Construction and Building Materials 189 (2018) 890–897

Contents lists available at ScienceDirect

Construction and Building Materials

journal homepage: www.elsevier.com/locate/conbuildmat

Use of random forests regression for predicting IRI of asphalt pavements

Hongren Gong a, Yiren Sun b, Xiang Shu a, Baoshan Huang c,a,⇑
a
The University of Tennessee, Knoxville, TN 37996, United States
b
Dalian University of Technology, China
c
Tongji University, China

h i g h l i g h t s

Introducing random forests to predict IRI.

Identifying variables critical to the evolution of IRI.

RFR significantly outperformed the linear regression regarding R2 and RMSE.

The initial IRI is the most critical factor to the evolution of IRI.

a r t i c l e i n f o a b s t r a c t

Article history: Random forest is a powerful machine learning algorithm with demonstrated success. In this study, the
Received 18 May 2018 authors developed a random forests regression (RFR) model to estimate the international roughness
Received in revised form 4 September 2018 index (IRI) of flexible pavements from distress measurements, traffic, climatic, maintenance and struc-
Accepted 6 September 2018
tural data. To validate the model, more than 11,000 samples were collected from the database of long-
Available online 18 September 2018
term pavement performance (LTPP) program, with 80% randomly sampled data for training and 20% of
them for testing the RFR model. The performance of the RFR model was then compared with that of
Keywords:
the regularized linear regression model. The results showed that the RFR model significantly outper-
Roughness
Pavement
formed the linear regression model, with coefficients of determination (R2 ) greater than 0.95 in both
Decision tree the training and test sets. The variable importance score obtained from the RFR revealed that the initial
Random forests IRI was the most important factor affecting the development of the IRI. In addition, the transverse crack-
Machine learning ing, fatigue cracking, rutting, annual average precipitation and service age had important influences on
Management the IRI. Other distresses such as longitudinal cracking, edge cracking, aggregate polishing, and potholes
LTPP exerted little impact on the evolution of the IRI.
Regression tree Ó 2018 Elsevier Ltd. All rights reserved.
Ride quality

1. Introduction and serviceability of pavements [1,3,4]. In effect, there have been

efforts made to relate the roughness to the structural capacity of
Pavement roughness is considered the most important indica- pavements [4,5]. Often, a huge amount of efforts and resources
tor of pavement condition by the driving public, especially for are needed to collect and monitor the roughness condition of pave-
pavements with high speed limits [1,2]. Assuring a reasonable level ment surfaces. According to Pierce et al.[6], depending on the func-
of pavement smoothness allows users to drive with comfort and tional class of roads, the surveys for roughness and other pavement
safety, but this requires highway agencies to monitor the perfor- performance indicators are conducted once or twice a year. Nowa-
mance of the entire road network frequent enough and adopt the days, even though it is relatively convenient and efficient to con-
most proper maintenance strategies where needed. In addition, duct performance survey with a high-speed automatic road
the roughness is often used as an critical performance indicator analyzer (ARAN) and other automatic distress collecting devices,
for pavement maintenance, as it represents the overall condition the reduction in the frequency of roughness survey will still result
in huge cost saving. The IRI, defined as the ratio of the accumulated
vehicle suspension motion to the distance it traveled during the
same time period, was proposed to provide a unified quantitative
⇑ Corresponding author at: The University of Tennessee, Knoxville, TN 37996,
United States.
method to compare roughness measured by different devices [1].
E-mail addresses: hgong2@vols.utk.edu (H. Gong), sunyiren@dlut.edu.cn The estimation of the roughness from distress data and other
(Y. Sun), xshu@utk.edu (X. Shu), bhuang@utk.edu (B. Huang). pavement information is not new, many attempts have been made

https://doi.org/10.1016/j.conbuildmat.2018.09.017
0950-0618/Ó 2018 Elsevier Ltd. All rights reserved.
 H. Gong et al. / Construction and Building Materials 189 (2018) 890–897 891

