Automated Asphalt Pavement Maintenance Classification

using Machine Learning

Group 2:
Asif Akhtab Ronggon
Isha Shrestha
Mostafa Omara
Sagar Parajuli
Saumya Jain
Saurav Upadhyaya
Sidhi Kumari Yadav
Yifeng Cao

Second Spring 2026

Submitted to:
Dr. Jianhua Chen
Machine Learning CSC 7333

Table of Contents

ABSTRACT...................................................................................................................................... 4
INTRODUCTION............................................................................................................................. 5
PROJECT OBJECTIVES..................................................................................................................... 6
SCOPE OF WORK............................................................................................................................ 7
DATA COLLECTION......................................................................................................................... 8
 Source of the Data..............................................................................................................8
 Data Period and Area of Interest........................................................................................8
 Description and Structure of the Dataset............................................................................9
 Preliminary Observations on Data Quality.........................................................................9

Second Spring 2026

Machine Learning CSC 7333

ABSTRACT

Newly paved asphalt roads undergo environmental and traffic loading that cause surface
irregularities, negatively impacting safety, ride quality, and vehicle maintenance costs. Effective
pavement management depends on monitoring these distresses and the resulting roughness. In
Louisiana, pavement condition data is collected every two years and maintained within a
Pavement Management System (PMS) database. However, a significant gap exists between
observing and classifying damage surfaces, as they are mostly identified manually, a labor-
intensive, time-consuming process. This project automates this problem by creating a robust
machine learning model to classify asphalt pavement condition into three functional levels: low,
medium, and high. We will use different Python libraries, such as pandas, NumPy, and scikit-
learn, along with Jupyter Notebook and Google Colab environment.

The methodology involves synthesizing raw distress indicators—such as cracking, rutting,

potholes, and patching—to compute the PCI via weighted factor analysis, while International
Roughness Index (IRI) measurements are integrated directly from the PMS. To optimize
classification accuracy, the study evaluated the efficacy of several algorithms, including
multinomial logistic regression, decision trees, random forest ensembles, and XGBoost. The
performance of these models was rigorously assessed using metrics such as RMSE, MAE,
$R^2$, and F1-macro scores to ensure a high level of predictive reliability.

The results indicate that an optimized model stack provides the most effective solution: the
XGBRegressor is the superior choice for PCI regression (R 2 of 0.9988), the
RandomForestRegressor performs best for IRI regression (R2 of 0.9799), and the XGBClassifier
delivers the highest accuracy for final maintenance classification (98.77%). Furthermore, feature
importance analysis reveals that alligator cracking is the primary determinant for transitioning
from routine upkeep to major structural rehabilitation. These findings offer transportation
authorities a highly accurate, automated framework for classifying pavement damage and
prioritizing infrastructure investments.

Second Spring 2026

Machine Learning CSC 7333

INTRODUCTION

Nearly 92% of roads in the United States are paved with asphalt, typically designed for a service
life of 10 to 15 years. Over time, newly paved asphalt roads undergo environmental and traffic
loading that result in surface irregularities, negatively impacting safety, ride quality, and vehicle
maintenance costs.

The various pavement distress includes alligator cracking, rutting, random cracking, potholes,
patching and roughness. Alligator cracking is characterized by interconnecting cracks forming a
pattern similar to an alligator's skin, this is a load-related distress typically caused by the repeated
application of heavy traffic loads or a weakened base layer. Rutting refers to longitudinal surface
depressions in the wheel paths, often caused by the permanent deformation of any of the
pavement layers or the subgrade under heavy loads. Random cracking are cracks that occur non-
parallel to the direction of travel or due to environmental factors such as temperature cycles
rather than direct traffic loading. Potholes are bowl-shaped depressions in the pavement surface,
while patching refers to portions of the pavement that have been removed and replaced with new
material to repair previous distresses. Effective pavement management depends on the
continuous monitoring of these distresses and the resulting roughness. In Louisiana, pavement
condition data is collected every two years and maintained within a Pavement Management
System (PMS) database.

