Construction and Building Materials 116 (2016) 281–288

Contents lists available at ScienceDirect

Construction and Building Materials

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

Performance models for hot mix asphalt pavements in urban roads

Ufuk Kırbasß a,⇑, Mustafa Karasßahin b
a
Ondokuz Mayis University, Department of Civil Engineering, Samsun, Turkey
b
Istanbul University, Department of Civil Engineering, Istanbul, Turkey

h i g h l i g h t s g r a p h i c a l a b s t r a c t

Pavement deterioration models were

developed for HMA paved road
sections in urban roads.

The models used were deterministic
regression analysis, MARS and ANN.

The ANN method was found to be the
most appropriate model for
predicting deterioration.

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

Article history: The correct budget allocation for road maintenance, which represents a significant infrastructure invest-
Received 23 November 2015 ment in urban roads, requires the accurate prediction of the deterioration of bituminous hot mix asphalt
Received in revised form 14 April 2016 (HMA). In this study, three different deterioration models have been developed that can predict the
Accepted 26 April 2016
future performance of pavements in urban HMA paved roads. First, the current condition of the pave-
Available online 4 May 2016
ments was measured by using the pavement condition index (PCI), which is approved by the PAVER sys-
tem. Then, three different models were developed to predict deterioration in the PCI as a function of
Keywords:
pavement age, and applied to urban road networks in Samsun (Turkey). The models used were determin-
Flexible pavement
Deterioration model
istic regression analysis, multivariate adaptive regression splines (MARS) and artificial neural networks
Performance prediction (ANN). Variations of each model were explored and the one with the highest computational efficiency
PCI was employed for ranking pavement sections with respect to rehabilitation needs. Results indicated that
Hot mix asphalt the three approaches had comparable prediction accuracies and R-squared values, although predictions
provided by the ANN model were more accurate compared with the other models. The article provides a
detailed comparison of the performance of the three models.
Ó 2016 Elsevier Ltd. All rights reserved.

1. Introduction

Pavement management, in its broadest sense, encompasses all

⇑ Corresponding author at: Ondokuz Mayis University, Department of Civil
Engineering, Kurupelit Campus, 55139 Atakum/Samsun, Turkey.
E-mail addresses: ufukkirbas@gmail.com (U. Kırbasß), mkarasahin@istanbul.edu.
tr (M. Karasßahin).
the activities involved in planning, design, construction, mainte-
nance, evaluation, and rehabilitation of the pavement portion of
a public works program [1]. The Pavement Management System
(PMS) concept that has been in place since the 1950 s has been
adapted over time in line with new innovations and technological