on this topic [7–15]. However, the predictive abilities of the models regression tree, the random forests are robust against over-fitting,
reported are still not of high applicability. Overall, this was caused have higher accuracy, and handle large data sets well [24,25].
either by using insufficient data samples or an inadequate model Although random forests have proven to be one of the most power-
formulation. Paterson[7] presented a linear regression model to ful machine learning algorithms, their applications in the field of
estimate the IRI from the cracking, rutting, pavement age, structure transportation and infrastructure engineering are limited. Through
number, equivalent single axle loads (ESAL), thickness of cracked a random forest classification model, Zaklouta et al.[26] achieved an
layers, and the number of potholes. Lin et al.[15] proposed a neural accuracy rate of 97.2% in traffic sign identification. Abdel-Aty et al.
network (NN) model to estimate the IRI from various distress [27] employed the integrated importance measure in the random
indicators using a total of 125 samples of data. The NN used in their forests to identify variables with significant association with their
study consists of three layers, including an input layer, a hidden binary target variable (crash or non-crash). Harb et al. [28] as well
layer and a output layer, each with 14, 6 and 1 nodes, respectively. utilized this feature to rank the impacts of drivers, vehicles, and
Based on the NN in the work of Lin et al. [15], Choi et al. [10] also environmental characteristics on precrash evasive actions.
employed the NN to predict the IRI using a dataset of 117 samples,
with 92 and 25 of them for training and testing the NN, 2.2. Long-term pavement performance program
respectively. The variables involved in their analyses consisted of
structure number, material parameters characterizing asphalt con- As the world’s largest and most comprehensive pavement per-
crete, thickness of asphalt layer and traffic data. Using a dataset of formance database, the Long-Term Pavement Performance (LTPP)
580 observations, Chandra et al.[16] employed the linear regres- program was established to study how pavement performance is
sion, nonlinear regression, and the NN to predict the IRI from rut- affected by various design predictors, environmental effects, traffic
ting, cracking, potholes, patches and raveling measurements. The loads, materials, construction quality, and maintenance practices
NN was reported to work better than the two regression methods, [29,30]. It is managed by the Federal Highway Administration
which has an R2 of 0.835 in the training set and 0.65 in the testing (FHWA), and has included pavement performance data of more
set. However, the NN models used in these studies tend to be over- than 2500 sections. The LTPP program consists of two complemen-
parameterized and are trained with very limited data, and thus tary programs, the general pavement study (GPS) and specific
suffer severely from overfitting issues [17,18]. pavement study (SPS). The GPS uses in-service pavement sections
Furthermore, in the Mechanistic-Empirical Pavement Design to study general performance of different pavement types, whereas
Guide (MEPDG), the IRI has been used as a critical performance cri- the SPS seeks to investigate the influence on performance of speci-
terion for pavement design, along with fatigue cracking (alligator fic predictors, such as drainage, layer thickness, rehabilitation or
and wheel-path longitudinal cracking), transverse cracking, and maintenance treatments [30]. The sections in the SPS are specifi-
rutting depth. In the MEPDG, the IRI prediction model is defined cally constructed for their own purposes. In general, each LTPP test
as a linear function of the initial IRI after construction, total area section is 152-meter long and 12-foot wide per lane. Since one of
of fatigue cracking (including alligator, longitudinal and reflective the major purposes of the LTPP program is to make high quality
cracking in the wheel path), length of transverse cracking, average data available for research, extensive studies have been exploring
rut depth, pavement age, average annual freezing index, average the LTPP database to understand the performance of various main-
annual precipitation, and plasticity index of soil [19]. This IRI pre- tenance and rehabilitation strategies [31–35], help adapt the
diction model was established with 1926 observations from the nationally calibrated performance models of the MEPDG to local
LTPP database with an R2 of 0.56, which indicates that it explains conditions [36–38], and optimize the maintenance decision
only 56% of the variation in the data and thus under-fits the data making process [39–44].
[19]. Similar performances of the locally calibrated IRI models were
also reported by Darter et al.[20] (R2 ¼ 0:67), Souliman et al. [21]
3. Data preparation
(R2 ¼ 0:675), and Guo and Timm [22] (R2 ¼ 0:686). Hence, a
method of higher predictive capacity for IRI estimation will benefit This study aims to develop a random forest regression (RFR)
both pavement management and design. model to predict the IRI from various types of distresses, pavement
structural information, traffic history, and climatic conditions. In
2. Background order to produce a model of good generalization ability, a large
dataset with sufficient observations is needed for training and test-
2.1. Random forests ing the RFR model. To this end, datasets with more than 12,300
samples of distress data, 28,700 samples of rutting data, and
Tree based algorithms including classification and regression 19,900 samples of IRI data for asphalt pavement from the most
trees (CART) have demonstrated success because of their excellence recent release of the LTPP database, Standard Data Release 31
in interpretation, visualization, and abilities to handle complex (SDR31) was used. The following gives a brief description for the
non-linearity between predictors and response [23]. However, it eighteen variables considered in this study:
also suffers from issues such as over-fitting, not robust to outliers,
and poor in predictive capacity compared with methods like the a) Structure
least square regression [24]. To tackle these problems, a common
Pavement total thickness (htot) in inches, including both
method is to aggregate the results of many trees; the random thicknesses of asphalt and underlying layers;
forests are one of the methods with good predictive performance.
Total thickness of asphalt layers (hac) in inches.
Random forests or random decision forests are ensemble learn- b) Performance
ing methods for classification, regression and other tasks. They work
fatigue: Fatigue cracking is a series interconnected
by constructing a group of randomly created decision trees and fore- cracks caused by the fatigue of an asphalt surface or a
casting the class that is the mode of the classes (classification) or the stabilized base under repeated traffic loading, in squared
mean (regression) of the individual trees. It is considered an meter (m2 ).
improvement to the bootstrap aggregation tree methods (bagging),
lwp and lnwp: Wheel path and non-wheel path longitu-
which use only bootstrapped samples for classification or regression dinal cracking, respectively, which are cracks parallel to
without predictor sampling. Compared with an individual the pavement centerline, in meters (m).
892 H. Gong et al. / Construction and Building Materials 189 (2018) 890–897

