Article

Proposal of an Integrated Method of Unmanned Aerial Vehicle

and Artificial Intelligence for Crack Detection, Classification, and
PCI Calculation of Airport Pavements
Valerio Perri 1,2 , Misagh Ketabdari 1 , Stefano Cimichella 2 , Maurizio Crispino 1 and Emanuele Toraldo 1, *

1 Department of Civil and Environmental Engineering, Politecnico di Milano, Piazza Leonardo da Vinci 32,
20133 Milan, Italy; [email protected] (V.P.); [email protected] (M.K.);
[email protected] (M.C.)
2 Logistics Command—Infrastructure Service, Air Force, Viale Dell’Università 4, 00185 Rome, Italy;
[email protected]
* Correspondence: [email protected]

Abstract: Assessing the condition of airport pavements is essential to ensure operational

safety and efficiency. This study presents an innovative, fully automated approach to
calculate the Pavement Condition Index (PCI) by combining UAV-based aerial photogram-
metry with advanced Artificial Intelligence (AI) techniques. The method follows three
key steps: first, analyzing orthophotos of individual pavement sections using a custom-
trained AI model designed for precise crack detection and classification; second, utilizing
skeletonization and semantic mask analysis to measure crack characteristics; and third,
automating the PCI calculation for faster and more consistent evaluations. By leveraging
high-resolution Unmanned Aerial Vehicle (UAV) imagery and advanced segmentation
models, this approach achieves superior accuracy in detecting transverse and longitudinal
cracks. The automated PCI calculation minimizes the need for human intervention, reduces
errors, and supports more efficient, data-driven decision-making for airport pavement
management. This study demonstrates the transformative potential of integrating UAV
and AI technologies to facilitate infrastructure maintenance and enhance safety protocols.
Academic Editors: Salvatore Leonardi
and Natalia Distefano
Keywords: pavement condition index (PCI); UAV-based aerial photogrammetry; airport
Received: 28 February 2025 pavement management; automated crack detection; AI-driven infrastructure analysis
Revised: 25 March 2025
Accepted: 27 March 2025
Published: 3 April 2025

Citation: Perri, V.; Ketabdari, M.;

1. Introduction
Cimichella, S.; Crispino, M.; Toraldo,
E. Proposal of an Integrated Method The integration of automation into pavement management systems, particularly in
of Unmanned Aerial Vehicle and airport infrastructure, marks a transformative step forward in civil engineering. Traditional
Artificial Intelligence for Crack methods for inspecting airport pavements, which rely heavily on manual labor and visual
Detection, Classification, and PCI assessments, are often inefficient, costly, and prone to human error. In contrast, modern
Calculation of Airport Pavements.
technologies such as Artificial Intelligence (AI) and Unmanned Aerial Vehicles (UAVs) offer
Sustainability 2025, 17, 3180. https://
more precise, efficient, and scalable solutions for pavement monitoring and assessment.
doi.org/10.3390/su17073180
Moreover, this approach contributes to sustainability by reducing the environmental im-
Copyright: © 2025 by the authors.
pact of traditional pavement evaluation methods by minimizing the need for on-ground
Licensee MDPI, Basel, Switzerland.
This article is an open access article
equipment and reducing carbon emissions associated with.
distributed under the terms and This research introduces a novel, fully automated system for calculating the Pavement
conditions of the Creative Commons Condition Index (PCI) of airport pavements by utilizing computer vision and machine
Attribution (CC BY) license learning techniques. Specifically, a custom-trained YOLO11 segmentation model developed
(https://creativecommons.org/ by the authors is adopted to detect and classify various types of pavement cracks on
licenses/by/4.0/).