http://dx.doi.org/10.1016/j.conbuildmat.2016.04.118
0950-0618/Ó 2016 Elsevier Ltd. All rights reserved.
282 U. Kırbasß, M. Karasßahin / Construction and Building Materials 116 (2016) 281–288
developments. In the 1990 s in particular, several books and
research reports were published explaining the theory of the
PMS concept, its levels, and its data collection and assessment
techniques. This period also saw the publication of a PMS standard,
known as the American Society Testing and Materials (ASTM)
E1889-97(2009). Due to limitations in annual budgets for the exe-
cution of pavement rehabilitation projects, the authority must
choose first those alternatives with highest priority, in order to
supply the demands of optimization in the network [2]. As a key
component of pavement management, pavement performance
models play a crucial role in PMS, where forecasting results pro-
vide key information for highway agencies to make decisions on
the overall maintenance and budget plan [3]. The effectiveness of
the strategies is therefore highly dependent on the accuracy of
the predictions [4]. The development of good models for predicting
performance, in terms of some indexes versus age or accumulated
axle load applications, has been a major challenge for pavement
engineers [5].
Information on several pavement condition characteristics is
critical to performing management functions. The characteristics
include roughness, skid resistance, structural capacity, and distress
[6]. The performance of a pavement, as presented herein, is mea-
sured by the PCI, which indicates the present condition of pave-
ment in terms of structural integrity and surface operational
condition [7]. The PCI provides an indication of the current perfor-
mance of a pavement using pavement distress data, which includes
different distress types, severity levels, and densities, and takes the
form of a numerical rating, with 0 being the worst possible condi-
tion and 100 being the best. This study made use of PAVER soft-
ware to obtain PCI values for twenty road sections, which
experience different volumes of traffic, in the Samsun region. This
study aimed to develop and compare deterioration models using
three different modeling techniques (deterministic regression
analysis, MARS and ANN) for HMA paved urban roads. These mod-
els are to be used in the planning resurfacing maintenance priori-
ties at network level and to forecast the outcomes under different
maintenance strategies and budget scenarios [8].
2. Review of deterioration models
The accuracy of these models depends on the type of algo-
rithms, number of effective variables, number of measured data
and accuracy of the instruments and measurements. Various algo-
rithms and statistical methods have been developed to predict
pavement condition [9].
The main factors influencing the loss of performance (service-
ability) of a pavement are traffic loads, age, and climate effects
(change in temperature and humidity) [1]. The age of the pave-
ment in a road section is the single most important variable in
explaining the loss of performance of the pavement, because it also
serves as a summary indicator of all traffic and climate effects. This
is why there are many studies in the literature, which model pave-
ment deterioration over time, using deterministic and stochastic
methods. The methods used in these studies include, linear or
non-linear regression [10–13], autoregression and time series
[4,14] methods, MARS method [15], artificial intelligence modeling
techniques, specifically the ANN method [16], and stochastic meth-
ods, and the Markov chain [17–19] and Bayesian theory [16,20] in
particular [21]. Research usually focuses on the development of
models that can predict deterioration over time in an index, such
as the PCI, which measures the current condition of the pavement,
the international roughness index (IRI), or the pavement service-
ability index (PSI). Most of these studies use data collected within
the framework of the Long-Term Pavement Performance (LTPP)
Program, run by the Federal Highway Administration (FHWA),
which contains data on more than 2000 pavement sections, mostly
from the United States, that has been collected over a period of
more than 20 years and shared with researchers. There are also
studies that compare results obtained using the multiple deter-
ministic and/or stochastic modeling approaches [8,9].
Most researchers focus on forecasting models that use, as
inputs, multiple independent variables developed with the pur-
pose of predicting the future value of a pavement performance
index. Some of the independent variables used among others,
include the age of the pavement, environmental factors, different
types and severities of pavement cracks, patching, pavement defor-
mations of different types and severities, traffic loads, pavement
design, pavement thickness, and rut depth. Because there are many
variables that potentially affect pavement performance, and
because some of the variables are not quantifiable, most forecast-
ing models use artificial intelligence methods. Those frequently
used are, Fuzzy logic [22,2,23] and ANN [3,24,25], which facilitate
the combined use qualitative and quantitative data, and ANFIS
[26], which is a combination of these two methods. MARS, which
facilitates modeling the relationship between multiple indepen-
dent variables, is another method preferred by researchers [27].
On the other hand, there are also studies that use linear or non-
linear regression methods to predict pavement performance [28–
30].
The World Bank’s Highway Design and Maintenance Models
(HDM) are comprehensive models for prioritizing pavement main-
tenance and rehabilitation. These models simulate total life-cycle
conditions and costs, and provide economic decision criteria for
multiple road design and maintenance alternatives [31]. Each
country has unique local conditions, including climatic, technolog-
ical, operational and traffic conditions, as well as differing stan-
dards and measuring units; therefore, the adaptation and
calibration of pavement deterioration models to local conditions
are indispensable [32,33]. Some studies adapt HDM models by tak-
ing local construction materials and environmental conditions into
account [34–36].
It is well known that predicting the future performance of pave-
ments has multiple benefits in the allocation of a limited budget.
The proper allocation of the budget requires identifying the opti-
mum mathematical solution for the performance of the pavements
on the road network. There are thus many studies that aim to opti-
mize the parameters that describe pavement performance on road
sections. In these studies, researchers use methods that are based
on soft computing evolutionary computation, fuzzy decision tree,
logistic regression method, data-mining, multi-criteria decision
making applications, genetic algorithm, deterministic and stochas-
tic approaches and hybrid techniques [37–40]. As prioritization is a
decision making process, the statistical models are not very
responsive. Therefore, in the future, highway agencies will become
open to adopting new Decision-Aid Tools (DATs) that minimize the
costs relating to their field of activity [41].
3. Pavement performance
The availability of a reliable index, which is capable of reflecting
the actual conditions of distress on a pavement section, is essential
for pavement management, decision making, prioritization, plan-
ning etc. The PAVER system is an evaluation procedure for the
assessment of the current performance of the pavement by evalu-
ating data gathered according to the identification guide with the
code ASTM D 6433-11 in sample areas of 225 ± 90 m2
(2500 ± 1000 ft2), chosen randomly in line with distribution princi-
pals [6]. The evaluation is made by reading the resulting values
from the deduct value tables, which were prepared in accordance
to the deterioration type and with the help of the value derived
 U. Kırbasß, M. Karasßahin / Construction and Building Materials 116 (2016) 281–288
by proportioning the surface deterioration data, gathered at the
low (L), medium (M) and high (H) severity levels, to the size of
the sample unit. Extracting the obtained deduct values from 100
results in a PCI value, which defines the performance of the pave-
ment with a numerical value of 0–100 [7].
The pavement distress data of 21 different urban highways in
99 different sections in the Turkish city of Samsun were obtained,
as defined in the ASTM D 6433-11 standard. Road sections to be
assessed were selected from among those sections on which no
pavement rehabilitation work was conducted after construction.
The data covers a total 55,750 m length of road with a total surface
area of 337,867 m2. The data was evaluated to identify the current
performance of the pavement according to the PAVER system.
From the roads selected for the surface condition evaluation,
1415 sample units were first identified, of which 1375 were ana-
lyzed for the collection of surface data. Besides gathering surface
corruption data for 97.2 percent of the originally identified areas,
all the investigated areas were subjected to surface distress evalu-
ations, ensuring the precision of the project, according to the
requirements of the PAVER system. Fig. 1 shows pavements dis-
tress examples from the roads assessed.
Distress types gathered in three different categories which
names are ‘‘axle load”, ‘‘climate” and ‘‘others”. At the design phase
thicknesses of the layers of a pavement that composed road body
determined depending upon the repeat number of chosen standard
axle load. If more traffic uses the road than expected or heavier
axle loads uses the road more than calculated axle loads, then ‘‘axle
load” caused distresses would be observed. Besides, seasonal tem-
perature or daily temperature differences also causes distresses at
the pavement which are named ‘‘climatological distresses”. Table 1
shows the ‘‘distress type reasons” and ‘‘ASTM distress definition
guide”. Furthermore, pavements goes bad in time with such rea-
sons are; construction faults, maintenance faults, material specifi-
cations, leaking of oil, gas and other chemical fluids, salting at
defrosting works.
Fig. 1. Examples of pavement distress due to (a) patching and utility cut patching, (b) railroad crossing, (c) pothole, (d) corrugation, (e) alligator cracking, and (f) weathering
(surface wear).
283
Most studies conducted on in-service pavements have been
unable to identify any structural effects of pavement strength or
traffic loading, and have related the performance deterioration
(in other words roughness progression) directly to time and pave-
ment age [42]. In an urban road network, it is estimated that traffic
has a negligible effect since there is little heavy vehicle traffic vol-
ume. Although it is well known that weather conditions play an
important role in the deterioration of pavements in an urban road
network, weather forecasts spanning multiple years are notori-
ously unreliable. With this purpose, to ensure usability, only the
variable of the age of pavement was used as an independent vari-
able in the following models. To improve the predictive power of
the models, outlying data points were excluded from the analyses.
PCI values that fit the standard normal distribution requirement of
1  a = 0.90 (90%) were included. Because standard normal distri-
bution requires that 90% of the data points are covered at a 90%
level of confidence (z = 1.645), a total of 10 data points were
excluded from the analyses.
3.1. Traffic evaluation
It is known that traffic data is not collected regularly on urban
road in many countries. Therefore, the traffic volumes passing
through the assessed pavement sections is not known in the study.
However, traffic characteristics of the road sections were investi-
gated. For this purpose, volume of traffic counts was made on cer-
tain days of the week in 15-min periods. In this way, the peak hour
traffic volumes were determined in the morning, afternoon and
evening hours for each section. The peak hour traffic volumes be
enlarged by factor K, Annual Average Daily Traffic (AADT) volumes
were estimated. Using values of 0.09 (highway capacity manual
recommends [43]), 0.10 and 0.125 as the K factor, the results were
compared. Vehicle types were evaluated in 5 different categories.
General Directorate of Highways (Turkey) adopted by values were
used as equivalent single axle load (ESAL). The number of effective
284 U. Kırbasß, M. Karasßahin / Construction and Building Materials 116 (2016) 281–288

