Research Article

Transportation Research Record

1–13
Ó National Academy of Sciences:
Statistics and Artificial Intelligence-Based Transportation Research Board 2020
Article reuse guidelines:
Pavement Performance and Remaining sagepub.com/journals-permissions
DOI: 10.1177/0361198120915889

Service Life Prediction Models for journals.sagepub.com/home/trr

Flexible and Composite Pavement

Systems

Orhan Kaya1, Halil Ceylan2,3, Sunghwan Kim1, Danny Waid4,

and Brian P. Moore5

Abstract
In their pavement management decision-making processes, U.S. state highway agencies are required to develop performance-
based approaches by the Moving Ahead for Progress in the 21st Century (MAP-21) federal transportation legislation. One of
the performance-based approaches to facilitate pavement management decision-making processes is the use of remaining ser-
vice life (RSL) models. In this study, a detailed step-by-step methodology for the development of pavement performance and
RSL prediction models for flexible and composite (asphalt concrete [AC] over jointed plain concrete pavement [JPCP]) pave-
ment systems in Iowa is described. To develop such RSL models, pavement performance models based on statistics and artifi-
cial intelligence (AI) techniques were initially developed. While statistically defined pavement performance models were
found to be accurate in predicting pavement performance at project level, AI-based pavement performance models were
found to be successful in predicting pavement performance in network level analysis. Network level pavement performance
models using both statistics and AI-based approaches were also developed to evaluate the relative success of these two mod-
els for network level pavement performance modeling. As part of this study, in the development of pavement RSL prediction
models, automation tools for future pavement performance predictions were developed and used along with the threshold
limits for various pavement performance indicators specified by the Federal Highway Administration. These RSL models will
help engineers in decision-making processes at both network and project levels and for different types of pavement manage-
ment business decisions.

U.S. state highway agencies are required to develop
performance-based approaches in their pavement man-
agement decision-making processes based on the Moving
Ahead for Progress in the 21st Century (MAP-21) federal
transportation legislation (1). One such performance-
based approach to facilitate the pavement management
decision-making process is to use a remaining service life
(RSL) model. RSL for pavements can be defined as the
time span between the present time and the time when a
significant rehabilitation treatment or reconstruction
should occur (2). Although application of a structural
overlay or reconstruction would normally be regarded as
a sign for termination of pavement service life, minor
maintenance treatments or thin overlays are often not
considered as such signs (2). Models for predicting the
RSL of pavements have been developed and are being
used as part of the pavement management process (3).
Multiple advantages of RSL have been reported in the
literature (4), with key positive features that include the
following:
1
Department of Civil, Construction and Environmental Engineering
(CCEE), Iowa State University, Ames, IA
2
FAA PEGASAS (Partnership to Enhance General Aviation Safety,
Accessibility and Sustainability) Center of Excellence (COE) on General
Aviation, CCEE, Iowa State University, Ames, IA
3
Program for Sustainable Pavement Engineering and Research (PROSPER),
CCEE, Iowa State University, Ames, IA
4
Iowa County Engineers Association (ICEA) Service Bureau, West Des
Moines, IA
5
Research and Analytics, ICEA Service Bureau, Ames, IA
Corresponding Author:
Orhan Kaya, okaya@iastate.edu
2 Transportation Research Record 00(0)
 Provides the time, expressed in years, before reha-
bilitation is required for any given road section
 Easy to understand (especially for public)
 Can be a multi-conditional measure developed
from any type of functional, structural data, or
both
 Allows agencies to distinguish between two road
sections with the same current condition (i.e., the
same current international roughness index)
 Provides deeper insight by converting ‘‘condition
measures’’ into an ‘‘operational performance’’
measure that tells how well or long the road will
continue serving the public
 Can be an ideal tool to address the transportation
planning and performance management criteria
requirements of the MAP-21 legislation
Performance curves or pavement performance models
are used to evaluate how a pavement’s performance
changes over the time. They could be developed using
various pavement performance indicators (international
roughness index [IRI], distresses, and so on). Pavement
performance models can be categorized into two groups,
deterministic and probabilistic, based on their prediction
results (5–8). Deterministic models estimate a single con-
dition value for a given time during a pavement’s design
life, while probabilistic models estimate the probability
of a condition value for a given time (5). Most state high-
way agencies use deterministic models as part of their
pavement management systems for various reasons,
including: (a) ease of explaining such models to users
and (b) ease of incorporating such models into pavement
management systems (9).
RSL models are broadly categorized as either
mechanistic or empirical models (10). In mechanistic
models, mechanistic-based pavement performance mod-
els are used (based on engineering principles), while in
empirical models data from observed historical data and
other parameters are analyzed mostly through statistical
approaches. Pavement performance models based on
artificial neural networks have been developed, mostly to
model relationships between pavement performance data
and several input parameters related to pavement struc-
tural design, traffic, and so forth (11, 12). However, there
are not many studies found in the literature where net-
work level pavement performance models based on arti-
ficial neural networks are developed to predict
performance change of pavement sections over time in a
quick and efficient way (13). Moreover, no study was
found in the literature in which efficient network level
and project level RSL models are developed separately.
Threshold limits are determined performance indica-
tor values at which a significant rehabilitation treatment
or reconstruction is needed (3). Performance indicators
and threshold limits are agency-specific parameters used
for rehabilitation decision-making processes. Both per-
formance models and threshold limits are components
used in the development of RSL models.
Objectives
In this study, a detailed step-by-step methodology in the
development of a framework for project and network
level pavement performance and RSL prediction models
for flexible and composite (asphalt concrete [AC] over
jointed plain concrete pavement [JPCP]) overlays is
explained using real pavement performance data
obtained from the Iowa Department of Transportation
(DOT) pavement management information system
(PMIS) database. Two approaches investigated for
developing pavement performance models are: a statis-
tics approach for use in project level pavement manage-
ment and an artificial intelligence-based approach for
use in network level pavement management. Using the
developed statistics and AI-based models, future pave-
ment performance predictions are successfully calculated
for each pavement section used in this study. Network
level pavement performance models are also developed
using a statistically defined approach with the same input
parameters used in those developed using AI-based
approaches to evaluate the relative success of these
approaches in network level pavement performance
modeling.
Automation tools based on Microsoft Excel have
been developed for both project and network level pave-
ment performance modeling and analysis to facilitate
development of pavement performance and RSL models,
to make future pavement performance predictions, and
to estimate RSL for any given pavement section. These
tools, that make use of real pavement performance data
to produce realistic future condition predictions, can eas-
ily be incorporated into pavement management processes
and help engineers make better-informed performance-
based pavement infrastructure planning decisions and
optimize agency resource expenditures.
Description of Overall Approaches and
Data Preparation
Figure 1 depicts the stages of development of the pave-
ment performance and RSL models followed in this
study. Initially, statistics and AI-based approaches were
investigated for developing pavement performance mod-
els in use of project level and network level pavement
management, respectively. Both project and network
level pavement performance models were developed for
flexible and composite (AC over JPCP) pavement sys-
tems in Iowa. Project level pavement performance
Kaya et al 3