Sustainability 2025, 17, 3180 https://doi.org/10.3390/su17073180

Sustainability 2025, 17, 3180 2 of 19

runways, aprons, and taxiways from orthophotos. The system is further enhanced by
deploying lightweight drones equipped with high-resolution cameras, capable of surveying
large areas quickly and safely, significantly reducing the time and costs typically associated
with pavement inspections.
The objective of this end-to-end methodology is to improve the accuracy and consis-
tency of pavement distress detection and classification, eliminating the need for human
intervention. Traditional inspection methods often demand extensive manual effort and
are prone to variability and errors due to subjective judgment.
By integrating UAVs with the AI segmentation model, this system revolutionizes
pavement inspections with a highly efficient, accurate, and automated workflow. UAVs
enable rapid data collection over large areas without long interference with airport opera-
tions, ensuring high-resolution imagery and uniform quality. This allows for the detection
of even the smallest defects, such as fine cracks, with exceptional precision. Geotagging
capabilities further enhance the system by accurately localizing defects and mapping main-
tenance areas with precision. This synergy not only improves the accuracy of pavement
condition evaluations but also accelerates maintenance planning, ultimately reducing costs
and minimizing downtime for airport operations.

2. Literature Review
Traditional methods for pavement inspection, relying on manual visual assessments,
are time-consuming, inconsistent, and prone to human error due to subjective evaluations.
Recognizing these limitations, significant advancements have been made in recent decades
to develop automated systems for pavement condition evaluation [1–35].
For example, Chen et al. [2] introduced the Shanghai Airport Pavement Management
System (SHAPMS), a GIS- and GPS-based platform designed for geospatial analysis and
maintenance planning at major airports. While SHAPMS optimizes data collection and
maintenance strategies, its implementation is hindered by the labor-intensive setup of
spatial databases and reliance on advanced hardware. Similarly, Lima et al. [3] proposed a
methodology integrating GPS, laser scanning, and video imaging for pavement inspections,
demonstrating faster and more consistent results than traditional methods. However,
challenges like limited image resolution and difficulties in detecting specific distress types
remain unresolved.
Innovative approaches like LiDAR-based Mobile Mapping Systems—MMS (Habib
et al. [4]) and IoT-integrated construction monitoring systems (Kong et al. [5]) have also
advanced the field, offering precise measurements and real-time monitoring capabilities.
Yet, these systems often rely on costly equipment and skilled personnel, limiting their
applicability in resource-constrained environments. Likewise, Kovačič et al. [6] presented a
sustainable runway management model for regional airports, utilizing geodetic measure-
ments to monitor deformations. While effective, the reliance on high-precision tools poses
challenges for broader adoption.
Recent studies have explored AI and UAV integration to address these limitations.
Pietersen et al. [8] demonstrated a drone-based imaging system using Convolutional Neural
Networks (CNNs) for crack detection and PCI calculations, achieving high accuracy but
requiring diverse datasets to enhance generalizability. Similarly, Zhang et al. [10] devel-
oped the AM-Mask R-CNN algorithm, incorporating attention mechanisms for improved
segmentation under challenging conditions. Despite achieving superior performance, the
system demands high computational resources and controlled conditions.
Other efforts include Shon et al. [11], who introduced an Autonomous Condition
Monitoring-Based Pavement Management System (ACM-PMS), utilizing connected vehi-
cles to minimize operational costs, and Malekloo et al. [12], who developed a cost-efficient
Sustainability 2025, 17, x FOR PEER REVIEW 3 of 19

Sustainability 2025, 17, 3180 3 of 19

vehicles to minimize operational costs, and Malekloo et al. [12], who developed a cost-
efficient
AI-enabledAI-enabled system
system using using dashcam
dashcam imageryimagery for distress
for distress classification.
classification. These stud-
These studies high-
ies highlight
light the potential
the potential of automation
of automation to revolutionize
to revolutionize pavementpavement
managementmanagement butface
but often of-
ten face challenges
challenges related
related to to diversity,
dataset dataset diversity, environmental
environmental conditions,conditions, and scalability.
and scalability.
To address the gaps in these prior studies, this research proposes an accessible and
computationally efficientsolution.
computationally efficient solution.AAYOLO11
YOLO11 Custom
Custom Model
Model is deployed
is deployed on aon a widely
widely used
used free platform
free platform toorthophotos
to infer infer orthophotos of individual
of individual pavementpavement
sections.sections.
The modelThedetects
model and
de-
tects and visualizes
visualizes longitudinallongitudinal and transverse
and transverse crackssemantic
cracks through through masks,
semantic masks, generat-
generating reports
ing
withreports with
detailed detailed information
information on distress
on distress types types and quantification.
and quantification. These data areThese
thendata
usedare
to
then usedthe
calculate to calculate
Pavement theCondition
PavementIndex
Condition
(PCI) Index (PCI) by parameters
by evaluating evaluating parameters such
such as distress
as distress
severity, severity,
density, anddensity,
deductand deduct
values. Thisvalues. Thissimplifies
approach approachimplementation,
simplifies implementation,
minimizes
minimizes
computationalcomputational
requirements,requirements,
and makesand makes advanced
advanced pavement pavement
evaluationevaluation
accessibleac-
to
cessible to non-professional
non-professional users, overcoming
users, overcoming key limitations
key limitations of existing of existing systems.
systems.