trans: Transverse cracking is a type of crack that extends
age: Time passed since the first roughness survey for
across the centerline, in meters (m). each section, in years.

edge: edge cracking only applies to pavements with c) Traffic
unpaved shoulders, in meters (m).
kesal: The annual average of equivalent single axle load

patch: It is a type of deterioration occurring in a patch, (ESAL) in the LTPP lane, in thousands (kesal). It is noted
which is an area where the original pavement has been that other traffic data could also be used, such as the
removed and replaced with either similar or different annual average daily traffic (AADT). However, the avail-
materials, in meters squared (m2 ). ability of the AADT is different from those of the distress,

ravel: Raveling is the wearing away of the pavement sur- rutting and IRI data, and only a small portion of the sec-
face caused by the dislodging of aggregate particles due tions selected in the study also have the AADT data.
to stripping and by the loss of asphalt binder due to Hence, the AADT was excluded from the modeling.
hardening. d) Climatic

block: Block cracks are those dividing the asphalt surface
freeze: Annual average freeze index is defined as the neg-
into approximately rectangular pieces, in meters squared ative of the sum of all average daily temperature below 0
(m2 ). °C in a year, in °C/day;

pothole: Potholes are bowl shaped holes of various sizes
precip: Annual average precipitation, in inches.
on the pavement surface, and they are caused by a bro-
ken pavement surface due to alligator cracking, localized In the LTPP database, the cracking, patching and potholes are
disintegration, or freeze–thaw cycles. recorded in three severity levels, namely, high, medium and low.

polish: The occurrence of polished aggregate, which For each type of these distresses, the values of different severity
shown as a portion of the aggregates extending above levels were added up. Regarding the IRI and rutting data, the mea-
the asphalt surface is either very small or without rough sured values are the average of the measurements from the left and
or angular particles to provide good skid resistance, in right wheel paths.
meters squared (m2 ). On most occasions, pavement surface defects such as cracking,

rut: A rut is a surface depression in the wheel path, and patches, potholes, and profile related indicators including IRI and
significant rutting can leads to structural failures and a rutting are collected with different devices, at different times, in dif-
potential for hydroplaning, in millimeters (mm). ferent frequencies, and recorded in different tables. For a section, if

IRI0 : The first IRI measurement after construction (IRI0 ), the survey frequencies of the IRI and rutting were found different
namely, the IRI measured at year 0, in m/km. The IRI0 is from those of the distresses, the IRI and rutting measurements were
considered as a critical parameter in the IRI transfer func- interpolated through regression to match those of the distresses.
tion of the MEPDG [45]. Fig. 1 presents a plot matrix for the variables considered, the

IRI: IRI value at a specific time of each section (average of lower triangle of the matrix shows scatter plots of each pair of
IRI measurements at right and wheel paths), in m/km. variables, while the upper triangle shows the correlations of the

1 2 3 4 0 400 800 0 300 700 0 10 20 0 10 20 500 1500

IRI

3
0.67 0.30 0.30 0.12 0.15 0.38 0.14 0.15 0.048 0.17 0.086

1
IRI0
3

0.21 0.067 0.069 0.051 0.17 0.02 0.25 0.066 0.15 0.022
1

rut

20
0.074 0.0041 0.015 0.029 0.06 0.0043 0.0099 0.031 0.021

0
fatigue
600

0.038 0.099 0.14 0.20 0.033 0.015 0.01 0.11

0

lwp
200

0.05 0.16 0.0046 0.051 0.033 0.062 0.084

0

lnwp
0 400

0.46 0.34 0.099 0.12 0.14 0.07

300

trans
0.25 0.049 0.047 0.20 0.12
0

age
15

0.011 0.047 0.051 0.028

0

2000

kesal
0.017 0.16 0.031
0

hac
15

0.014 0.0027
0

1500

freeze
0.16
0

precip
500

1 3 0 15 30 0 150 300 0 200 0 1500 0 1000

Fig. 1. A plot matrix for the variables considered. Scales of variables are the same as their definitions. Due to the limitation in space, some variables were omitted from this
plot.
 H. Gong et al. / Construction and Building Materials 189 (2018) 890–897 893