Table 1
PAVER distress classification for roads and parking areas [6].

Code Distress Unit of Defined Severity Cause Code Distress Unit of Defined Severity Cause
Measure Levels Measure Levels
1 Alligator cracking m2 Yes Load 11 Patching m2 Yes Other
2 Bleeding m2 Yes Other 12 Polished m2 No Other
aggregate
3 Block cracking m2 Yes Climate 13 Potholes number Yes Load
4 Bumps and sags m Yes Other 14 Railroad m2 Yes Other
crossings
5 Corrugation m2 Yes Other 15 Rutting m2 Yes Load
6 Depression m2 Yes Other 16 Shoving m2 Yes Other
7 Edge cracking m Yes Load 17 Slippage m2 Yes Other
cracking
8 Joint reflection cracking m Yes Climate 18 Swell m2 Yes Other
9 Lane/shoulder drop-off m Yes Other 19 Ravelling m2 Yes Climate
10 Longitudinal and transverse m Yes Climate 20 Weathering m2 Yes Climate
cracking

traffic lane used is determined as 2 on all road sections evaluated Table 3

in this study. ESAL values were calculated for a year on road section Statistical evaluation results.

examined. Traffic volumes in some sections, for example are ESAL K = 0.09 W 18 K = 0.10 W 18 K = 0.125 W 18
shown in Table 2. The statistical evaluation results for all pavement Min 500,095 450,085 360,068
sections are shown in Table 3. Max 789,328 710,395 568,316
Although there are important differences in sections of the traf- Mean 611,878 550,690 440,552
fic volume, due to the heavy vehicle traffic is relatively small, ESAL Std.Dev. 87,823 79,040 63,232
values of sections has no significant difference are shown in Tables
2 and 3. Therefore, traffic volume in the study was excluded. Even
if deterioration models were constructed using other independent
variable as traffic etc., models could not be used in a healthy way
for which regular data is not in urban roads.

3.2. Normality of the dependent variable

One of the preconditions in a regression analysis is that the

independent variable to be modeled has a normal distribution
[44]. To see whether the PCI values, which serve as the dependent
variable in this study, have a normal distribution, a one-sample
Kolmogorov Smirnov test was conducted. The significance value
(2-tailed) was found to be 0.176 > 0.05, indicating that the PCI
had a normal distribution with a 95% level of confidence. In addi-
tion, the histogram in Fig. 2 shows that the PCI values were not
severely skewed.

4. Modeling approaches

In the following sections, deterministic regression analysis,
MARS and ANN modeling approaches are described and their
results presented and evaluated.
4.1. Deterministic regression analysis
Empirical approaches using statistical regression techniques
have been utilized extensively for a long time in the prediction of
extremely complicated pavement responses and performance indi-
cators. A non-linear regression model was used in this study to
Fig. 2. Histogram of the PCI variable showing normal distribution.
describe PCI progression as a function of potential explanatory
variables available in the dataset, such as the age of the pavement.
Linear and nonlinear regression models (i.e., sigmoid or reversed S-
shaped curves) have previously been used by various researchers/
highway agencies [12]. In this study, a model for the prediction of
pavement distress was developed using non-linear regression and
data collected from a total of 89 road sections with hot-mix asphalt

Table 2
Traffic volumes and ESAL values examples of some sections.