Figure 1. Schematic image of development stages of pavement performance and remaining service life models. FHWA = Federal
Highway Administration; RSL = remaining service life.

models were developed for each pavement section in  ypredicted= Model predictions for IRI or other
each pavement type, while network level pavement per- pavement performance indicators
formance models were developed for each pavement per-
formance indicator, or a condition matrix (i.e., distresses Once pavement performance models were developed
and IRI) for each pavement type. for the two pavement types, remaining service lives for the
Success of the pavement performance prediction mod- pavement sections were calculated using threshold limits
els in mimicking measured pavement performance indica- for various performance indicators. Based on the Federal
tors was quantified using a line-of-equality coefficient of Highway Administration (FHWA)’s Final Rule (effective
correlation (R2) (Equation 1), an absolute average error February 17, 2017) regarding implementation of the per-
(AAE) (Equation 2) and a standard error of the estimates formance management requirements of MAP-21 and the
(SEE) (Equation 3). Higher R2 and lower AAE and SEE Fixing America’s Surface Transportation Act (1, 14), con-
values are signs of accurate model prediction. dition of pavements is required to be determined based on
 2 the following metrics: IRI, percent cracking, rutting, and
Pn faulting (Table 1). IRI was used as a construction trigger
j=1 ymeasured
j  ypredicted
j
R2 = 1  P  2 ð1Þ for the rehabilitation decision-making process in project
n and network level RSL calculations. RSL is determined
j=1 ymeasured
j  ymeasured
mean
  based on the year when future performance predictions
Pn  measured  reach the ‘‘poor’’ condition threshold for the correspond-
j = 1  yj  ypredicted
j 
AAE = ð2Þ ing condition metric (See Table 1).
n The Iowa DOT collects pavement condition data and
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
uP  2 stores them in its PMIS, and information on pavement
u n  ypredicted
t j = 1 ymeasured j j structural design features and traffic volume are also
SEE = ð3Þ available as part of the PMIS. Iowa DOT’s PMIS data-
n
base has been used as data source in this study. This
where database includes all information on traffic, distress, and
construction information related to the pavement
 n = Data set size sections.
 j = Case number in the data set The number of pavement sections and the total num-
 ymeasured= Measured IRI or other pavement per- ber of data points for each pavement type used in this
formance indicator measurements study are as follows:
4 Transportation Research Record 00(0)
Table 1. Pavement Condition Rating Thresholds Determined by
FHWA for Flexible and Composite Pavement Systems (14)
Condition metric Performance level
IRI (in./mile) Good
Fair
Poor
Percent cracking Good
Fair
Poor
Rutting (in.) Good
Fair
Poor
Note: FHWA = Federal Highway Administration; IRI = international
roughness index.
 35 sections for flexible pavements (430 data
points)
 60 sections for composite (AC over JPCP) pave-
ments (644 data points).
The pavement sections used in this study represent a
variety of geographical locations across Iowa with vari-
ous traffic levels, thicknesses, and ages. Distributions of
locations, traffic levels, thicknesses, and ages for these
pavement sections and other detailed information can be
found in another study by Kaya (15).
While analyzing pavement condition data points for
each pavement section in the PMIS database, it was
realized that in some pavement sections, measured
pavement condition values for some pavement perfor-
mance indicators remained the same over several years,
after which an increase in those pavement condition
values was observed. Note that Iowa DOT has a pave-
ment performance collection practice whereby pave-
ment performance measurements for some sections are
collected in even years while for some others they are
collected in odd years. Because pavement condition
data was not collected or recorded every year for some
sections, pavement condition measurements reported
for previous years had been recorded as pavement con-
dition measurements for upcoming years for these sec-
tions in PMIS. In such cases, a systematic data
preparation methodology similar to one described in
the literature for previous studies was developed (5, 16):
A linear increase was achieved between the first year
when pavement condition data points started to be the
same over several years and the year when an increase
in those pavement condition values was observed.
Applying this data preparation methodology, more rea-
listic pavement condition records can be obtained, and
in turn, more accurate pavement performance models
can be developed.
Development and Accuracy Evaluations of
Statistics-Based Pavement Performance
Threshold Model
\95 A statistically defined sigmoid pavement deterioration
95–170 curve-based approach was used for development of the
.170 project level pavement performance model in this study.
\5% Sigmoidal equations have been particularly used in statis-
5–20% tics model development because: (a) they have a low ini-
.20%
\0.20 tial slope and an increasing slope with time and (b) they
0.20–0.40 follow a trend in which pavement condition always gets
.0.40 worse and damage is irreversible. Both these features of
sigmoidal models cause these models to mimic pavement
deterioration behavior observed in project level studies
(5, 17, 18). Since sigmoidal equations have been found to
model pavement deterioration successfully when there is
single pavement deterioration trend (project level), a sig-
moidal equation for each pavement section in each pave-
ment type was optimized, with each equation having
different coefficients. IRI was used as a performance
indicator in project level pavement performance models.
Equation 4 is the generalized sigmoidal equation used
for IRI calculation.
C2
IRI = C1 + ð4Þ
1 + eðC3 + C4 3 ageÞ
where C1, C2, C3, and C4 are coefficients that represent
contributions of different input parameters.
Sigmoidal curves were fitted to measured IRI values
by minimizing the square of differences value between
measured and predicted IRI values. The fitting process
was carried out by manipulating prediction coefficients
(Equation 4) to produce minimum error.
Figure 2 shows examples of IRI prediction models for
flexible and composite (AC over JPCP) pavement types.
Using these models, future IRI predictions can be calcu-
lated for these pavement types.
As part of this study, a Microsoft Excel Macro-based
automation tool was developed, automatically updating
and improving pavement performance prediction models
as more data were added into the model development data-
set. The benefit of this tool is that, as engineers add more
data into the model development dataset, they will be able
automatically to refine performance prediction models and
make decisions using the most recent and more accurate
pavement performance models. Another benefit of using
this tool is that pavement performance prediction models
can be developed using very few data points.
Development and Results of Statistics-
Based Pavement RSL Model
Once a pavement performance model has been devel-
oped for each pavement section, as explained in the
Kaya et al 5

