regression analysis of the problem at
hand, they can be used by converting the available
data into small classification tasks. Figure 2
represents a typical decision tree. In this study, the
decision trees have been optimised using Bayesian
optimisation and hence are here referred to as
optimised trees.
The decision of the models is based on the rules
as defined in Figure 2. The figure describes the
route to follow to determine if a person is fit. In the
first step arithmetic operation on the age of the input
variable age determines which route the model has
to follow. Step 2, based on the input response
model, determines if the input variable to be
compared is “Eats a lot of pizza” or “Exercises in
the morning”. The answer to the questions decides if
the person is fit or unfit.
2
0,5
1
1,5
2
2,5
3
0 500 1000
IRI
kesal_year
0,5
1
1,5
2
2,5
3
0,5 1,5 2,5
IRI
initial IRI for pavement
2
0,5
1
1,5
2
2,5
3
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14
IRI
AGE
WSEAS TRANSACTIONS on COMPUTER RESEARCH
DOI: 10.37394/232018.2023.11.10
Ankit Sharma, Praveen Aggarwal
E-ISSN: 2415-1521
113
Volume 11, 2023
Fig. 2: Example decision tree
3.3 Gaussian Process Regression
It is a non-parametric Bayesian approach to solving
regression problems. It is able to widely capture the
relationship between predictors and responses. The
complexity of the data is mapped using the kernel
functions. The kernel function quantifies the
similarity between the input variables. The
similarity between the variables is used to generate
the covariance matrix. The covariance defines how
the output of the model is correlated to the input
variables. Bayesian optimisation has been used for
choosing the best hyperparameters for which the
model performs best.
3.4 Optimisation
Optimisation of a machine learning model has to be
performed when the parameters defined for the
model have large sets, and many possible
combinations of the hyperparameters are possible.
The Bayesian optimisation technique is used to look
for the best possible values of the input
hyperparameters for the machine learning model.
This optimisation is preferred when the selection of
hyperparameters is huge, and it is not
computationally possible to find the best-performing
hyperparameters in the polynomial time. One of the
key benefits of Bayesian optimisation ought to be its
capability to balance the trade-off between
exploration and exploitation in the available solution
space. It uses a probabilistic model of the function
to guide the search for optimal solution and balance
the exploration of new inputs and exploitation of
known good inputs.
4 Methodology
The application of machine learning algorithms has
been done in the MATLAB program. The
descriptive statistics of the variables were analysed
to find out the different machine learning algorithms
that can be applied to predict IRI accurately. The
Regression learner has been used for the application
of linear regression, optimised trees, and optimised
GPR. To validate the results, the Cross-validation
method of validation has been used along with 5-
fold cross-validation. The data was divided into five
equal but random parts by the machine learning
models and was used for improving the training of
the model.
The Bayesian optimisation technique was
selected for the optimisation of the hyperparameters
for different models. The trained models were then
compared based on the model performance indices.
5 Results
The performance of all the models has been
compared to each other in terms of performance
parameters as discussed in, [8], [9], [10], [11]:
Root Mean Square Error (RMSE):
(2)
Mean Average Error (MAE):
(3)
R-squared:
(4)
Mean Square Error (MSE):
(5)
= predicted values by the model
= observed values for those inputs
n= total number of observations
= average of the measured values.
Table 2. depicts the parameters for the developed
models
RMSE
R-squared
MSE
MAE
Linear
regression
0.1976
0.83
0.039
0.13
Optimised
Trees
0.2182
0.80
0.047
0.14
Gaussian
Process
Regression
0.1576
0.89
0.024
0.09
From Table 2, it can be observed that the
optimised trees have performed the lowest of all the
models trained. The Classification and regression
trees algorithm, [12], performs better when the data
is not complex, and the data to be modelled is
categorical variables. In the dataset of pavement, the
dataset is in tabular format, and optimised trees are
hence performing badly in performance parameters.
WSEAS TRANSACTIONS on COMPUTER RESEARCH
DOI: 10.37394/232018.2023.11.10
Ankit Sharma, Praveen Aggarwal
E-ISSN: 2415-1521
114
Volume 11, 2023
The scatter plot for actual and predicted values from
the validation results from linear regression has
been displayed in Figure 3 (a).
The linear regression model performed
better than optimised trees with an r-squared value
(0.83). The root mean square value for the model
was 0.1976. actual vs predicted scatter plot for the
validation dataset has been shown in Figure 3(a).
The optimised GPR model performed best with
an R-squared value of 0.89. The performance of the
model was at par with the previous studies
conducted in the prediction of [3], [13], [14], [15].
Figure 3(c) represents the scatter plot between
actual and predicted IRI values from the optimised
GPR model.
(a) Actual vs predicted linear regression
(b) Actual vs predicted Optimised trees
(c) Actual vs predicted for Optimised GPR
Fig. 3: True versus predicted IRI value
6 Conclusion
This study analyses the applicability of machine
learning models to the prediction of the road
roughness of the pavement using various input
parameters, i.e. structural number,
HPMS16_CRACKING_PERCENT_AC,
MEPDG_CRACKING_PERCENT_AC,
MEPDG_CRACKING_LENGTH_AC,
MEPDG_LONG_CRACK_LENGTH_AC,
ME_PERCENT_WHEEL_PATH_CRACK,
kesal_year, unbound granular base thickness
average, asphalt concrete thickness average,
MEAN_ANN_TEMP_AVG,
TOTAL_ANN_PRECIP, initial IRI for pavement,
and AGE. The different models that were utilised in
this study are linear regression, optimised trees, and
optimised GPR. The different performance indices
used are RMSE, R-square, MAE, and MSE. The
following conclusions are drawn from the study and
listed below:
1. The use of the Bayesian method yields machine
learning models that perform at par with past
studies. The use of Bayesian optimisation enables
the user to automatically search for optimised
hyperparameters for the model based on the
performance indices.