variables. The values of the correlation were mapped to the size of predictors, i.e., X 1 ; X 2 ; . . . ; X p , the procedure for solving this prob-
the text. It is seen that the IRI is highly correlated to the initial lem with a binary regression tree can be divided into four steps:
roughness (IRI0 ), while the distresses such as transverse cracking,
fatigue cracking, block cracking, longitudinal cracking (wheel and 1. Start with all cases in one region, which is the root node.
non-wheel paths) and rutting have strong impacts on the IRI. This 2. At each internal node of the tree, a test is applied to one of the
is in agreement with the IRI’s transfer function in the MEPDG. The predictor, X j .
traffic (ESAL), temperature (freeze index), and service age were 3. Depending on the outcome of the test, observations are
also highly correlated with the IRI. assigned to either the left or the right sub-region (branch) of
the tree.
4. Repeat Step 3 until a terminal node or leaf is reached, at which a
4. Random forests regression
prediction is made.

The random forest is an advancement of decision tree based

Fig. 2 shows a simple decision tree for illustration purposes. For
methods. To help explain the methodology of random forests, it
simplicity, this regression tree was grown with only three vari-
will be helpful to understand the concepts behind an individual
ables, including the IRI0 , trans and fatigue, and 1000 samples. As
decision tree.
Fig. 2 displays, the tree is in effect upside-down, with the root
being the top node and is growing downwards. Decision was made
4.1. Decision tree at each node until a terminal node or leaf is reached. Each leaf con-
tains a prediction. In this example, the top node assigned observa-
The decision tree is the theoretical basis of the random forests, tions that satisfies IRI0 < 1 to the left node, and this left node was
which has been one of the most widely used methods for inductive further split into two sub-nodes at IRI0 < 0:67. At the node that
inference [46]. A decision tree recursively splits the feature space met IRI0 < 0:67, two splits were made at the node trans < 14,
into a set of rectangular regions, and uses the mean or mode as and the prediction for observations falling into the right region (a
the prediction for the observations in the region [47,23]. Because leaf node) of this branch was 0.88. Likewise, for observations in
the rules used to split the feature space can be represented in a the right branch of the tree (IRI0 P 1), a further split was made
tree, it is also known as the decision tree method. For a regression at IRI0 < 1:5, and the prediction of those observations in the region
problem, observations with similar response values are split into with IRI0 > 1:5 was 2. As indicated in this tree, the initial rough-
the same region and a constant value (mean) is predicted within ness (IRI0 ) is the most important variable for estimating the IRI.
each region. Appropriate splitting variables and splitting points
are selected by minimizing the loss function, and for regression 4.2. Random forests
problems the mean squared error (MSE) is often used. Once the
loss function is minimized, the pair of splitting variable and point The RF is an advancement of the single classification and regres-
can be chosen. Assuming that Y is a response to be predicted by p sion trees (CART) method, which follows a simple non-parametric

Fig. 2. An illustration of the regression tree for IRI estimation (trans = transverse cracking; fatigue = alligator cracking; IRI0 =initial IRI).
894 H. Gong et al. / Construction and Building Materials 189 (2018) 890–897