Figure 2. Results of international roughness index (IRI) prediction model and equation examples for (a) a new flexible pavement section
(US 61, mile post [MP] 167.95 to 174.74, North; average annual daily truck traffic [AADTT]: 1,154; construction year: 1999) and (b) a
composite (AC over JPCP) pavement section (US 30, MP 310.08 to 318.84, West; traffic [AADTT]: 1,264; restoration year: 2000). AAE =
absolute average error; AC = asphalt concrete; JPCP = jointed plain concrete pavement; SEE = standard error of the estimates.

previous section, the RSL of each pavement section can
be calculated using threshold limits for the pavement
performance indicators. In this study, IRI was used as a
performance indicator for project level RSL calculations
because: (a) it quantifies functional performance of pave-
ment systems, the aspect most road users care about, (b)
it has also been adopted as a standard for the Federal
Highway Performance Monitoring System (19), and (c)
it is also one of the condition metrics identified for use
by FHWA (14). The same threshold level recommended
by FHWA for poor pavement condition (i.e., an IRI
value of 170 in./mile) was selected as the threshold value
in this study for project level RSL calculations (14).
The RSL for each pavement section was calculated by
the following steps:
1. Statistically defined pavement performance mod-
els were developed for each pavement section in
each pavement type (i.e., flexible and composite).
2. Using the developed pavement performance mod-
els, future IRI predictions were calculated for
each pavement section.
3. Whether future IRI predictions reached the
threshold limit (170 in./mi) was checked.
a. If yes, the RSL value for each pavement sec-
tion was calculated by subtracting the present
year from the year when IRI predictions first
reached the threshold limit.
b. If no, meaning that, based on available mea-
sured IRI data, future IRI predictions had
not reached 170 in./mile over a long period of
analysis time (i.e., 40 years). In other words,
these pavement sections performed very well
in relation to smoothness criteria. Adding
more data points (i.e., future performance
measurements) would change the model and
increase its accuracy. In these cases, the RSL
value for each pavement section was calcu-
lated by subtracting the current age of pave-
ment from 40 years.
Figure 3 shows the distribution of RSL for flexible
and composite (AC over JPCP) pavement sections inves-
tigated in this study, respectively.
Development and Accuracy Evaluations of
Artificial Intelligence-Based Pavement
Performance Model
An approach based on artificial intelligence (AI) was
investigated for developing pavement performance mod-
els for use in network level pavement management. AI
techniques, such as artificial neural networks (ANNs),
have been widely used to model complex pavement engi-
neering problems (20, 21). ANN-based models are very
useful tools for modeling pavement deterioration when
considering many pavement sections with various traffic,
thickness (network level), or deterioration trends. They
are also very fast tools with which thousands of pave-
ment scenarios for which various traffic, thickness, and
conditions can be solved in seconds. Both of these
6 Transportation Research Record 00(0)

(a) (b)
40 40
38
Remaining service life (Years)

Remaining service life (Years)

35 36
34
32
29

29
29
30
28
28
28
30
27
26
26

26
26

26
26
28

25
25

25

25

25
24
26
25

23

23
24
22

19
20 20
18
14

14
16
15

12
14
12
10 10

8
8

6
5

6
5

3
4

2
1

0
0 0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 0 5 10 15 20 25 30
Pavement section ID Pavement length (Miles)
(c) (d)
40
39

Remaining service life (Years)

35 36
33
Remaining service life (Years)

31

31

31

31

31
33
30
30

30

30
30
30
30
29

29

29

29
30
28
28
28

28

28
30
27
27

27
27
26

26

26
25
25

25
25
25

27
24

24

25 24
21
20
18
15

15 15
12
11

12
10 9
7

6
5
4

4
3

3
2

2
1

1
0

0
0

0 0

0 5 10 15 20 25 30
11
13
15
17
19
21
23
25
27
29
31
33
35
37
39
41
43
45
47
49
51
53
55
57
59
1
3
5
7
9

Pavement section ID Pavement length (Miles)

Figure 3. Distribution of remaining service life for flexible pavement sections: (a) based on pavement section ID and (b) based on
pavement length; and for composite (AC over JPCP) pavement sections: (c) based on pavement section ID and (d) based on pavement
length. AC = asphalt concrete; JPCP = jointed plain concrete pavement.