Despite the availability of this data, a significant gap exists between the observation and
classification of damage surfaces. Currently, maintenance levels are identified through manual
inspection—a process that is notoriously labor-intensive, time-consuming, and subject to
engineering bias. This project automates the workflow by creating a robust machine learning
framework to classify asphalt pavement conditions into four distinct, functional maintenance
levels: Defer Action, Pavement Preservation, Pavement Rehabilitation, and Pavement
Reconstruction.

To develop this framework, the study utilizes a comprehensive suite of Python libraries,
including pandas, NumPy, and scikit-learn, implemented within Jupyter Notebook and Google
Colab environments. The methodology synthesizes raw surface distress indicators—specifically

Second Spring 2026

Machine Learning CSC 7333

alligator cracking, rutting, potholes, and patching—to compute a standardized Pavement

Condition Index (PCI). This value, alongside International Roughness Index (IRI)
measurements, serves as the primary determinant for the following action bands:

 Class 1 (Defer Action): PCI > 95

 Class 2 (Pavement Preservation): 65 ≤ PCI ≤ 95

 Class 3 (Pavement Rehabilitation): 50 ≤ PCI < 65

 Class 4 (Pavement Reconstruction): PCI < 50

By evaluating the efficacy of diverse algorithms—including Multinomial Logistic Regression,

Decision Trees, and Random Forest Ensembles—this study establishes an optimized model stack
that identifies the XGBRegressor for PCI prediction and the XGBClassifier for final maintenance
categorization. This approach provides transportation authorities with a highly accurate,
automated tool to streamline infrastructure management and prioritize cost-effective
interventions across the roadway network.

PROJECT OBJECTIVES

The primary objective of this project is to develop an automated machine learning framework
that classifies asphalt pavement maintenance needs to replace labor-intensive manual
identification processes. By utilizing data from Louisiana's Pavement Management System
(PMS), the study aims to achieve the following specific goals:

1. Create a robust system to categorize roadway segments into four distinct action levels:
Defer Action, Pavement Preservation, Pavement Rehabilitation, and Pavement
Reconstruction based on refined PCI thresholds.
2. Develop a methodology to compute the Pavement Condition Index (PCI) by analyzing
weighted distress indicators, such as alligator cracking, rutting, potholes, and patching,
while integrating International Roughness Index (IRI) measurements.

Second Spring 2026

Machine Learning CSC 7333

3. Evaluate and select the most effective algorithms for specific tasks, identifying the
XGBRegressor for PCI prediction, the RandomForestRegressor for IRI assessment, and
the XGBClassifier for final maintenance categorization.
4. Utilize feature importance analysis within tree-based models to isolate specific distress
variables, such as alligator cracking, that serve as the primary determinants for major
structural rehabilitation.

SCOPE OF WORK

The scope of this project involves the development, evaluation, and implementation of a machine
learning framework designed to automate pavement maintenance decisions using Louisiana’s
infrastructure data. The work includes:

 Extracting pavement condition data from the Louisiana Pavement Management System
(PMS) database, followed by rigorous data filtering, cleaning, merging, and class
balancing to ensure dataset integrity.

 Computing the Pavement Condition Index (PCI) through weighted factor analysis of
distress indicators and establishing four distinct maintenance action bands: Defer Action,
Preservation, Rehabilitation, and Reconstruction.

 Building and training a variety of machine learning models, including multinomial

logistic regression, decision trees, random forest ensembles, and XGBoost, to predict
PCI, IRI, and final maintenance classifications.

 Conducting a comparative analysis of model efficacy using metrics such as RMSE,

MAE, R2, accuracy, and F1-macro scores to determine the optimal model stack for each
specific task.

 Performing feature importance analysis on tree-based models to identify the specific

surface distresses that most significantly influence the transition from routine
maintenance to major structural rehabilitation.

Second Spring 2026

Machine Learning CSC 7333

 Applying the selected models to individual roadway segments across the network to
provide visual analysis, such as average PCI by district and identification of the worst-
performing routes requiring reconstruction.

DATA COLLECTION

The reliability of any data-driven pavement-management framework depends directly on the quality,
coverage, and structure of the underlying dataset. This section describes the source of the data used in this
study, the temporal and spatial extent it covers, and the variables it contains.