regression approach [24]. The forest is developed by taking B boot- the RFR model, there are three major hyperparameters to deter-
strap versions of samples (with replacement), and meanwhile mine: the total number of trees (B), the number of variables (m)
drawing a random subset of m from the p predictors. The boot- in each tree, and number of samples to incorporate to grow each
strapping process is also called row subsampling, while the predic- tree (N). Chen and Guestrin [49] observed that the row subsam-
tors subsampling is called column subsampling. Both row and pling has smaller effect on the model performance than column
column subsampling help reduce the risk of overfitting and thus subsampling. Hence, in this study, the optimal N was determined
boost the model accuracy. In this study, the randomForest pack- after obtaining the optimal value for B and m. In the randomForest
age implemented in the statistical language, R, was used to develop R package, the default value for the m is p=3. In addition, the whole
the random forests model [48]. In this package, the default m-value data set (11,715 observations) was randomly split into two data
is one-third (1/3) of the total number of predictors (p), as the cre- sets, 80% (9372 observations) to train the RFR model, and 20%
ator of the RF algorithm Breiman [24] reported, with which for (2343 observations) for testing purpose.
most occasions is adequate to generate predictions of desirable
accuracy. For classification problems, Breiman [24] recommends 1 Xi
pffiffiffi MSE ¼ ðy  y^i Þ2
to use m ¼ q. n n i
Pn ð1Þ
The procedure for building a RF can be elaborated as follows: ðy  y^i Þ2
R2 ¼ 1  Pin i
2
i ðyi  yÞ
1. Assume there are B trees in the forest.
2. For each tree, draw a bootstrap sample of size N from the train- ^i is the predicted value, and y
where yi is the response, y  is the mean
ing set (row subsampling). of the response variable.
3. For a training set comprised of p predictors, m out of p predic- The MSE is the most frequently used metric for model assess-
tors are chosen at random as candidates for splitting (column ment in regression problems. This study used this metric along
subsampling).
with the R2 to assess the RFR’s goodness-of-fit and predictive per-
4. Pick the best variable and split-point among the m predictors,
formance. Eq. (1) gives a definition for the two metrics. To find the
and split each node into two sub-nodes.
best combination of hyperparameters for the RFR model, namely,
5. Predict new data by aggregating the predictions of the B trees,
the number of trees (B), the number of predictors (m) and samples
that is, averaging the predictions from each individual tree of
(N) in each tree, a grid search in combination with a cross valida-
the forest.
tion was used. To reduce the computation cost, the number of sam-
ples in each tree (N) was determined first by a separate grid search,
5. Model development and assessment which consists of using 50%, 75% and 100% of the samples. It is
found that the inclusion of all samples generates the smallest
5.1. RFR parameter optimizing MSE (Table 1). After choosing the N, another grid search with a
4-fold cross-validation was used to find the best combination of
In this study, the R package randomForest developed by Liaw B and m. Fig. 3 shows the results from this grid search. It is found
and Wiener [48] was used to develop the RFR model. In training that the forest grown with 500 trees and 11 predictors yield the
smallest MSE (4:52  104 ). It is also noted that when the number
Table 1
MSE for different number samples (B ¼ 500; m ¼ 11).
of predictors is greater than 11, the models began to overfit. Due to
the curse of dimensionality, models with more trees and predictors
Proportion of samples MSE
tend to have higher risks in overfitting the data. In addition, a sim-
50% 0.00573 pler model also reduces the computation cost. Hence, the final RF
75% 0.00177 model was constructed with 500 trees, 11 predictors and all sam-
100% 0.000447
ples (no row subsampling).

19 ● ● ● ● ● ● ●
MSE
15 ● ● ● ● ● ● 0.02 ●
Number of Predictors

0.04

11 ● ● ● ●
● ● ● ●

0.06

7 ● ● ● ● ● ● ●
0.08
5 ● ● ● ● ● ● ●

3 ● ● ● ● ● ● ●
1

10 50 100 200 500 1000 10000

Number of trees
Fig. 3. RFR model hyperparameters grid-search.
 H. Gong et al. / Construction and Building Materials 189 (2018) 890–897 895

5.2. Model assessment (IRI0 ). In other words, a pavement with a smoother surface in the
early stage tends to sustain the smoothness better. As to the trans-
Fig. 4 displays the predicted IRI values versus the measured ones verse cracking, it was found that it was more important to the
in both the training and testing sets. As indicated, the R2 in the development of pavement roughness than the fatigue and longitu-
training set is as high as 0.998, while it is 2% lower in the testing dinal cracks. As the IRI is the accumulated suspension motion of a
set (R2 ¼ 0:975). A lower R2 in the testing set is usually an indication vehicle, such as an ARAN, compared to the fatigue and longitudinal
cracking, the transverse cracking is more likely to be caught by the
of overfitting. However, given the high R2 and small MSE achieved,
ARAN and exert more impacts on the IRI. As expected, the fatigue
this is not a serious problem for the RFR model in the study.
cracking and rutting showed strong influence on the IRI. It was also
noted that the precipitation was critical to the IRI, and sections in
5.3. Importance of variables those areas with more precipitation tend to keep their smoothness
better. In addition, the traffic volumne (kESAL), temperature
As a regression method, the RFR is more predictive than the (freeze index), and age of pavements were significant factors to
individual regression tree. However, this was achieved at a cost the IRI as well.
of sacrificing the interpretability of the tree based methods. Similar
to a single regression tree, it is convenient to evaluate the variable
importance in the RFR. Fig. 5 shows the relative importance of vari- 6. Comparison of RFR with linear regression
ables in the RFR model. As shown, the initial roughness IRI0 is the
most important in predicting the IRI. The plot matrix presented in Fig. 6 gives the predicted IRI values from a linear regression
Fig. 1 also shows that the IRI is proportional to the initial roughness method versus the corresponding measurements. This simple

5 5
Training Data Testing Data
(R2 = 0.998, MSE = 5e-04) (R2 = 0.974, MSE = 0.006)
n = 9372 n = 2343
4 4
Predicted IRI (m/km)

Predicted IRI (m/km)

3 3

2 2

1 1

0 0
0 1 2 3 4 5 0 1 2 3 4 5
Actual IRI (m/km) Actual IRI (m/km)

Fig. 4. Predicted IRI using random forest.