Branch name Sec. no Peak hour traffic K = 0.09 ESAL K = 0.10 ESAL K = 0.125 ESAL
P. car Medium commercial vehicle Bus Truck Truck + Trailer
100th Year Ave. Sec. .1 470 39 14 18 1 506,060 455,454 364,363
Mustafa Kemal Ave. Sec. .2 1063 86 29 9 0 694,059 624,653 499,722
19 Mayıs Ave. Sec. .4 518 81 33 6 0 697,194 627,474 501,980
_
Ismet _
Inönü Ave. Sec. .3 300 54 22 8 1 528,358 475,522 380,418
 U. Kırbasß, M. Karasßahin / Construction and Building Materials 116 (2016) 281–288 285

pavement, which are part of the urban road network of the pro- 100
vince of Samsun. The current condition of the road sections was R2 = 0.7100
measured using the PAVER system. The dependent variable of 80
the prediction model was the PCI value, and the independent vari-
able was the age of the pavement (Age), calculated as the amount 60

PCI
of time that passed between the construction of the road segment,
and the date on which pavement performance was measured. 40
Many different mathematical models were considered for the
non-linear regression analysis, but polynomial regression was pre- 20
ferred because it is highly similar to the structure of the theoretical
model outlined in important studies, such as those by Haas [5] and 0
Shahin [6], and because it has a higher predictive accuracy com- 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
pared with other approaches. Because the best condition in the Age (years)
PAVER system receives a score of 100 and the current PCI values
Fig. 3. Pavement deterioration model graphical representation.
are calculated by deducting the total deduct value from 100, which
is a measure of the amount of distress on the pavement surface, the
constant of the model was 100. The model for the prediction of Table 5
pavement deterioration and statistical parameters are shown in ANOVA Table.
Table 4. Fig. 3 displays the graphical representation of the model. Sum of Squares df Mean Square F Sig.
Adjusted R2 of the model, which is a measure of predictive accu-
Regression 21170.359 3 7056.786 69.360 .000
racy, was found to be 0.70, an acceptable level [45]. It is also nota-
Residual 8648.023 85 101.741
ble that the level of significance of the model is very low. ANOVA Total 29818.382 88
test results of the model are reported in Table 5.

4.2. MARS approach
The aim of the MARS procedure is to combine recursive parti-
tioning and spline fitting in a way that best retains the positive
aspects of both, while being less vulnerable to their unfavorable
properties. This has been accomplished, at least to some extent
[46]. MARS has been successfully applied to many areas of science
and technology in recent years. MARS algorithm uses expansions in
piecewise linear, actually, truncated basis functions of the form;
½þðx  sÞþ ; ½ðx  sÞþ
where [q]+: = max {0, q} and s is a univariate knot. Each function is
piecewise linear, with a knot at the value s, and the corresponding
couple of function is called a reflected pair [47]. The MARS algo-
rithm searches over all possible univariate hinge locations and
across interactions among all variables.
The approach is analogous to the use of splines. The general
form of a MARS predictor is as follows:
X
P X
B
Y ¼ b0 þ ½bjb ðþÞMaxð0; xj  Hb;j Þ þ bjb ðÞMaxð0; Hbj  xj Þ
j¼1 b¼1
for P predictor variables and B basis function. The basis functions
max(0, x  H) and max(0, H  x) are univariate and do not have
to each be present if their b coefficients are 0. The H values are
called ‘hinge’ or ‘knot’ [15].
The MARS algorithm, in the forward stepwise part of the model,
searches for model constants for basis functions. In the backward
stepwise part, the basis functions that reduce the predictive
Table 4
Regression model for pavement deterioration and statistical parameters.
accuracy of the model are eliminated. The lack of fit measure used
is based on the generalized cross-validation criterion (GCV). The
GCV criterion is the average residual error multiplied by a penalty
to adjust for the variability associated with the estimation of more
parameters in the model [27].
In the prediction model, the PCI values of the road sections were
used as the dependent variable, and pavement age as the predictor
variable. A threshold value of 0.0005 and penalty value of 2 were
used, and the GCV error value was found to be 108.993. At the
ð1Þ end of the MARS analysis, the model for the prediction of pavement
deterioration consisted of four terms created using three basis
functions. The PCI prediction model that was developed using
MARS and 89 data points was as follows:
PCI ¼ 45:11898  15:56164  BF1 þ 7:19860  BF2 þ 7:71304  BF3
ð3Þ
where;
BF1 = max (0; Age - 7.32)
BF2 = max (0; 7.32- Age)
BF3 = max (0; Age - 5.33)
ð2Þ
The knots values of the three basis functions that resulted from
the algorithm were 5.33 and 7.32 for the Age variable. This method,
which divides data clusters into regions using knots, can eliminate
independent variables when necessary and within certain limits,
unlike non-linear regression analysis. This prevents the creation
of meaningless output (prediction) values when independent vari-
ables with extreme values are included in the model as inputs. In
other words, the MARS algorithm involves a mechanism that elim-
inates overfit data and other inconsistencies in the data [27].
Table 6 reports the model developed.