features of ANN models make them useful tools to be used each pavement section using developed ANN models.
in the development of network level pavement performance This tool calculates future pavement performance predic-
modeling. In this study, an ANN-based pavement perfor- tions for any pavement performance indicator (i.e., IRI,
mance model was developed for each pavement perfor- each distress) of any pavement section.
mance indicator (i.e., distress, IRI) and for each pavement The following steps were used in the development of
type: flexible and composite (AC over JPCP). In the model this tool:
development, 80% of all data points in each pavement type
was used, and out of this set of data points 48%, 8%, and 1. ANN models were developed in the MATLABÒ
24% were used as training, testing, and validation datasets, environment using six training algorithms and a
respectively. The remaining 20% of the data points were variable number of hidden neurons (from five to
not used in model development but rather were used as an 60).
independent testing dataset. 2. The ANN model producing highest accuracy was
ANN models must have the following capabilities: selected as the final model for the given pavement
performance indicator.
 High accuracy: They must successfully produce 3. Weights and biases for the final ANN model were
results very similar to those from measured extracted into the automation tool.
distresses. 4. Using these extracted weights and biases, through
 Physically meaningful future Distress predictions: matrix multiplications, future distress predictions
Distress predictions must increase in the future were calculated for the given thickness, accumu-
unless a maintenance or repair activity occurs. lated equivalent single axle load (ESAL) traffic,
age, and previous two years’ pavement perfor-
An automation tool based on Microsoft Excel Macro mance records for any pavement performance
for network level pavement performance prediction was indicator. Compound truck traffic growth of 1%
developed that predicts future pavement performance for per year was assumed in calculating future traffic.
Kaya et al 7

Table 2. Summary of Input and Output Parameters Used in Development of Five ANN Models for Flexible and Composite Pavements

Model name Input parameters Output parameter

Rutting AC thickness, traffic (accumulated ESALs), age, rut (i–2) year, Rut (i) year
rut (i–1) year
Longitudinal cracking AC thickness, traffic (accumulated ESALs), age, longitudinal Longitudinal cracking (i) year
cracking (i–2) year, longitudinal cracking (i–1) year
Transverse cracking AC thickness, traffic (accumulated ESALs), age, transverse Transverse cracking (i) year
cracking (i–2) year, transverse cracking (i–1) year
IRI (approach 1) AC thickness, traffic (accumulated ESALs), age, IRI (i–2) year, IRI (i) year
IRI (i–1) year
IRI (approach 2) Age, rut (i) year, longitudinal cracking (i) year, transverse IRI (i) year
cracking (i) year, IRI (i–2) year, IRI (i–1) year

Note: AC = asphalt concrete; ANN = artificial neural network; ESALs = equivalent single axle loads; IRI = international roughness index.

As part of this study, an ANN model for each pave-
ment type was developed for rutting, longitudinal crack-
ing, transverse cracking, and IRI predictions. Thirty-five
flexible pavement sections with 360 data points of each
pavement performance indicator were used in model
development and independent testing. Of the 360 data
points, 172, 30, 86, and 72 data points were used, respec-
tively, as training, testing, validation, and independent
testing datasets. Sixty composite pavement sections with
524 data points of each pavement performance indicator
were used in model development and independent test-
ing. Of the 524 data points, 251, 42, 126, and 105 data
points were used, respectively, as training, testing, valida-
tion, and independent testing datasets. Table 2 sum-
marizes input and output parameters used in the five
ANN models developed for flexible and composite pave-
ments. As can be seen in Table 2, AC thickness, traffic
(accumulated ESALs), age, and previous two years’
pavement performance records were used in rutting,
longitudinal cracking, transverse cracking, and IRI
(approach 1) model development. On the other hand, in
approach 2, an IRI model was developed using age, mea-
sured distress values (rutting, longitudinal cracking, and
transverse cracking in this case), and previous two years’
measured IRI data. In approach 2, rutting and longitudi-
nal and transverse cracking values predicted by the
ANN model, along with other input parameters, were
used as inputs to predict future IRI.
Figure 4 compares measured pavement condition
records and ANN model predictions for flexible and
composite (AC over JPCP) pavements using: (a) rutting,
(b) longitudinal cracking, (c) transverse cracking, (d) IRI
(approach 1), and (e) IRI (approach 2) ANN models,
respectively. While the ANN models accurately predicted
corresponding pavement performance indicators, the IRI
models produced more accurate predictions compared
with the rutting, longitudinal cracking, and transverse
cracking models because of their higher R2 and lower
AAE and SEE values. The IRI models developed using
approach 1 and approach 2 produced similar accuracies.
In all cases investigated, high R2 and low AAE and SEE
values were obtained for all training, testing, validation
and independent testing datasets.
Development and Results of AI-Based
Pavement RSL Model
Once AI-based network level pavement performance
models were developed for each pavement performance
indicator or condition metric, as explained in the previ-
ous section, the RSL for each pavement section in a road
network could be calculated using these performance
models and corresponding threshold limits for the pave-
ment performance indicators. In this study, IRI was used
as the performance indicator for network level RSL cal-
culations because, as stated earlier, this condition metric
was determined by FHWA (1, 14) (Table 1). RSL is
determined based on the year when future performance
predictions reach the ‘‘poor’’ condition threshold.
The RSL value for each pavement section in a road
network was calculated based on the following steps:
1. Using developed AI-based pavement performance
models, future pavement condition predictions
were calculated for each pavement section.
2. Whether future pavement condition predictions
reached threshold limits was checked for each cor-
responding condition metric.
a. If yes, RSL value for each pavement section
was calculated by subtracting the present
year from the year when pavement condition
predictions first reached the threshold limit.
b. If no, based on available pavement condition
data, this means that future pavement condi-
tion predictions do not reach corresponding
condition metric threshold over a long
8 Transportation Research Record 00(0)