100
Variable Importances

23.55

19.75

18.97

11.77

9.72

9.19

8.27

7.91

6.37

6.04
6.4

3.67

0.001
1.33

0.51

0.13

0.06
1.8

e e t e h ac h total -nwp g-wp atch avel Edge olish holes hove

I0
IR svers atigu Ru tation L x
Ag ESA Inde Bloc
k
P R
a n F i pi k eze ng Lon P Pot S
r
Pre
c Lo
T Fre

Fig. 5. Variable importance plot for random forests regression.

896 H. Gong et al. / Construction and Building Materials 189 (2018) 890–897

5 5
Training Data Testing Data
(R2 = 0.625, MSE = 0.0859) (R2 = 0.615, MSE = 0.0855)
n = 9372 n = 2343
4 4
Predicted IRI (m/km)

Predicted IRI (m/km)

3 3

2 2

1 1

0 0
0 1 2 3 4 5 0 1 2 3 4 5
Actual IRI (m/km) Actual IRI (m/km)

Fig. 6. Comparison of predicted and measured IRI by regularized linear regression.

model was selected because the transfer function for the IRI in the The RFR model presented herein is not limited to the prediction
MEPDG is also a linear function. To control the potential overfitting of the IRI, it can also be integrated into the Mechanistic-Empirical
issues in the linear regression, a regularized version of the linear Pavement Design (MEPDG) framework, and thus to generate more
regression method was used, namely, the ridge regression. The accurate pavement design. This is because the linear regression
ridge regression shrinks the regression coefficients by imposing a model used as a transfer function for the IRI in the MEPDG is often
penalty on their size, which is implemented by minimizing a incapable of capturing sufficient information and dealing with
penalized residual sum of squares [23]. The penalized coefficients nonlinearity between the predictors and response. Moreover, as
often contribute to reducing the variance and diminish overfitting, the IRI has long been used as a performance indicator in the pave-
and a detailed formulation of the ridge regression can be found in ment management system, the RFR model could also be adopted as
[23]. As indicated in Fig. 6, the R2 values in the training and testing an efficient device for maintenance decision making.
sets of the ridge regression model are 0.625, and 0.615, respec-
tively. A small R2 value is an indication of the underfitting, and also
implies that the model is inadequate to capture the variation in the References
data [23]. From this perspective, the ridge regression performed
much worse than the RFR. This is probably caused by the inability [1] M. Sayers, T. Gillespie, C. Queiroz, The international road roughness
experiment: a basis for establishing a standard scale for road roughness
of linear regression model in capturing the nonlinearity between measurements, Transp. Res. Rec. 1084 (1986) 76–85.
the IRI and the predictors. [2] Y.H. Huang, Pavement Design and Analysis, Pearson/Prentice Hall, 2004.
[3] L. Aultman-Hall, E. Jackson, C. Dougan, S.-N. Choi, Models relating
pavement quality measures, Transp. Res. Rec.: J. Transp. Res. Board 2004
7. Conclusions (1869) 119–125.
[4] G. Rada, R. Perera, V. Prabhakar, L. Wiser, Relating ride quality and structural
adequacy for pavement rehabilitation and management decisions, Transp. Res.
The present study proposed a new approach to predict the
Rec.: J. Transp. Res. Board 2304 (2012) 28–36.
international roughness index (IRI) from the distress, traffic, cli- [5] G. Sollazzo, T. Fwa, G. Bosurgi, An ANN model to correlate roughness and
mate, and pavement structure data. Using a large data set from structural performance in asphalt pavements, Constr. Build. Mater. 134 (2017)
the LTPP, a random forests regression (RFR) model was introduced 684–693.
[6] L.M. Pierce, G. McGovern, K.A. Zimmerman, Practical guide for quality
to establish the relationship between the roughness and various management of pavement condition data collection, Final Document
indicators. The performance of the RFR model was compared with DTFH61-07-D-00028, Appl. Pavement Technol. (2013).
that of a regularized linear regression. The critical variables affect- [7] W. Paterson, A transferable causal model for predicting roughness progression
in flexible pavements, Transp. Res. Rec. 1215 (1989) 70–84.
ing the development of roughness were determined. Based on the [8] Y. Wang, Ordinal logistic regression model for predicting AC overlay cracking, J.
analyses in the study, the following conclusions were drawn: Perform. Constructed Facil. 27 (3) (2012) 346–353.
[9] M. Irfan, M.B. Khurshid, S. Labi, Determining the service life of thin hot-mix
asphalt overlay by means of different performance indicators, Transp. Res. Rec.