3 2
Table 6
PCI ¼ 100  0:0659  Age þ 1:2521  Age  11:6722  Age ð1Þ PCI deterioration model coefficients, knots and basis functions.

Coefficients Knots
Coeff of correlation : 0.8426
Approximate R2 : 0.7100 Intercept 45.11898
Standard deviation of error : 9.9134 Term.1 15.56164 7.32
Absolute mean of error : 7.6799 Term.2 7.19860 7.32
Arithmetic mean of error : 0.0018 Term.3 7.71304 5.33
286 U. Kırbasß, M. Karasßahin / Construction and Building Materials 116 (2016) 281–288
The regression (R2) value of the model was found to be 0.7238,
and adjusted R2 was 0.7107. MARS is known for its ability to accu-
rately predict a single dependent variable with multiple indepen-
dent variables, but its predictive accuracy with a single
dependent variable also seems to be high. It is clear that this math-
ematical modeling technique, which identifies knots at the most
appropriate points and optimizes data distribution in the form of
splines, can be used to predict future values of the PCI. Fig. 4 shows
the theoretical representation of the spline form of the model.
4.3. ANN method
ANNs, or simple neural networks, are computing systems made
up of a number of simple and highly interconnected elements that
process information by its dynamic-state response to external
inputs [25]. The recognition of this type of model as a neural net-
work is because of the similarity of its configurations to the archi-
tecture of the human brain [16]. ANNs are mathematical systems
that consist of a series of process units (neurons) that are con-
nected to each other with associated weights. A process unit is
an equation that is often referred to as a transfer function. This pro-
cess unit received signals from other neurons, combines and trans-
forms them, and gives a numerical output. Process units
correspond to actual neurons and are interconnected in a network.
These structures make up the neural networks. Significant progress
has been reported in pavement deterioration modeling of simu-
lated data using ANN.
Many different ANN network structures and learning methods
are used in the literature. Multi-layer feed-forward (MLF) networks
are the most widely used network structure in the prediction of
pavement deterioration. MLF neural networks, trained with a
back-propagation learning algorithm, are the most popular neural
networks. The ANN modeling consists of two steps: the first step is
to train the network; the second step is to test the network with
data that were not used in the training step. During the training
stage the network uses the inductive-learning principle to learn
from a set of examples called the training set [24]. Most ANNs con-
tain some form of ’learning rule’ that modifies the weights of the
connections according to the input patterns with which it is
presented.
A typical feed-forward ANN contains three layers, including one
input layer, several hidden layers, and one output layer. It should
be noted that there is no theoretical limit on the number of hidden
layers but typically there is just one or two [48]. Fig. 5 depicts using
artificial neural network architecture.
Age of pavement was used as the input to the ANN network cre-
ated, and the output was the PCI. Randomly selected from the 89
data points, 63 (70%) were used for training, 13 (15%) were used
100
R2 = 0.7238
80
60
PCI
40
20
0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Age (years)
Fig. 4. A graphical representation of a regional split applied to the PCI data scatter
plot.
Age PCI
Fig. 5. Artificial neural network architecture.
for validation, and the remaining 13 (15%) were used for testing.
To obtain the highest predictive accuracy possible, multiple hidden
layers and multiple neuron alternatives at each layer were consid-
ered and the most appropriate neural network architecture was
created. A network architecture with one input layer (single neu-
ron), one hidden layer (three neurons) and one output layer (single
neuron) was identified as the optimal topology for the prediction
of deterioration in HMA paved urban roads. The network, opti-
mized for predictive accuracy, can be described as having a
(1:3:1) architecture. In the analysis stage, the Levenberg-
Marquardt back-propagation training algorithm was used as the
learning algorithm. All variables were normalized between 0 and
1 to achieve standardization before training the ANN model. The
parameter value for the training function was 0.001, and a decrease
factor of 0.1 and increase factor of 10 were used. The R2 value,
which captures the predictive accuracy of the model, was found
to be 0.7118 for the training set, 0.8523 for the validation set,
0.8192 for the testing set, and 0.7456 for all input data. Fig. 6 pro-
vides a graphical representation of the pavement deterioration
model for urban HMA paved roads, created using synthetic Age
data.
5. Comparison of models
The Age variable, which was used as the independent variable
in the models examined, arguably contains many parameters, such
as traffic and climate, on an urban road. This is because in most
urban regions, road drainage systems are constructed, housing
development is completed, heavy vehicle traffic is limited, and
traffic increase is minimal and constant. This study found that
the three modeling approaches examined have similar levels of
prediction accuracy. However, the model with the highest predic-
tive accuracy was the ANN (R2 = 0.7456), followed by MARS
(R2 = 0.7238) and deterministic regression (R2 = 0.7100).
Because the data used in the models came from road sections
on which no rehabilitation work was conducted after construction,
100
R2 = 0.7456
80
60
PCI
40
20
0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Age (years)
Fig. 6. Graphical representation of model results with synthetic data.
 U. Kırbasß, M. Karasßahin / Construction and Building Materials 116 (2016) 281–288 287