0.5 AAETraining = 0.01 SEETraining = 0.01

(a)
AAETesting = 0.01 SEETesting = 0.01
AAEValidation = 0.01 SEEValidation = 0.01
0.4 AAEInd.Testing= 0.01 SEEInd.Testing= 0.01

Predicted Rutting (in)

R2Training = 0.974
R2Testing = 0.971
0.3 R2Validation = 0.962
R2Ind.Testing= 0.972

0.2
Training
Testing
Validation
0.1
Independent Testing

0.0
0.0 0.1 0.2 0.3 0.4 0.5
Measured Rutting (in)
(b) (c)
10,000 7,000
Predicted Longitudinal Cracking

AAETraining = 142 SEETraining = 248 AAETraining = 95 SEETraining = 201

9,000 AAETesting = 141 SEETesting = 211 AAETesting = 138 SEETesting = 281

Predicted Transverse Cracking

6,000
8,000 AAEValidation = 133 SEEValidation = 236 AAEValidation = 123 SEEValidation = 252
AAEInd.Testing= 144 SEEInd.Testing= 220 AAEInd.Testing= 105 SEEInd.Testing= 190
7,000 R2Training = 0.974 5,000 R2Training = 0.978
R2Testing = 0.976 R2Testing = 0.975
(ft/mile)

6,000
R2Validation = 0.972 4,000 R2Validation = 0.956
R2Ind.Testing= 0.981

(ft/mile)
5,000 R2Ind.Testing= 0.978
4,000 3,000
Training
Training
3,000 Testing
2,000 Testing
Validation
2,000 Validation
Independent Testing
1,000 1,000 Independent Testing

0
0 2,000 4,000 6,000 8,000 10,000 0
0 1,000 2,000 3,000 4,000 5,000 6,000 7,000
Measured Longitudinal Cracking (ft/mile)
Measured Transverse Cracking (ft/mile)
(d) (e)
250 AAETraining = 1.17 SEETraining = 1.90 250 AAETraining = 1.19 SEETraining = 1.86
AAETesting = 1.21 SEETesting = 1.85 AAETesting = 1.65 SEETesting = 2.19
AAEValidation = 1.50 SEEValidation = 3.05 AAEValidation = 1.76 SEEValidation = 3.01
200 AAEInd.Testing= 1.32 SEEInd.Testing= 2.40
Predicted IRI (in/mile)

200 AAEInd.Testing= 1.59 SEEInd.Testing= 2.48

Predicted IRI (in/mile)

R2Training = 0.994 R2Training = 0.995

R2Testing = 0.997 R2Testing = 0.985
150 R2Validation = 0.984 R2Validation = 0.983
R2Ind.Testing= 0.989 150
R2Ind.Testing= 0.989

100 100
Training Training
Testing Testing
Validation
50 50 Validation
Independent Testing Independent Testing

0 0
0 50 100 150 200 250 0 50 100 150 200 250
Measured IRI (in/mile) Measured IRI (in/mile)
(f) (g)
0.5 AAETraining = 0.01 SEETraining = 0.01 10,000 AAETraining = 98 SEETraining = 173
AAETesting = 0.01 SEETesting = 0.01
AAETesting = 102 SEETesting = 177
AAEValidation = 0.01 SEEValidation = 0.01 9,000
Predicted Longitudinal Cracking

AAEValidation = 111 SEEValidation = 178

0.4 AAEInd.Testing= 0.01 SEEInd.Testing= 0.01
Predicted Rutting (in)

8,000 AAEInd.Testing= 117 SEEInd.Testing= 178

R2Training = 0.976 R2Training = 0.986
R2Testing = 0.934 7,000 R2Testing = 0.970
0.3 R2Validation = 0.956 R2Validation = 0.987
R2Ind.Testing= 0.969 6,000
R2Ind.Testing= 0.985
(ft/mile)

5,000
0.2 4,000
Training Training
Testing 3,000 Testing
Validation Validation
0.1 2,000
Independent Testing Independent Testing
1,000

0.0 0
0.0 0.1 0.2 0.3 0.4 0.5 0 2,000 4,000 6,000 8,000 10,000
Measured Rutting (in) Measured Longitudinal Cracking (ft/mile)

Figure 4. Comparison between measured pavement condition records and artificial neural network (ANN) model predictions using: (a)
and (f) rutting, (b) and (g) longitudinal cracking, (c) and (h) transverse cracking, (d) and (i) IRI (approach 1), and (e) and (j) IRI (approach 2)
ANN models for flexible and composite pavements, respectively. AAE = absolute average error; IRI = international roughness index; SEE
= standard error of the estimates.
Kaya et al 9

(h) 10,000 AAETraining = 138 SEETraining = 216

AAETesting = 140 SEETesting = 202
9,000

Predicted Transverse Cracking

AAEValidation = 139 SEEValidation = 220
8,000 AAEInd.Testing= 150 SEEInd.Testing= 286
R2Training = 0.980
7,000 R2Testing = 0.977
6,000
R2Validation = 0.982
R2Ind.Testing= 0.972

(ft/mile)
5,000

4,000
Training
3,000 Testing
Validation
2,000
Independent Testing
1,000

0
0 2,000 4,000 6,000 8,000 10,000
Measured Transverse Cracking (ft/mile)

(i) (j)
250 AAETraining = 1.26 SEETraining = 2.00 250 AAETraining = 1.86 SEETraining = 2.80
AAETesting = 1.73 SEETesting = 3.38 AAETesting = 2.22 SEETesting = 4.11
AAEValidation = 2.30 SEEValidation = 4.44 AAEValidation = 1.91 SEEValidation = 2.93
200 AAEInd.Testing= 1.62 SEEInd.Testing= 2.55 AAEInd.Testing= 1.65 SEEInd.Testing= 2.19
Predcited IRI (in/mile)

