This article was downloaded by: [Florida Atlantic University]

On: 11 November 2014, At: 01:09

Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House,
37-41 Mortimer Street, London W1T 3JH, UK

International Journal of Pavement Engineering

Publication details, including instructions for authors and subscription information:
http://www.tandfonline.com/loi/gpav20

Identification of Homogeneous Sections from Road Data

a b
Rahul Misra & Animesh Das
a
Department of Civil Engineering , Indian Institute of Technology , 781 039, Guwahati, India
b
Department of Civil Engineering , Indian Institute of Technology , 208 016, Kanpur, India
Published online: 31 Jan 2007.

To cite this article: Rahul Misra & Animesh Das (2003) Identification of Homogeneous Sections from Road Data, International
Journal of Pavement Engineering, 4:4, 229-233, DOI: 10.1080/10298430410001672237

To link to this article: http://dx.doi.org/10.1080/10298430410001672237

PLEASE SCROLL DOWN FOR ARTICLE

Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”) contained
in the publications on our platform. However, Taylor & Francis, our agents, and our licensors make no
representations or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of the
Content. Any opinions and views expressed in this publication are the opinions and views of the authors, and
are not the views of or endorsed by Taylor & Francis. The accuracy of the Content should not be relied upon and
should be independently verified with primary sources of information. Taylor and Francis shall not be liable for
any losses, actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities whatsoever
or howsoever caused arising directly or indirectly in connection with, in relation to or arising out of the use of
the Content.

This article may be used for research, teaching, and private study purposes. Any substantial or systematic
reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any
form to anyone is expressly forbidden. Terms & Conditions of access and use can be found at http://
www.tandfonline.com/page/terms-and-conditions
 The International Journal of Pavement Engineering, Vol. 4 (4) December 2003, pp. 229–233

Identification of Homogeneous Sections from Road Data

RAHUL MISRAa,† and ANIMESH DASb,*

a
Department of Civil Engineering, Indian Institute of Technology, Guwahati 781 039, India; bDepartment of Civil Engineering,
Indian Institute of Technology, Kanpur 208 016, India

(Received 10 August 2002; Revised 1 January 2004)

For highway design or maintenance projects, a long road stretch needs to be delineated into some
homogeneous sections, on the basis of various pavement responses, such as, subgrade CBR value, field
moisture content, soil type, deflection, pavement distress parameters, roughness, etc. Cumulative
Downloaded by [Florida Atlantic University] at 01:09 11 November 2014

difference approach (CDA) for identification of homogeneous sections is suggested in the AASHTO
Guide for Design of Pavement Structures [AASHTO (1986) Guide for Design of Pavement Structures,
AASHTO, Washington, DC, pp. III-17– III-20 and Appendix J; AASHTO (1993) Guide for Design of
Pavement Structures, AASHTO, Washington, DC, pp. III-19 –III-24 and Appendix J]. This paper
identifies some of the limitations associated with this method and suggests improved yet simplistic
methodology for identification of homogeneous sections based on a combined approach of
classification and regression tree (CART) and exhaustive search. Examples are drawn from synthetic
and realistic data.

Keywords: Homogeneous section; Cumulative difference approach; Change-point; AASHTO Guide;

CART method

INTRODUCTION This paper also attempts to suggest simplistic procedure for

delineation of data into homogeneous sections where the
Pavement condition parameters, like subgrade strength, shortcomings of CDA are taken care of. Examples are
deflection due to test load, skid resistance, or level of presented to show the effectiveness of the proposed
various distresse, vary along a road’s length. These data methodology over the CDA.
are used for design of new roads, or rehabilitation of
existing roads. If design is based on the data of a particular THE CDA
parameter of the whole stretch taken together, it will lead
to an uneconomical design. Therefore, pavement response In the CDA, the term Zx is defined as the difference in the
data along a long stretch of road needs to be delineated cumulative area values at a given distance (x) between
into sections which are “relatively uniform”, referred to as the actual response and the average response (AASHTO,
homogeneous sections, and the design performed 1993). Mathematically,
individually for each of these homogeneous sections. 0X 1
n
This results in economy in design without compromising
Xn
B x i CX
nt
reliability level. Zx ¼ ai 2 @ i¼1 A ai ð1Þ
For identification of homogeneous sections, cumulative i¼1 Lp i¼1
difference approach (CDA) is suggested in the AASHTO’s
Guide (1986, 1993). CROW (1998) also suggests a similar and
methodology. However, there are situations where CDA

becomes sensitive to local variations of data (Divinsky r i21 þ r i
et al., 1997) and fails to delineate homogeneous sections ai ¼ xi ; ð2Þ
2
effectively. Also this method does not offer a choice of
the number of homogeneous sections a designer intends to where, xi is the distance along the road section up to i-th
delineate. These issues have been addressed in this paper. response measurement; n is the n-th pavement response