2. In terms of the R-squared value of the
predictions, the model’s optimised trees performed
the lowest, and optimised GPR performed the best.
WSEAS TRANSACTIONS on COMPUTER RESEARCH
DOI: 10.37394/232018.2023.11.10
Ankit Sharma, Praveen Aggarwal
E-ISSN: 2415-1521
115
Volume 11, 2023
3. From the scatter plot of actual vs predicted, it is
evident that the predictions by optimised GPR are
close to the actual observed values for IRI. Hence
optimised GPR model can be effectively used for
accurate prediction of IRI using given parameters.
References:
[1] Inkoom S, Sobanjo J, Barbu A, Niu X (2019)
Pavement Crack Rating Using Machine
Learning Frameworks: Partitioning, Bootstrap
Forest, Boosted Trees, Naïve Bayes, and K -
Nearest Neighbors. J Transp Eng Part B
Pavements 145:04019031.
https://doi.org/10.1061/JPEODX.0000126.
[2] Marcelino P, de Lurdes Antunes M, Fortunato
E, Gomes MC (2020) Transfer learning for
pavement performance prediction. Int J
Pavement Res Technol 13:154–167.
https://doi.org/10.1007/s42947-019-0096-z.
[3] Zeiada W, Dabous SA, Hamad K, et al (2020)
Machine Learning for Pavement Performance
Modelling in Warm Climate Regions. Arab J
Sci Eng 45:4091–4109.
https://doi.org/10.1007/s13369-020-04398-6.
[4] Karballaeezadeh N, Danial MS, Moazemi D,
et al (2020) Smart structural health monitoring
of flexible pavements using machine learning
methods. Coatings 10:1–18.
https://doi.org/doi.org/10.20944/preprints2020
04.0029.v1.
[5] The MathWorks Inc. (2022). MATLAB
version: 9.12.0.2039608 (R2022a) Update 5,
Natick, Massachusetts: The MathWorks Inc.
https://www.mathworks.com
(Accessed: March 25, 2019).
[6] E. Elkins G, Ostrom B (2019) Long-Term
Pavement Performance Information
Management System User Guide,
LTPP_IMS_USER_GUIDE_2019_V8.pdf ,
Available at :
https://infopave.fhwa.dot.gov/InfoPave_Repo
sitory/files/LTPP_IMS_USER_GUIDE_2019
_V8.pdf.
[7] Transportation Officials. (2008). Mechanistic-
empirical pavement design guide: a manual of
practice. AASHTO, available at :
https://fenix.tecnico.ulisboa.pt/downloadFile/
563568428712666/AASHTO08.pdf
(Accessed: March 30, 2019).
[8] Chai T, Draxler RR (2014) Root mean square
error (RMSE) or mean absolute error (MAE)?
-Arguments against avoiding RMSE in the
literature. Geosci Model Dev.
https://doi.org/10.5194/gmd-7-1247-2014.
[9] Botchkarev A (2019) A New Typology
Design of Performance Metrics to Measure
Errors in Machine Learning Regression
Algorithms. Interdiscipl. J Information,
Knowledge, Manag 14:045–076.
https://doi.org/10.28945/4184.
[10] Krijnen WP (2006) Some results on mean
square error for factor score prediction.
Psychometrika. https://doi.org/10.1007/s
11336-004-1220-7.
[11] Miles J (2014) R Squared, Adjusted R
Squared. In: Wiley StatsRef: Statistics
Reference Online
[12] Ziegler A, König IR (2014) Mining data with
random forests: Current options for real-
world applications. Wiley Interdiscip Rev
Data Min Knowl Discov.
https://doi.org/10.1002/widm.1114.
[13] Sharma A, Sachdeva SN, Aggarwal P (2021)
Predicting IRI Using Machine Learning
Techniques. Int J Pavement Res Technol.
https://doi.org/10.1007/s42947-021-00119-w
[14] Sollazzo G, Fwa TF, Bosurgi G (2017) An
ANN model to correlate roughness and
structural performance in asphalt pavements.
Constr Build Mater 134:684–693.
https://doi.org/10.1016/j.conbuildmat.2016.12
.186.
[15] Abdelaziz N, Abd El-Hakim RT, El-Badawy
SM, Afify HA (2020) International
Roughness Index prediction model for
flexible pavements. Int J Pavement Eng
21:88–99.
https://doi.org/10.1080/10298436.2018.14414
14.
Contribution of Individual Authors to the
Creation of a Scientific Article (Ghostwriting
Policy)
-Ankit Sharma collected data and carried out model
preparation
-Praveen Aggarwal helped in finalising the
manuscript
Sources of Funding for Research Presented in a
Scientific Article or Scientific Article Itself
No external source of funding
Conflict of Interest
The authors have no conflict of interest to declare.
Creative Commons Attribution License 4.0
(Attribution 4.0 International, CC BY 4.0)
This article is published under the terms of the
Creative Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en
_US
WSEAS TRANSACTIONS on COMPUTER RESEARCH
DOI: 10.37394/232018.2023.11.10
Ankit Sharma, Praveen Aggarwal
E-ISSN: 2415-1521
116
Volume 11, 2023