The RFR has achieved excellent predictive performance in both 2108 (1) (2009) 37–45.
the training and testing sets, which was evidenced by the R2 in [10] J.-H. Choi, T.M. Adams, H.U. Bahia, Pavement roughness modeling using back-
the two data sets, 0.998 and 0.974, respectively. propagation neural networks, Comput.-Aided Civ. Infrastruct. Eng. 19 (4)
(2004) 295–303.

The comparison between the regularized linear regression [11] A. Ahmed, S. Labi, Z. Li, T. Shields, Aggregate and disaggregate
(ridge regression) and RFR indicated that the RFR outperformed statistical evaluation of the performance-based effectiveness of long-term
pavement performance specific pavement study-5 (LTPP SPS-5) flexible
the linear regression in terms of MSE and R2 . Moreover, the RFR pavement rehabilitation treatments, Struct. Infrastruct. Eng. 9 (2) (2013)
addressed the overfitting issue significantly better than the lin- 172–187.
ear regression model. [12] M. Irfan, M.B. Khurshid, S. Labi, W. Flora, Evaluating the cost effectiveness of
flexible rehabilitation treatments using different performance criteria, J.

The variable importance score of the RFR model showed that
Transp. Eng. 135 (10) (2009) 753–763.
the initial roughness was the most critical factor affecting the [13] J.N. Meegoda, S. Gao, Roughness progression model for asphalt pavements
development of the IRI. Also, the transverse cracking, rutting, using long-term pavement performance data, J. Transp. Eng. 140 (8) (2014)
block cracking, and fatigue cracking were strongly associated 04014037.
[14] P. Nitsche, R. Stütz, M. Kammer, P. Maurer, Comparison of machine learning
with the IRI. However, due to their lower occurrences, the methods for evaluating pavement roughness based on vehicle response, J.
patch, ravel, polishing and potholes had little impact on the IRI. Comput. Civ. Eng. 28 (4) (2012) 04014015.
 H. Gong et al. / Construction and Building Materials 189 (2018) 890–897 897