100 predict the future performance of road sections upon which PCI
ANN
measurements are made, a mathematical simulation needs to be
80 MARS performed. That is to say, in order to predict future performance,
Regression the difference between the PCI value generated by the model when
60 the current age of the road section is input and the observed PCI
PCI

value needs to be added to or subtracted from the PCI value pre-

40
20
0 and Kruskal-Wallis H tests performed statistical significance val-
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Age (years)
Fig. 7. Deterioration curves predicted by the three models.
the lifespan of HMA paved roads in an urban road network that
receives routine maintenance can be said to be about 15–16 years.
For a visual comparison of the results obtained using the three
models, deterioration curves predicted by all three models are
shown on the same graph. Fig. 6 shows the regression curve, the
points at which the data clusters were split using the knot values
of the MARS model, connected through lines (splines), and the
results of the ANN model obtained by creating synthetic values.
Fig. 7 shows that deterioration trends predicted by the three
modeling approaches are similar to one another. In addition, the
performance deterioration (distress) trends predicted by the three
models are very similar to the schematic curves, frequently
referred to in the literature, that represent theoretical expectations
about performance deterioration [1,5,6]. On the other hand, this
study found that the ANN method is superior to the other models
in terms of its ability to model data clusters. Fig. 8 shows the pre-
diction accuracies of the three models, defined as the difference
between the predicted and the observed PCI values.
In the PAVER system, a newly constructed pavement with no
construction defects is assigned a PCI score of 100. Thus, in the
regression analysis, the constant of the model was 100. It should
be noted that to predict using the regression method, the future
performance of road sections upon which PCI measurements are
conducted, the current PCI value should be inserted as the constant
instead of 100. On the other hand, MARS and ANN methods do not
have a model constant. Thus, if MARS or ANN is to be used to
dicted by the model.
Finally, in this study, statistical similarity between estimated
PCI value, using three different modeling technique, were investi-
gated. For this purpose, the data were compared between ANOVA
ues. Statistical significance values using both methods were found
to be less than 0.05. According to the results, included with every
three modeling techniques is seen that statistically significant rela-
tionship between PCI values. Table 7 reports the statistical test
results.
6. Conclusions
Using three different modeling approaches, Deterministic
Regression, MARS and ANN, in this study developed models for
predicting pavement deterioration in HMA paved road sections in
an urban road network, and compared the prediction accuracies
of the models. The PCI index required by the PAVER system -
which is the most commonly used system for pavement manage-
ment in urban roads-, was used as the dependent variable in the
models. The independent variable was the age of the pavement
(Age), which was assumed, in urban environments, to represent
traffic and climate parameters as well. Data used in the models
were collected in the Middle East, from road sections constructed
using local construction materials, and on which no rehabilitation
work was conducted after construction. The ANN method was
found to be the most appropriate model for predicting deteriora-
tion in urban roads. Nevertheless, the three models had prediction
accuracies that were close to each other.
The models developed in this study can be used in the manage-
ment of pavements in urban areas with similar climate and traffic
conditions. The study also provides a comparative assessment of
different mathematical modeling techniques in the prediction of
pavement deterioration, which can be useful in the design of future
studies on the topic.
Acknowledgements

100 This work was supported by the Scientific Research Projects

_
Program of Istanbul University (Project No 21794).
Regression
80
MARS References
PCI (Pred.)