200
R2Training = 0.996 Predcited IRI (in/mile) R2Training = 0.992
R2Testing = 0.982 R2Testing = 0.980
150 R2Validation = 0.978 R2Validation = 0.991
150
R2Ind.Testing= 0.993 R2Ind.Testing= 0.995

100 100
Training Training
Testing Testing
Validation Validation
50 50
Independent Testing Independent Testing

0 0
0 50 100 150 200 250 0 50 100 150 200 250
Predicted IRI (in/mile) Measured IRI (in/mile)

Figure 4. (Continued)

period of analysis time (i.e., 40 years). In Comparisons between Statistics and AI-
other words, this means that these pavement Based Network Level Pavement
sections perform very well in relation to the Performance Models
corresponding condition metric, although
adding more data points (i.e., future perfor- Network level IRI performance models were developed
mance measurements) would increase accu- using the statistically defined approach for both flexible
racy of the predictions. In these cases, RSL and composite (AC over JPCP) pavement types to evalu-
value for each pavement section was calcu- ate their relative success in modeling network level pave-
lated by subtracting the current age of pave- ment performance in comparison with those developed
ment from 40 years. using ANN-based approaches. The input parameters
used for developing statistics-based network level IRI
Figure 5 shows the distribution of RSL for 35 flexible models for both pavement types were the same as the
and 60 composite (AC over JPCP) pavement sections ANN-based network level IRI (approach 1) models:
when: (i) an IRI threshold limit of 170 in./mile was used
as the threshold limit and (b) an ANN-based network  Input parameters: AC thickness, traffic (accumu-
level IRI model (approach 1) was used as the pavement lated ESALs), age, IRI (i–2) year, IRI (i–1) year
performance model in the calculation of RSL values.  Output parameter: IRI (i) year
10 Transportation Research Record 00(0)

(a) (b)
40 40
38
Remaining service life (Years)

Remaining service life (Years)

36
35 34

30
32
30 30
28
26
25 24
21

21
20

20
22
19
19
18

20 20
16
16

16

18
15

15
15
15

15
14
16
15 14

12

12

12
11

11
12
10
10
10

9
8

8
8

6
6

5
5

4
4
2

0
0
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 0 5 10 15 20 25 30
Pavement section ID Pavement length (Miles)
(c) (d)
40 40
38
Remaining service life (Years)

Remaining service life (Years)

36
35 34

31
30
30
30
30
32
29

30 30
28

26

26
25

26
25
23

24
22
20 20
18
18

18

18
15
15

16
14

14

14

14

15
13

13
13

14
12

12
11

11

11

12
10
10
10

10
10

10
10

10
10
9
9
9

9
9
9
9
8

8
8

8
6
5 4
3

3
2

2
0

0
0

0
0

0
0
0 5 10 15 20 25 30
17

41
11
13
15

19
21
23
25
27
29
31
33
35
37
39

43
45
47
49
51
53
55
57
59
1
3
5
7
9

Pavement section ID Pavement length (Miles)

Figure 5. Distribution of remaining service life for flexible pavement sections (a) based on pavement section ID and (b) based on
pavement length, and for composite pavement sections (c) based on pavement section ID and (d) based on pavement length. International
roughness index (IRI) model (approach 1) threshold limit of 170 in./mile used in (b) and (d).

The same generalized sigmoidal equation (Equation Figure 6a compares the accuracies of statistics and
4) was also used in the development of network level sta- ANN-based network level IRI models for flexible pave-
tistical models, and the same methodology, error minimi- ments. As can be seen in the figure, the ANN model pro-
zation, was used in the optimization of network level duced greater accuracy with higher R2 and lower AAE
statistical models. and SEE values than the statistics model.