 Source of the Data

The dataset analyzed in this project was obtained from the Louisiana Transportation Research Center
(LTRC), which maintains the pavement condition records of the Louisiana Department of
Transportation and Development (LA DOTD). The records were extracted through the IVision
software platform, the enterprise asset-management system used by LA DOTD to store, query, and
distribute pavement-condition survey results across the state highway network.

IVision consolidates the raw distress, roughness, and inventory data collected biennially during LA
DOTD’s automated pavement-condition surveys. These surveys are conducted using instrumented
vehicles equipped with high-resolution line-scan cameras, laser profilers, inertial sensors, and GPS
receivers, all of which enable continuous measurement of pavement surface condition along every
state-maintained route. The platform therefore serves as the official Pavement Management System
(PMS) of record for Louisiana and provides the authoritative data source used by state engineers for
network-level maintenance planning. For the purposes of this study, LTRC granted access to the most
recent PMS extract, which was exported from IVision in tabular form and consolidated into a single
spreadsheet file.

 Data Period and Area of Interest

The dataset corresponds to the 2025 pavement-condition survey cycle, as indicated by the “2025_”
prefix attached to every distress, roughness, and condition field in the file. Because the Louisiana
PMS operates on a biennial inspection schedule, the reported values reflect the most recent network-
wide survey campaign and represent the current operating condition of the state’s asphalt and
concrete pavements.

Geographically, the dataset covers the entire Louisiana state-maintained roadway network under the
jurisdiction of LA DOTD. The records span all nine administrative districts of the department—

Second Spring 2026

Machine Learning CSC 7333

Districts 02, 03, 04, 05, 07, 08, 58, 61, and 62—which together encompass the interstates, U.S.
routes, and Louisiana state routes maintained by the state. In total, the dataset contains 404,196
uniquely identified roadway segments distributed across 2,313 distinct routes, providing a
comprehensive, statewide snapshot of pavement performance that is well suited for training and
evaluating machine-learning models.

 Description and Structure of the Dataset

The raw dataset is organized as a flat tabular file in which each row corresponds to a single roadway
segment, uniquely defined by its route, its begin-chainage, and its end-chainage. A total of fifteen
(15) attributes are reported for every segment. These attributes fall into four logical groups:

 Segment geometry and identification : BeginChainage, EndChainage, IDSegment,

RSEG_ID.
 Administrative attributes: ROUTE_, RDISTRICT, DISTRICT_.
 Pavement type : VPAVETYPE.
 2025 pavement-condition measurements : 2025_ALCR, 2025_RNDM, 2025_RUFF,
2025_PTCH, 2025_RUT, 2025_AVG_IRI, 2025_AVG_PCI.
Each of the five 2025 distress indices (alligator cracking, random cracking, roughness, patching, and
rutting) is reported on a 0 – 100 scale, where a value of 100 represents the absence of the
corresponding distress and lower values represent increasing severity or extent. These indices
collectively serve as the inputs from which the composite Pavement Condition Index (PCI) is derived
through weighted-factor analysis.

 Preliminary Observations on Data Quality

A preliminary inspection of the raw file revealed two characteristics that informed the subsequent
data-preparation workflow:

 Sentinel values. A value of −1 appears throughout the distress, roughness, and IRI fields
to denote segments for which a valid measurement could not be obtained (for instance,
segments that were not surveyed, or segments whose pavement type is not applicable to a
given distress type). These sentinel entries must therefore be treated as missing values
rather than legitimate measurements.
 Missing PCI records. Approximately 85,417 records (about 21 percent of the dataset) do
not contain an average PCI value. These correspond primarily to segments whose
pavement type does not support PCI computation (e.g., brick, gravel, miscellaneous, or
composite surfaces) or whose distress inputs were incomplete for the 2025 cycle.

Second Spring 2026

Machine Learning CSC 7333

These observations motivated the filtering, cleaning, merging, and class-balancing procedures
described in the following sections of this report, which were necessary to produce a consistent
dataset suitable for training the proposed machine-learning models.

Second Spring 2026