DOI 10.1007/s12205-010-0123-7
www.springer.com/12205
···································································································································································································································
Abstract
This paper describes the application of two decision tree algorithms, Logistic Regression Trees with Unbiased Selection (LOTUS)
and Classification Rule with Unbiased Interaction Selection and Estimation (CRUISE), to identify prevailing distress types that affect
pavement maintenance decision making. The maintenance history of the selected sections within Interstate highway I-90 in the State
of Wisconsin is modeled to build decision trees. The historical field distress survey data include the types, quantities, severities, and
locations of distresses. The primary objective of this paper is to present the decision tree algorithms that may be incorporated into the
decision making process and help state highway agencies make rational decisions during the course of pavement management. The
paper discusses the model development process and prevailing distress identification procedures employing the decision tree
algorithms. A case study is conducted on another Interstate highway, I-43, with a view to furthering the analysis of the results from
the two algorithms. These analysis models and the results will provide the state highway agencies and practitioners with a means to
identify prevailing distress types in a specific location and facilitate making decisions on future maintenance practices.
Keywords: Pavement Management System (PMS), maintenance decision support, highway maintenance factors, field distress
survey, decision tree algorithm
···································································································································································································································
*Graduate Research Assistant, Dept. of Civil and Environmental Engineering, University of Wisconsin-Madison, Madison, WI 53706, USA (E-mail:
myungookkang@wisc.edu)
**Member, Senior Staff Engineer, MMI Engineering, Houston, TX 77077, USA (E-mail: minkwank@gmail.com)
***Assistant Professor, Dept. of Construction Science and Organizational Leadership, Purdue University Calumet, Hammond, IN 46323, USA (Correspond-
ing Author, E-mail: leej@calumet.purdue.edu)
− 123 −
Myungook Kang, Minkwan Kim, and Joo Hyoung Lee
selection process are then identified in accordance with each weighted and factored in a formula to compute the PDI that
failure mode. These days, many State agencies established their ranges from 0 to 100, with 0 representing an excellent condition
own assessment system and evaluation criteria in order to judge and 100 being worst (Shober and Friedrichs, 1998). Typically,
the level of distress and corresponding treatments. Some examples the PDI of a pavement gradually increases as the pavement ages;
include PASER (Pavement Surface Evaluation and Rating) it dramatically drops when maintenance on the pavement is
published by the State of Wisconsin (Walker, 2002) and OPI conducted.
(Organizational Performance Index) and Maintenance Quality
Survey Manual available in the States of Ohio and Indiana 3. Data Modeling
(ODOT, 2001; INDOT, 2006). The State agencies specify the
treatments on the basis of existing pavement distresses, antici- In an effort to explore prevailing distress types that triggered
pated traffic loading, environmental conditions, cost of the treat- maintenance activities, total 112 rigid pavement sections are
ment, and the performance history of different treatments. selected and 1,196 inspection data from 1985 to 2005 in the PIF
Generally, the distress types and the history of maintenance are are analyzed. Table 1 shows detailed information of the collected
location specific in that it is essential to investigate prevailing data and the descriptions of variables. In this study, the levels of
distress types in a specific segment of an Interstate highway. This distresses, i.e. none, moderate and severe, are considered inde-
paper thus focuses on physical distresses on rigid pavement and pendent variables. Maintenance is a dependent variable that tells
pavement condition assessment instruments in the State of whether maintenance is conducted in the selected sections. This
Wisconsin. According to Distress Identification Manual provided is determined by abrupt drop of the PDI at a certain point in time.
by Federal Highway Administration (FHWA, 2003), typical Total seven observable distress indicators are recorded for
jointed concrete pavement has four distress categories: cracking, rigid pavement in the PIF. These are slab breakup (Sbkup),
joint deficiencies, surface defects, and miscellaneous distresses. distressed joint (Djt), patching (Ptch), surface distress (Sur),
Considering the federal guideline, the State of Wisconsin uses longitudinal cracking (Ljt), transverse faulting (Flt), and Rutting
seven distress indicators for rigid pavement. The distress indi- (Rut). Slab breakup is excluded in the analysis because none of
cators include roughened pavement, deteriorated joints, damaged the selected segments exhibits this distress type. Instead, Inter-
patching, and distressed surfaces such as cracks, faulting, and national Roughness Index (IRI) is added. Although the IRI is
rutting. In this study selected Interstate highway segments within generally not considered as a distress measurement, it is one of
I-90 are modeled and analyzed using decision tree algorithms in the indices considered during decision making process because it
an effort to explore prevailing distress types based on the main- affects riding quality as well as driver’s safety. The condition of
tenance history. These analyses will support the state agency to pavement section is represented by categorical values in accor-
make decisions on future maintenance practices. A case study is dance with the severity and extent of a distress type, while the
conducted on another Interstate highway, I-43, with a view to IRI is expressed by a numerical value in meter per kilometer.
furthering the analysis of the results from the two algorithms. Table 2 shows the extent of distressed joint associated with
corresponding categorical values as an example. The variable
2. Pavement Evaluation Process for State High- represents the extent of each distress type, while the category
ways in Wisconsin refers to the severity of the distress type. The collected data are
then analyzed to determine whether there is a correlation
The State of Wisconsin has a dual pavement management between certain distress types and the maintenance event. Fig. 1
system: Pavement Information File (PIF) and Wisconsin Infor- shows the data distribution of each variable in the I-90 model to
mation System for Local Road (WISLR). The PIF is for state be analyzed. IRI is expressed by a numerical value in meter per
highways and the WISLR is for local roads. Both of them kilometer.
archive the historical pavement performance data at specific One of the conventional techniques used to determine the
locations. Focused on the state highway system, this study dependency between independent variables and a dependent
analyzes the PIF for further discussions. variable is a regression analysis. There exist, however, some
Since the early 1980s, Wisconsin Department of Transporta- technical issues that limit the use of the traditional regression
tion (WisDOT) has monitored the pavement condition of the analysis. First, the stepwise regression technique will find one
state highway system. The assessed condition of 1.6-km (one- most significant variable and its dependency with the dependent
mile) section is documented in the PIF. The collected historical variable in terms of P-value (Montgomery and Runger, 2007),
pavement performance data include distress type, the severity whereas this study needs to rank the independent variables
and extent of each distress type, and Pavement Distress Index according to the level of dependency with the dependent
(PDI) at each section. The PDI is a double digit index that variable. Second, the logistic regression analysis is known as a
represents overall observable pavement surface distresses by powerful tool to handle data with binary response variables.
severity and extent (Friedrichs, 1999). The severity represents Nonetheless, it is difficult to interpret the result of the logistic
how serious the damage is while the extent tells how much of the regression analysis when it involves a number of independent
section is affected. The extent and severity of the distresses were variables. It is also known that the logistic regression analysis is