Flexible Pavement Case Composite (AC over JPCP) Pavement Case

A globally optimized sigmoid equation (Equation 5) was A globally optimized sigmoid equation (Equation 6) was
developed by correlating the coefficients of the sigmoidal developed by correlating the coefficients of the sigmoidal
equation (C1, C2, C3, and C4) with the input parameters equation (C1, C2, C3, and C4) with the input parameters
for the whole dataset of model development (35 flexible
for the entire dataset of model development [60 compo-
pavement sections [360 data points])
site pavement sections (524 data points)]
C2
IRI = C1 + ð5Þ C2
1 + eð17:57 + 0:93 3 ageÞ IRI = C1 + ð6Þ
1 + eð17:57 + 0:93 3 ageÞ
where
where
 C1=7.52E-7 3 ACC Traffic – 2.11 3 AC Thickness
+ 1.04 3 IRI (i–2) year + 0.32 3 IRI (i–1) year  C1=1.37E-7 3 ACC Traffic – 2.12 3 AC Thickness
 C2=20.04 3 ACC Traffic – 2.00 3 AC Thickness + 0.82 3 IRI (i–2) year + 0.30 3 IRI (i-1) year
+ 2.94 3 IRI (i–2) year + 3.90 3 IRI (i–1) year  C2=20.04 3 ACC Traffic – 2.00 3 AC Thickness
+ 2.94 3 IRI (i–2) year + 3.90 3 IRI (i–1) year
A model with the model architecture of 5-5-1 (number
of inputs-number of hidden neurons-number of outputs) A model with the model architecture of 5-5-1 was used
was used as the network level ANN model. as the network level ANN model.
Kaya et al
Figure 6. Accuracy comparisons between statistics and ANN-based network level international roughness index (IRI) models for (a)
flexible and (b) composite pavements. AAE = absolute average error; ANN = artificial neural network; SEE = standard error of the estimates.
Figure 6b compares the accuracies of the statistics and
ANN-based network level IRI models for composite
pavements. As can be seen in the figure, the ANN model
produced more accuracy with higher R2 and lower AAE
and SEE values than the statistical model.
Conclusions and Recommendations
Overall Conclusions
In this study, a detailed step-by-step methodology for
development of a framework for pavement performance
and RSL prediction models was established and
explained using real pavement performance data
obtained from the Iowa DOT PMIS database. To
develop RSL models, project level and network level
pavement performance models were initially developed
using two approaches: a statistically defined approach
for project level model development and an AI-based
approach for network level model development. Then,
using threshold limits for various pavement performance
indicators (IRI for project level and network level mod-
els) and FHWA-specified threshold limits for pavement
performance indicators, RSL models were developed for
flexible, and composite (AC over JPCP) pavements.
A statistically defined sigmoid pavement deterioration
curve-based approach was used for development of the
project level pavement performance model. Sigmoidal
equations have been particularly used in the statistics
model development because: (a) they have a low initial
slope that increases with time, and (b) they follow a trend
in which pavement condition always gets worse and dam-
age is irreversible, and both these features make these
models mimic the pavement deterioration behavior
observed in field studies. Sigmoidal equations were found
11
to model pavement deterioration successfully when there
is a single pavement deterioration trend (project level).
One of the benefits of project level pavement perfor-
mance models is that they can be developed using very
little data. Therefore, they can be extensively used when
only limited data on pavement condition or structural
and traffic are available for pavement sections.
AI-based pavement performance models were used for
development of the network level pavement performance
model in this study. AI techniques such as ANN-based
models have been found to be good tools for modeling
pavement deterioration when there are many pavement
sections with various traffic, thickness, and other various
deterioration trends (network level). They are also very
fast tools that can solve thousands of pavement scenarios
with various traffic, thickness, and conditions in seconds
(near real time). Both these features of ANN models make
them good tools for use in development of network level
pavement performance modeling.
As part of this study, network level pavement perfor-
mance models were also developed using statistics and
ANN-based approaches, with identical input parameters
used in both approaches to evaluate their relative success
for network level pavement performance modeling. It
was found that network level ANN-based pavement per-
formance models produced greater accuracy with higher
R2 and lower AAE and SEE values compared with net-
work level statistical models.
As part of this study, Microsoft Excel-based automa-
tion tools were developed for both project level and net-
work level pavement performance modeling and analysis:
 The project level pavement performance modeling
and RSL calculation tool is capable of developing
project-based statistical models for predicting
12
future pavement performance as well as calculat-
ing RSL values based on user-defined threshold
limits. It is also capable of automatically updating
and improving pavement performance prediction
models because it allows more data to be added
into the model development dataset. The benefit
of this tool is that, as engineers add more data into
the model development dataset, they will be able
to refine performance prediction models automati-
cally and make decisions using more recent and
more accurate pavement performance models.
 The network level pavement performance
modeling tool is capable of making pavement per-
formance predictions based on pre-developed
ANN-based pavement performance models.
While having only thickness, traffic, age, and pre-
vious two years’ pavement performance records
for any pavement performance indicator, it can
make future pavement performance calculations
in less than one second for any pavement section.
It is also capable of producing pavement perfor-
mance predictions for thousands of pavement sce-
narios under various traffic, thickness, and other
conditions in seconds. The network level pave-
ment performance modeling tool is also capable
of (a) making future pavement performance pre-
dictions for some distresses (i.e., transverse crack-
ing, rutting, and longitudinal cracking), then (b)
using these predicted distress values as inputs in
making future IRI predictions.
The pavement performance modeling and prediction
models developed and explained in this study are very
powerful and versatile tools that can easily be adopted
by federal and state highway agencies, county engineer
offices, and so on for predicting the future performance
of their transportation infrastructure systems and can
easily be used as a decision-making tool in managing
their transportation infrastructure assets
Recommendations
This study can be further expanded by: (i) including other
pavement performance indicators (i.e., faulting for rigid
pavements, material-related distresses, etc.), (ii) defining
other agency-specific threshold limits, and (iii) prioritizing
some pavement performance indicators over others, and so
on, as part of RSL model development. Some state high-
way agencies use decision trees to determine when a major
rehabilitation or reconstruction is needed. Multi-objective
RSL models can be developed considering various pave-