60
ANN [1] AASHTO, Guide for Design of Pavement Structures, AASHTO, Washington,
40 1993.
(Reg.) Sq.R = 0.7100 [2] D. Moazami, H. Behbahani, R. Muniandy, Pavement rehabilitation and
maintenance prioritization of urban roads using fuzzy logic, Expert Syst.
20 (MARS) Sq.R = 0.7238 Appl. 38 (2011) 12869–12879, http://dx.doi.org/10.1016/j.eswa.2011.04.079.
[3] J. Yang, J.J Lu, M. Gunaratne, Q. Xiang, Overall Pavement Condition Forecasting
0 (ANN) Sq.R = 0.7456 Using Neural Networks – An Application to Florida Highway Network, in: 82th
0 20 40 60 80 100 Annual Meeting of Transportation Research Board, Washington D.C. (Annual
Meeting CD-ROM), 2003.
PCI (Obs.) [4] Z. Luo, Pavement performance modelling with an autoregression approach, Int.
J. Pavement Eng. 14 (2013) 85–94, http://dx.doi.org/10.1080/
Fig. 8. Prediction accuracies of the three models. 10298436.2011.617442.
[5] R. Haas, W.R. Hudson, J. Zaniewski, Modern Pavement Management, Krieger
Publishing Company, Florida, 1994.
[6] M.Y. Shahin, Pavement Management for Airports, Roads and Parking Lots,
Table 7 second ed., Springer, New York, 2005.
Statistical test results. [7] M.Y. Shahin, J.M. Becker, Development of performance prediction models for
airfield pavements, Transp. Res. Rec. 985 (1984) 25–33.
Significance Regression MARS ANN
[8] R. Hassan, O. Lin, A. Thananjeyan, A comparison between three approaches for
ANOVA test 0.000 0.000 0.000 modelling deterioration of five pavement surfaces, Int. J. Pavement Eng. (2015).
Kruskal-Wallis H test 0.021 0.015 0.021 Published online: 13 Apr 2015. http://dx.doi.org/10.1080/10298436.2015.
1030744.
288 U. Kırbasß, M. Karasßahin / Construction and Building Materials 116 (2016) 281–288
[9] H. Ziari, J. Sobhani, J. Ayoubinejad, T. Hartmann, Prediction of IRI in short and
long terms for flexible pavements: ANN and GMDH methods, Int. J. Pavement
Eng. (2015). Published online: 10 Mar 2015. http://dx.doi.org/10.1080/
10298436.2015.1019498.
[10] O. Salem, A. El-Assaly, S. AbouRizk, Performance prediction models of
pavement highway network in Alberta, in: 82th Annual Meeting of
Transportation Research Board, Washington D.C. (Annual Meeting CD-ROM),
2003.
[11] M.J. Khattak, C. Landry, J. Veazey, Z. Zhang, Rigid and composite pavement
index-based performance models for network pavement management system
in the state of Louisiana, Int. J. Pavement Eng. 14 (7) (2013) 612–628, http://dx.
doi.org/10.1080/10298436.2012.715643.
[12] E. Chou, H. Pulugurta, D. Datta, Pavement Forecasting Models, Federal
Highway Administration, 2008. Research Report, FHWA/OH-2008/3.
[13] C. Chen, J. Zhang, Comparisons of IRI-Based Pavement Deterioration Prediction
Models Using New Mexico Pavement Data, Geo-Front. (ASCE) (2011) 4594–
4603. http://dx.doi.org/10.1061/41165(397)470.
[14] K. Ahmed, G. Abu-Lebdeh, R.W. Lyles, Prediction of pavement distress index
with limited data on causal factors: an auto-regression approach, Int. J.
Pavement Eng. 7 (1) (2006) 23–35, http://dx.doi.org/10.1080/
10298430500502017.
[15] N.O. Attoh-Okine, K. Cooger, S. Mensah, Multivariate adaptive regression
(MARS) and hinged hyperplanes (HHP) for doweled pavement performance
modeling, Constr. Build. Mater. 23 (2009) 3020–3023, http://dx.doi.org/
10.1016/j.conbuildmat.2009.04.010.
[16] M. Parvini, Artificial neural network modeling of pavement performance using
expert judgement, Road Mater. Pavement Design 3 (4) (2002) 373–384, http://
dx.doi.org/10.1080/14680629.2002.9689931.
[17] M.U. Khan, M. Mesbah, L. Ferreira, D.J. Williams, Developing a new road
deterioration model incorporating flooding, Proc. Inst. Civ. Eng. Transport 167
(TR5) (2014) 322–333, http://dx.doi.org/10.1680/tran.13.00095.
[18] H. Pulugurta, Q. Shao, Y.J. Chou, Pavement condition prediction using Markov
process, J. Statist. Manag. Syst. 12 (5) (2009) 853–871, http://dx.doi.org/
10.1080/09720510.2009.10701426.
[19] N. Lethanh, B.T. Adey, Use of exponential hidden Markov models for modelling
pavement deterioration, Int. J. Pavement Eng. 14 (7) (2013) 645–654, http://
dx.doi.org/10.1080/10298436.2012.715647.
[20] E.S. Park, R.E. Smith, T.J. Freeman, C.H. Spiegelman, A Bayesian approach for
improved pavement performance prediction, J. Appl. Statist. 35 (11) (2008)
1219–1238, http://dx.doi.org/10.1080/02664760802318651.
[21] H.P. Hong, S.S. Wang, Stochastic modeling of pavement performance, Int. J.
Pavement Eng. 4 (4) (2003) 235–243, http://dx.doi.org/10.1080/
10298430410001672246.
[22] N.F. Pan, C.H. Ko, M.D. Yang, K.C. Hsu, Pavement performance prediction
through fuzzy regression, Expert Syst. Appl. 38 (2011) 10010–10017, http://dx.
doi.org/10.1016/j.eswa.2011.02.007.
[23] M. Karasßahin, S. Terzi, Performance model for asphalt concrete pavement
based on the fuzzy logic approach, Transport 29 (1) (2014) 18–27, http://dx.
doi.org/10.3846/16484142.2014.893926.
[24] S. Terzi, Modeling the pavement serviceability ratio of flexible highway
pavements by artificial neural networks, Constr. Build. Mater. 21 (2007) 590–