*Corresponding author. E-mail: adas@iitk.ac.in

†
Current address: Assistant Systems Engineer, Tata Consultancy Services, C-56, Phase-2, Noida 201 305, India. E-mail: rahul_iitg@yahoo.com

ISSN 1029-8436 print/ISSN 1477-268X online q 2003 Taylor & Francis Ltd
DOI: 10.1080/10298430410001672237
 230
FIGURE 1 Identification of homogeneous section by Cumulative
Difference Approach [the data points are taken from the AASHTO’s
Guide (1993) example].
measurement; nt is the total number of pavement response
Downloaded by [Florida Atlantic University] at 01:09 11 November 2014
measurements taken in the project; ri is the pavement
response value of the i-th measurement; Lp is the total
project length.
The Zx values are plotted versus distance along the road
length and the points where the slope of the Zx curve
changes its algebraic sign are taken as the border line bet-
ween the consecutive homogeneous sections (AASHTO,
1993). These points have been described as change-points
(Brodsky and Darkhovsky, 1993, Internet site of Cardiff
University),‡ and are referred to as such in the rest of the
paper.
A sample plot of the pavement response values as
read from in the AASHTO’s Guide (1993) and the
resulting Zx values, are plotted in Fig. 1. The change-
points, thus identified, are shown by vertical lines
intersecting the Zx curve.
Discussion on the CDA
In fact, the CDA is equivalent to finding those points at
which the response curve crosses the mean horizontal line,
which includes all the data-points. This can be explained
as follows. As long as the response plot lies below
the mean line, the Zx value keeps on decreasing, as the
difference between the cumulative and the average area is
negative. The slope of the Zx curve changes its sign only
when the response curve crosses the average line.
This is cross-checked in Fig. 2, where the homogeneous
sections delineated are the same as those delineated by the
CDA procedure shown in Fig. 1.
CDA, though a quick and straight-forward method for
identification of homogeneous sections, has certain
limitations, as identified in the following:
. As mentioned above, CDA is the same as finding
those points at which the response curve crosses the
mean horizontal line. But there may be cases when
more than one homogeneous sections with different
average response levels may exist consecutively,
‡
http://www.cf.ac.uk/maths/stats/changepoint, last accessed on 25th June, 2002)
R. MISRA AND A. DAS
FIGURE 2 Homogeneous sections delineated on the basis of the mean
value [the data points are taken from the AASHTO’s Guide (1993)
example].
FIGURE 3 Cumulative difference approach may fail to identify
existence of other homogeneous sections within a homogeneous section.
and all of them may either be completely above or
below the mean horizontal line. In such a case, Zx
curve does not change its sign, hence CDA fails to
identify the existence of other possible homogeneous
sections. Such a situation is shown in Fig. 3 (plotted
from synthetic data) with AB, BC, CD and DE as
individual homogeneous sections, where both AB and
BC lie below the mean line, and similarly, where
CD and DE lie above the mean line. In such a
situation, CDA could only identify AC and CE as
two homogeneous sections (because the slopes
change in magnitudes at points B and D, but not in
their signs).
. Figure 4 consists of a modified form of response
values of the AASHTO guide example, where the
values in between the points A and B are deliberately
elevated to higher level by adding fixed offset value
to all the data-points in this zone. It is interesting to
observe from Fig. 4 that CDA is now unable to
identify homogeneous sections which were pre-
viously delineated (refer Fig. 1), as the average of
the response has shifted to a larger value. Thus, as
per CDA, the data level of one part of the stretch
affects identification of homogeneous sections of
other part of the stretch, which is undesirable.
 IDENTIFICATION OF HOMOGENEOUS SECTIONS
FIGURE 4 In Cumulative difference approach data level of one part
of the stretch may affect identification of homogenous sections in the
other part.
. As well, CDA does not have any control over the
Downloaded by [Florida Atlantic University] at 01:09 11 November 2014
number of homogeneous sections a designer intends
to choose. Statistically, one should be able to find out
desired number of homogeneous sections from a
given response data of a road stretch, depending on
the reliability level chosen. There may not be any
restriction on maximum length of a homogenous
section, but from practical considerations, there must
be a provision for specifying minimum length of a
homogeneous section. Because, frequent changes in
specification along the road stretch may cause
inconvenience from the construction point of view.
Choice of minimum section length can not be
incorporated in CDA.
These limitations of CDA have prompted the authors to
look for an improved method for delineation of
homogeneous sections, which is easy to implement and
also works efficiently in all such situations. Divinsky et al.
(1997) pointed out that CDA being sensitive to the mean
value of data-points, some methodology based on standard
deviation needs to be devised. He suggested that the whole
data-set could be replaced by moving the standard
deviation of the chosen base-length operated on the same
data, and then the same CDA can be adopted to delineate
the homogeneous sections. However, for obvious reasons,
this method is sensitive to what base-length is chosen, as
shown by Divinsky et al. (1997), who chose a base-length
of five data-points for their study based on road roughness
data of Israel. Also, this method finally uses CDA for
delineation, so the limitations involved with CDA
discussed above are equally valid in this case.
PROPOSED METHOD
Some formulations are available in literature for
estimation of change-points. These are based on multiple
regression, generalized linear models, cluster analysis,
maximum likelihood estimation, autoregressive
moving average, exponential smoothening method,
Bayesian approach, etc. (Basseville and Nikiforov, 1993;
Brodsky and Darkhovsky, 1993; Montgomery et al.,
231
2001). Some of them involve complexities in terms of
implementation, or assumptions related to the distribution
pattern of the response data. The approach suggested in
the present case is simple, basic and does not involve any
assumption related to the distribution pattern of the
response data. The suggested approach is presented
through the following three sub-sections.
Development of Classification and Regression Trees
(CART)
The first step of the approach deals with creation of a large
collection of partitions of the original data set by
recursively dividing the parent section into two child
sections. This can be better explained by taking a simple
example. Figure 5 presents a plot of synthetic random data
where the first 20 data points have a different mean and
standard deviation to the other 20 points. The data-set may
be divided into two parts at any point (say at “A”,
as shown in the figure). A term, squared error,
may be calculated as the sum of squared difference
between the data points and the mean of each part, added
together. Thus,
X
A X
nt
Squared error ¼ ðr i 2 r L Þ2 þ ðr i 2 r R Þ2 ; ð3Þ
i¼1 i¼A
where, r L is the average response value of the left part of A,
r R is the average response value of the right part of A.
Now, by shifting the point A from one end of the data to
the other, it is possible to find a specific position where the
value of squared error is minimum. This is an exhaustive
search. This exhaustive search procedure can easily
divide a given stretch into two homogeneous sections.
However, this becomes an optimization problem when the
given section is to be delineated into homogeneous sections
of more than two in number. The solution in that case is
likely to require a very large computation time, specially
when the data size is large. Instead, the above procedure
could be operated again on the two sections individually
which have just been delineated, resulting in four
delineated sections and so on (Montgomery et al., 2001).
FIGURE 5 Exhaustive search for dividing the data-points into two
homogeneous sections.
 232 R. MISRA AND A. DAS
Downloaded by [Florida Atlantic University] at 01:09 11 November 2014