ment performance indicators with different priorities.
RSL results will allow agencies to distinguish between
two pavement sections with the same current condition
(i.e., the same current IRI) but having different
Transportation Research Record 00(0)
performance behaviors with time. This can be an ideal
approach to addressing the transportation planning and
performance management criteria requirements of the
MAP-21 legislation.
Note that RSL models are only intended to help engi-
neers in their decision-making processes. They consider
only a limited number of condition metrics (IRI, some
distresses, etc.) but may fail to consider other important
parameters, such as the structural capacity and integrity
of pavement systems. Engineers should consider various
parameters as well as RSL model results, combined with
their engineering judgment to determine when a pavement
section will fail and need major rehabilitation or recon-
struction. Note that calculated RSL results in this study
are based on a limited number of dataset elements, devel-
oped pavement performance models and FHWA-speci-
fied threshold limits. Adding more data points (i.e., future
performance measurements) would change the pavement
performance models as well as the calculated RSL results.
Acknowledgments
The authors gratefully acknowledge the Iowa Highway
Research Board and Iowa County Engineers Service Bureau for
supporting this study. The project technical advisory committee
(TAC) members from Iowa County Engineers Association
(ICEA), including Lee Bjerke, Zach Gunsolley, Todd Kinney,
Mark Nahra, John Riherd, Brad Skinner, and Jacob Thorius,
are gratefully acknowledged for their guidance, support, and
direction throughout the research. Special thanks are expressed
to Steve De Vries and Danny Waid, who developed the original
concept of this study.
Author Contributions
The authors confirm contribution to the paper as follows: study
conception and design: Halil Ceylan, Orhan Kaya, Sunghwan
Kim, Danny Waid, and Brian P. Moore; data collection: Orhan
Kaya and Sunghwan Kim; analysis and interpretation of results:
Orhan Kaya, Halil Ceylan, Sunghwan Kim, Danny Waid, and
Brian P. Moore: draft manuscript preparation: Orhan Kaya,
Halil Ceylan, and Sunghwan Kim. All authors reviewed the
results and approved the final version of the manuscript.
Declaration of Conflicting Interests
The authors declared no potential conflicts of interest with
respect to the research, authorship, and/or publication of this
article.
Funding
The authors received no financial support for the research,
authorship, and/or publication of this article.
References
1. H.R. 4348, Moving Ahead for Progress in the 21st Century
Act (MAP-21). An Act to Authorize Funds for Federal-Aid
Highways, Highway Safety Programs, and Transit
Kaya et al
Programs, and for Other Purposes. 112 Congress, 2nd Ses-
sion. Enacted October 1, 2012.
2. Federal Highway Administration. Pavement Health Track
Remaining Service Life (RSL) Forecasting Models, Techni-
cal Information. Federal Highway Administration,
Washington, D.C., 2018.
3. Elkins, G. E., T. M. Thompson, J. L. Groeger, B. Visin-
tine, and G. R. Rada. Reformulated Pavement Remaining
Service Life Framework. Federal Highway Administration,
Washington, D.C., 2013.
4. Mack, J. W., and R. L. Sullivan. Using Remaining Service
Life as the National Performance Measure of Pavement
Assets. Presented at 93rd Annual Meeting of the Transpor-
tation Research Board, Washington, D.C., 2014.
5. Chen, D., and N. Mastin. Sigmoidal Models for Predicting
Pavement Performance Conditions. Journal of Perfor-
mance of Constructed Facilities, Vol. 30, No. 4, 2016, p.
04015078.
6. Sundin, S., and C. Braban-Lexdoux. Artificial Intelligence-
Based Decision Support Technologies in Pavement Man-
agement. Computer Aided Civil and Infrastructure Engi-
neering, Vol. 16, No. 2, 2001, pp. 143–157.
7. Albuquerque, N. M., and M. Broten. Local Agency Pave-
ment Management Application Guide. Washington State
Department of Transportation, Olympia, WA, 1997.
8. Pavement Management Guide. American Association of
State Highway and Transportation Officials, Washington,
D.C., 2001.
9. Wolters, A. S., and K. A. Zimmerman. Current Practices
in Pavement Performance Modeling. Project 08-03 (C07),
Task 4 Report: Final Summary of Findings, Pennsylvania
Department of Transportation, Harrisburg, PA, 2010.
10. Yu, J. Pavement Service Life Estimation and Condition Pre-
diction. University of Toledo, Toledo, Ohio, 2005.
11. Abdallah, L., O. Melchor-Lucero, C. Ferregut, and S. Nazar-
ian. Artificial Neural Network Models for Assessing Remain-
ing Life of Flexible Pavements. Research Report 1711-2.
Texas Department of Transportation, Austin, TX, 2000.
12. Zaghloul, S., and M. Elfino. Pavement Rehabilitation
Selection Based on Mechanistic Analysis and Field
Diagnosis of Falling Weight Deflectometer Data - Virginia
Experience. Transportation Research Record: Journal of the
Transportation Research Board, 2000. 1730: 177–186.
13. Yang, J., J. J. Lu, M. Gunaratne, and Q. Xiang. Forecast-
ing Overall Pavement Condition with Neural Networks
13
Application on Florida Highway Network. Transportation
Research Record: Journal of the Transportation Research
Board, 2003. 1853: 3–12.
14. FHWA. Guidelines for Informing Decision Making to Affect
Pavement Performance Measures: Final Report. Publication
No. FHWA-HRT-17-090. Federal Highway Administra-
tion, McLean, VA, 2018.
15. Kaya, O. Investigation of AASHTOWare Pavement ME
Design/Darwin-MEä Performance Prediction Models for
Iowa Pavement Analysis and Design. MSc thesis. Iowa State
University, Ames, IA, 2015.
16. Beckley, M. E. Pavement Deterioration Modeling using His-
torical Roughness Data. MSc thesis. Arizona State Univer-
sity, Tempe, AZ, 2016.
17. Ercisli, S. Development of Enhanced Pavement Deterioration
Curves. MSc thesis. Virginia Polytechnic Institute and State
University, Blacksburg, VA, 2015.
18. Kaya, O., A. Rezaei-Tarahomi, H. Ceylan, K. Gopalak-
rishnan, S. Kim, and D. R. Brill. Developing Rigid Airport
Pavement Multiple-Slab Response Models for Top-Down
Cracking Mode using Artificial Neural Networks. Pre-
sented at 96th Annual Meeting of the Transportation
Research Board, Washington, D.C., 2017.
19. Miller, J. S., and W. Y. Bellinger. Distress Identification
Manual for the Long Term Pavement Performance Program.
No. FHWA-RD-03-031. Federal Highway Administration,
McLean, VA, 2003.
20. Kaya, O., A. Rezaei-Tarahomi, H. Ceylan, K. Gopalakrish-
nan, S. Kim, and D. R. Brill. Neural Network-Based Multi-
ple-Slab Response Models for Top-Down Cracking Mode in
Airfield Pavement Design. Journal of Transportation Engineer-
ing, Part B: Pavements, Vol. 144, No. 2, 2018, p. 04018009.
21. Kaya, O., N. Garg, H. Ceylan, and S. Kim. Development
of Artificial Neural Networks Based Predictive Models for
Dynamic Modulus of Airfield Pavement Asphalt Mixtures.
Proc., International Conference on Transportation and
Development, Pittsburgh, PA, 2018.
The contents of this paper reflect the views of the authors who
are responsible for the facts and accuracy of the data presented.
The contents do not necessarily reflect the official views and poli-
cies of the Iowa Highway Research Board, Iowa County
Engineers Service Bureau, or Iowa State University. This paper
does not constitute a standard, specification, or regulation.