[15] J.-D. Lin, J.-T. Yau, L.-H. Hsiao, Correlation analysis between international [32] S. Labi, G. Lamptey, S. Konduri, K. Sinha, Part 1: Pavement management:
roughness index (IRI) and pavement distress by neural network, in: Paper analysis of long-term effectiveness of thin hot-mix asphaltic concrete overlay
Publication at the 82th Annual Meeting of the Transportation Research Board, treatments, Transp. Res. Rec.: J. Transp. Res. Board 2005 (1940) 1–12.
Washington, DC, 2003. [33] H.L. Von Quintus, A.L. Simpson, A.A. Eltahan, Rehabilitation of Asphalt
[16] S. Chandra, C.R. Sekhar, A.K. Bharti, B. Kangadurai, Relationship between Concrete Pavements: Initial Evaluation of the SPS-5 Experiment-Final Report,
pavement roughness and distress parameters for indian highways, J. Transp. Tech. Rep., 2006.
Eng. 139 (5) (2012) 467–475. [34] Y. Wang, The effects of using reclaimed asphalt pavements (RAP) on the long-
[17] C.M. Bishop, Pattern Recognition and Machine Learning, Springer, 2006. term performance of asphalt concrete overlays, Constr. Build. Mater. 120
[18] I. Goodfellow, Y. Bengio, A. Courville, Deep Learning, MIT press, 2016. (2016) 335–348.
[19] E.C.D. ARA Inc, Guide for Mechanistic-Empirical Design of New and [35] H. Gong, B. Huang, X. Shu, Field performance evaluation of asphalt mixtures
Rehabilitated pavement Structures, Final report, NCHRP 1–37A, containing high percentage of RAP using LTPP data, Constr. Build. Mater. 176
Transporation Research Board of the National Academies, Washington D.C., (2018) 118–128.
2004. [36] A. Banerjee, J. Aguiar-Moya, J. Prozzi, Texas experience using LTPP for
[20] M.I. Darter, L. Titus-Glover, H.L. Von Quintus, Implementation of the calibration of the MEPDG permanent deformation models, Transp. Res. Rec.:
Mechanistic Empirical Pavement Design Guide in Utah: Validation, J. Transp. Res. Board 2094 (2009) 12–20.
Calibration, and Development of the UDOT MEPDG User’s Guide, Tech. Rep., [37] K. Hall, D. Xiao, K. Wang, Calibration of the mechanistic-empirical pavement
2009. design guide for flexible pavement design in Arkansas, Transp. Res. Rec.: J.
[21] M.I. Souliman, M.S. Mamlouk, M.M. El-Basyouny, C.E. Zapata, Calibration of the Transp. Res. Board 2226 (2011) 135–141.
AASHTO MEPDG for flexible pavement for arizona conditions, in: Proceedings [38] H. Gong, B. Huang, X. Shu, S. Udeh, Local calibration of the fatigue cracking
of the Transportation Research Board 89th Annual Meeting, Transportation models in the Mechanistic-Empirical Pavement Design Guide for Tennessee,
Research Board Washington, DC, USA, 22, 2010. Road Mater. Pavement Des. (2017) 1–9.
[22] X. Guo, D.H. Timm, Local calibration of MEPDG using national center for [39] W.T. Chan, T. Fwa, K.Z. Hoque, Constraint handling methods in pavement
asphalt technology test track data, in: Transportation Research Board 94th maintenance programming, Transp. Res. Part C: Emerg. Technol. 9 (3) (2001)
Annual Meeting, 15-1032, 2015. 175–190.
[23] J. Friedman, T. Hastie, R. Tibshirani, The elements of statistical learning, vol. 1, [40] K.T. Hall, C.E. Correa, A.L. Simpson, Performance of flexible pavement
Springer series in statistics, New York, 2001. maintenance treatments in the LTPP SPS-3 experiment, Transp. Res. Rec.: J.
[24] L. Breiman, Random forests, Mach. Learn. 45 (1) (2001) 5–32. Transp. Res. Board 1823 (03–4526) (2003) 47–54.
[25] T.K. Ho, Random decision forests, in: Document Analysis and Recognition, [41] S. Labi, K.C. Sinha, Effectiveness of highway pavement seal coating treatments,
1995, Proceedings of the Third International Conference on, vol. 1, IEEE, 1995, J. Transp. Eng. 130 (1) (2004) 14–23.
pp. 278–282. [42] Y. Wang, G. Wang, Y. Ahn, Impact of climate conditions on effectiveness of
[26] F. Zaklouta, B. Stanciulescu, O. Hamdoun, Traffic sign classification using kd asphalt pavement preservation techniques, Transp. Res. Rec.: J. Transp. Res.
trees and random forests, in: Neural Networks (IJCNN), in: The 2011 Board 2292 (2012) 73–80.
International Joint Conference on, IEEE, 2011, pp. 2151–2155. [43] H. Gong, Q. Dong, B. Huang, X. Jia, Effectiveness analyses of flexible pavement
[27] M. Abdel-Aty, A. Pande, A. Das, W. Knibbe, Assessing safety on Dutch freeways preventive maintenance treatments with LTPP SPS-3 experiment data, J.
with data from infrastructure-based intelligent transportation systems, Transp. Eng. 142 (2) (2015) 04015045.
Transp. Res. Rec.: J. Transp. Res. Board 2083 (2008) 153–161. [44] X. Chen, H. Zhu, Q. Dong, B. Huang, Optimal thresholds for pavement
[28] R. Harb, X. Yan, E. Radwan, X. Su, Exploring precrash maneuvers using preventive maintenance treatments using LTPP data, J.Transp. Eng. Part A:
classification trees and random forests, Accident Anal. Prev. 41 (1) (2009) 98– Syst. 143 (6) (2017) 04017018.
107. [45] AASHTO, Mechanistic-Empirical Pavement Design Guide-A Manual of Practice,
[29] W.O. Hadley, SHRP-LTPP overview: five-year report, Tech. Rep. SHRP-P-416, first edn., 2008.
Strategic Highway Research Program, National Research Council, 2101 [46] T.M. Mitchell et al., Machine learning, WCB, 1997.
Constitution Avenue N.W., Washington, DC 20418, 1994. [47] L. Breiman, J. Friedman, R. Olshen, C. Stone, Classification and regression trees.
[30] FHWA, The Long-Term Pavement Performance Program, Tech. Rep. FHWA- [48] A. Liaw, M. Wiener, Classification and regression by randomForest, R News 2
HRT-15-049, Federal Highway Administration, Turner-Fairbank Highway (3) (2002) 18–22. URL http://CRAN.R-project.org/doc/Rnews/.
Research Center, 6300 Georgetown Pike, McLean, VA 22101, 2015. [49] T. Chen, C. Guestrin, Xgboost: a scalable tree boosting system, in: Proceedings
[31] K. Hall, C. Correa, A. Simpson, Performance of flexible pavement rehabilitation of the 22nd ACM SIGKDD International Conference on Knowledge Discovery
treatments in the long-term pavement performance SPS-5 experiment, Transp. and Data Mining, ACM, 2016, pp. 785–794.
Res. Rec.: J. Transp. Res. Board 1823 (2003) 93–101.