FIGURE 6 Original tree (a) and sub-tree (b) (for eight delineated sections) for Alaska roughness data (GPS section 1008, 25th May, 1990, Repeat 1).

This can be recursively done on the given data until each TABLE I Sections delineated by the proposed method [distance unit]
section delineated is smaller than a given minimum length
specified based on practical considerations. AASHTO Modified AASHTO
Guide example (Fig. 1) Guide example (Fig. 4)
The result obtained can be represented in the form
of a binary tree which has been illustrated with the help 0.0– 5.0 0.0– 5.0
5.0– 25.0 5.0– 25.5
of a sample data in Fig. 6. The data (Internet site University 25.0–33.5 25.5– 30.5
of Michigan){ represents the roughness level of a particular 33.5–52.5 30.5– 39.0
road section (GPS section 1008, May 1990) in Alaska. 52.5–57.5 39.0– 52.5
57.5–65.0 52.5– 65.0
Figure 6a shows the recursive binary splitting of this data in 65.0–76.5 65.0– 76.5
the form of a binary tree, till it has reached the minimum
specified section length, which has been taken as 50 data-
points in this case.
The Algorithm

Selection of Best Sub-tree
Since the above mentioned binary tree has been
obtained by delineating sections with sizes close to the
minimum specified length, the number of sections
resulted may be considerably large. Now, depending on
the requirement of the number of sections to be
delineated, a designer may choose to suitably combine
some of these sections. This can be represented in the
form of a sub-tree derived from the original binary tree.
Thus, for a given number of homogeneous sections to be
delineated, one may obtain a number of such sub-trees
as possible solutions. The best solution can be identified
as that combination of sections which has the least sum
of standard deviations. Figure 6b now shows the best
sub-tree obtained for the sample data (as used in Fig. 6a),
when the number of sections delineated was restricted
to eight.
Thus, the algorithm can be summarized in the form of
following two stage process:
. The given section is recursively divided into homo-
geneous sections using binary splitting based on
a minimum squared error criterion. If any section
comes out to be less than the permissible minimum
section length, the process aborts further sub-division
for that section.
. The large collection of sections delineated as a result of
previous step is reduced to the required number
of sections by merging individual sections with
other sections. This is done by exhaustive search.
The combination, which gives the least sum of standard
deviations, is the intended delineation.
An algorithm of a similar yet more rigorous in approach
has been suggested in a recent paper by Gey and Lebarbier