593, http://dx.doi.org/10.1016/j.conbuildmat.2005.11.001.
[25] N.O. Attoh-Okine, Analysis of learning rate and momentum term in
backpropagation neural network algorithm trained to predict pavement
performance, Adv. Eng. Softw. 30 (4) (1999) 291–302, http://dx.doi.org/
10.1016/S0965-9978(98)00071-4.
[26] S. Terzi, Modeling for pavement roughness using the ANFIS approach, Adv.
Eng. Softw. 57 (2013) 59–64, http://dx.doi.org/10.1016/j.
advengsoft.2012.11.013.
[27] N.O. Attoh-Okine, S. Mensah, M. Nawaiseh, A new technique for using
multivariate adaptive regression splines (MARS) in pavement roughness
prediction, Proc. Inst. Civil Eng. Transport 156 (1) (2003) 51–55, http://dx.
doi.org/10.1680/tran.2003.156.1.51.
[28] A. Ahmed, S. Labi, Z. Li, T. Shields, Aggregate and disaggregate statistical
evaluation of the performance-based effectiveness of long-term pavement
performance specific pavement study-5 (LTPP SPS-5) flexible pavement
rehabilitation treatments, Struct. Infrastruct. Eng.: Main. Manag. Life-Cycle
Design Perform. 9 (2) (2013) 172–187, http://dx.doi.org/10.1080/
15732479.2010.524224.
[29] M. Mubaraki, Predicting pavement condition deterioration for the Saudi inter-
urban road network, in: Material Design, Construction, Maintenance, and
Testing of Pavements, GeoHunan International Conference, ASCE, 2009, pp.
56–61, http://dx.doi.org/10.1061/41045(352)9.
[30] L.E. Amador-Jiménez, D. Mrawira, Reliability-based initial pavement
performance deterioration modelling, Int. J. Pavement Eng. 12 (2) (2011)
177–186, http://dx.doi.org/10.1080/10298436.2010.535538.
[31] E. Fekpe, N.O. Attoh-Okine, Deterioration modelling for lateritic-base flexible
pavements, Constr. Build. Mater. 9 (3) (1995) 159–163, http://dx.doi.org/
10.1016/0950-0618(95)00008-4.
[32] C.R. Bennett, W.D. Paterson, A Guide to Calibration and Adaptation, HDM-4,
vol. 5, PIARC, 1999.
[33] A. Braga, D. Cygas, Adaptation of pavement deterioration models to Lithuanian
automobile roads, J. Civ. Eng. Manag. 3 (3) (2002) 214–220, http://dx.doi.org/
10.1080/13923730.2002.10531279.
[34] J.B. Odoki, M. Anyala, E. Bunting, HDM-4 adaptation for strategic analysis of UK
local roads, Proc. Inst. Civ. Eng. Transport 166 (TR2) (2013) 65–78, http://dx.
doi.org/10.1680/tran.9.00026.
[35] T.C. Martin, Heavy vehicle road wear on sealed unbound granular roads, Proc.
Inst. Civ. Eng. Transp. 164 (TR1) (2011) 13–22, http://dx.doi.org/
10.1680/tran.9.00012.
[36] D. Jorge, A. Ferreira, Road network pavement maintenance optimisation using
the HDM-4 pavement performance prediction models, Int. J. Pavement Eng. 13
(1) (2012) 39–51, http://dx.doi.org/10.1080/10298436.2011.563851.
[37] P. Chootinan, A. Chen, M.R. Horrocks, D. Bolling, A multi-year pavement
maintenance program using a stochastic simulation-based genetic algorithm
approach, Transp. Res. Part A 40 (2006) 725–743, http://dx.doi.org/10.1016/
j.tra.2005.12.003.
[38] S. Chi, J. Hwang, M. Arellano, Z. Zhang, M. Murphy, Development of network-
level project screening methods supporting the 4-year pavement management
plan in Texas, J. Manag. Eng. ASCE 29 (4) (2013) 482–494, http://dx.doi.org/
10.1061/(ASCE)ME.1943-5479.0000158.
[39] S. Chi, M. Murphy, Z. Zhang, Sustainable road management in texas: network-
level flexible pavement structural condition analysis using data-mining
techniques, J. Comput. Civ. Eng. 28 (1) (2014) 156–165, http://dx.doi.org/
10.1061/(ASCE)CP.1943-5487.0000252.
[40] J.S. Chou, T.S. Le, Reliability-based performance simulation for optimized
pavement maintenance, Reliab. Eng. Syst. Saf. 96 (2011) 1402–1410, http://dx.
doi.org/10.1016/j.ress.2011.05.005.
[41] S. Meneses, A. Ferreira, Flexible pavement maintenance programming
considering the minimisation of maintenance and rehabilitation costs and
the maximisation of the residual value of pavements, Int. J. Pavement Eng. 16
(7) (2015) 571–586, http://dx.doi.org/10.1080/10298436.2014.943207.
[42] W.D.O. Paterson, Road Deterioration and Maintenance Effects, Published for
The World Bank, Maryland, 1990.
[43] HCM2000, Highway Capacity Manual, Transportation Research Board,
Washington, 2000.
[44] J. Cohen, P. Cohen, S.G. West, L.S. Aiken, Applied Multiple Regression/
Correlation Analysis for the Behavioral Sciences, Lawrence Erlbaum
Associates Inc., New Jersey, London, 2003.
[45] D.N. Gujarati, D.C. Porter, Basic Econometrics, fifth ed., McGraw-Hill, New
York, 2009.
[46] J.H. Friedman, Multivariate adaptive regression splines, Ann. Statist. 19 (1)
(1991) 1–67. http://www.jstor.org/stable/2241837.
[47] G.W. Weber, I. _ Batmaz, G. Köksal, P. Taylan, F. Özkurt, CMARS: a new
contribution to nonparametric regression with multivariate adaptive
regression splines supported by continuous optimization, Inver. Probl. Sci.
Eng. 20 (3) (2012) 371–400, http://dx.doi.org/10.1080/
17415977.2011.624770.
[48] S. Sumathi, S. Paneerselvam, Computational Intelligence Paradigms: Theory &
Applications using MATLAB, Taylor and Francis Group, LLC, 2010.