{
http://www.umtri.umich.edu/erd/roughness, last accessed on 25th June, 2002
 IDENTIFICATION OF HOMOGENEOUS SECTIONS
FIGURE 7 Plot of the Alaska roughness data and various possible homogeneous sections (GPS section 1008, 25th, May, 1990, Repeat 1).
Downloaded by [Florida Atlantic University] at 01:09 11 November 2014
(2002), where they have used penalty functions when the
number of change points are unknown.
RESULTS AND DISCUSSION
Based on the above algorithm, a computer code has been
developed in C language, and operated on various
synthetic and realistic data. Table I shows the sections
delineated for the AASHTO example (Fig. 1) and modified
data (Fig. 4). An example on modified data (Column 2 of
Table I) shows that this could successfully identify
a separate homogeneous section from distance unit
39.0 –52.5, which was provided deliberately.
Figure 7 shows various number of homogeneous
sections that can be delineated depending on the
requirement of the designer. Corresponding squared errors
are also put in Fig. 7. The data points in this figure have
been taken from the same source as that of used in Fig. 6
(GPS section 1008, May 1990, Internet site of
University of Michigan),{ and the minimum section
length is chosen to be 50 data-points in this case. Thus the
proposed method has a better flexibility in terms of
adjustments the possible in specifying minimum section
length and the number of delineated sections desired.
As per CDA the sections delineated are 0 – 10, 10–
325, 325 – 575, 575 –760 and 760– 1000. These values are
also derived from the proposed method at the first part of
the algorithm, when the minimum section length is specified
as 5. Subsequently, some other homogenous sections than
those delineated by CDA are chosen by the proposed
method. Since the proposed method performs an
exhaustive search on least sum of standard deviations, it
can be said that the sections delineated are statistically
more significant than those of CDA.
Thus, the advantages of the proposed algorithm can be
identified as:
. The algorithm stated above is simple, fast, can work on
any random data of any size, and overcomes the
limitations associated with CDA.
233
. The algorithm does not involve any assumption of the
distribution pattern of the data-points.
. The algorithm takes into account the constraint of
minimum size of section to be delineated.
. It also provides the flexibility to choose the number of
sections to be delineated.
CONCLUSIONS
The limitations of CDA (AASHTO, 1993; CROW, 1998)
for identification of homogeneous sections from road data
of a given stretch are discussed. A combined approach
based on classification and regression tree and exhaustive
search, is suggested in this paper, so as to delineate
homogeneous sections more effectively. The proposed
method takes care of the limitations of the CDA, and is
simple, fast and flexible.
References
(1986) Guide for Design of Pavement Structures (AASHTO, Washington,
DC), pp III-17–III-20 and Appendix J.
(1993) Guide for Design of Pavement Structures (AASHTO, Washington,
DC), pp III-19–III-24 and Appendix J.
Basseville, M. and Nikiforov, N. (1993) The Detection of Abrupt
Changes—Theory and Applications (Prentice-Hall, Englewood
Cliffs, NJ), Information and Systems Sciences Series.
Brodsky, B. and Darkhovsky, B. (1993) Nonparametric Methods in
Change-Point Problems (Kluwer Academic Publishers, Dordrecht,
The Netherlands).
(1998) “Deflection Profile – Not a Pitfall Anymore”, The National
Information and Technology Centre for Transport and Infrastructure
(CROW), The Netherlands, Appendix II, pp.149–153.
Divinsky, M., Nesichi, S. and Livneh, M. (1997) “Development of a Road
Roughness Profile Delineation Procedure”, Journal of Testing and
Evaluation 25(4), 445 –450.
Gey, S. and Lebarbier, E. (2002) “A CART based Algorithm for Detection
of Multiple Change Points in the Mean for Large Samples”, preprint
prepublication no. 2002-10 of Université de Paris XI, http://www.
math.u-psud.fr/ , biblio/ppo/2002/fic/ppo_2002_10.ps, last accessed
on 25th June, 2002.
Montgomery, D.C., Peck, E.A. and Vining, G.G. (2001) Introduction to
Linear Regression Analysis, Wiley Series in Probability and
Statistics, 3rd ed. (Wiley, New York).