3. Methodology
This study
This study aims
aimstotodevelop
developan anintegrated
integratedAIAIand
andUAV
UAVsystem
systemforfor quickly
quickly and
and au-
auto-
tomatically assessing airport pavement conditions, and the key steps of the proposed
matically assessing airport pavement conditions, and the key steps of the proposed meth-
methodology
odology are presented
are presented in theinfollowing
the following flowchart
flowchart (see Figure
(see Figure 1). 1).

Figure 1. The flowchart

flowchart of the proposed integrated methodology.
methodology.

As illustrated in Figure 1, the methodology starts with the development of a custom

AI model capable of detecting and segmenting different different types of pavement distresses. To To
build this model, a large dataset of high-resolution images was collected from various
build
airport infrastructures using
using aalightweight
lightweightUAV.UAV.These
Theseimages, capturing
images, capturing various types
various of
types
distresses,
of form
distresses, thethe
form foundation
foundation forfor
training thethe
training AI AI
algorithm.
algorithm.
Then, each image
image was manually labeled by a qualifiedengineer
was manually labeled by a qualified engineer toto
ensure accuracy
ensure in
accuracy
identifying
in andand
identifying classifying pavement
classifying pavement defects. This This
defects. labeling step isstep
labeling crucial for thefor
is crucial AI the
model
AI
to learn how to detect and classify different types of distresses effectively.
model to learn how to detect and classify different types of distresses effectively.
Sustainability 2025, 17, 3180 4 of 19

To optimize the dataset for training and reduce computational demands, a tiling
preprocessing technique was applied. This ensured that the images were divided into
smaller sections suitable for the AI model’s specifications without losing important details
or missing any distresses during resizing.
Once the dataset was prepared, the AI model was trained on the labeled images.
Following the training phase, the model underwent validation to evaluate its performance
and quantify any errors in detecting and segmenting pavement distresses.
This systematic approach ensures the development of a reliable and efficient AI/UAV
system for automated airport pavement management. Once this custom AI model is
ready, it can be utilized as a general tool to process every type of asphalt pavement
through images.
Future scope of this study is to demonstrate the application of the AI custom model to
assess the Pavement Condition Index (PCI) of an adopted case study (i.e., an Italian airport
taxiway) using advanced photogrammetric techniques and orthophotos.
Instead of relying on standard pavement images, georeferenced orthophotos were
used due to their metric properties, which are essential for accurately evaluating distresses.
A lightweight UAV was deployed to conduct a photogrammetric survey of the selected
taxiway. Approximately 900 high-resolution images were captured, along with geotagging
of 45 targets. These data were used to create a digital model of the taxiway, divided into
orthophotos with a resolution of 6666 × 6666 pixels, representing 20 × 20-m sections of
the pavement.
The custom YOLO11 model was applied to each orthophoto for detecting and seg-
menting linear distresses, such as cracks. The semantic masks generated during this process
were used to measure the total length and average width of the cracks, key metrics for
PCI calculation. Using these metrics, the severity level, density, and deduct value for each
distress were determined, enabling the calculation of the PCI for each pavement section.
Finally, the overall PCI for the entire taxiway was obtained as the average of the PCI values
from all sections.
Furthermore, to validate the proposed integrated methodology, a visual inspection
was conducted on 10% of randomly selected pavement sections (simple units). The results
from this manual inspection were then compared with those obtained from the automated
method to assess the accuracy and reliability of the new approach.
This methodology highlights the effectiveness of integrating UAVs and AI for au-
tomated pavement management, offering a precise, efficient, and scalable solution for
evaluating pavement conditions.

3.1. Development of Custom AI Model

3.1.1. UAV Image Acquisition
The first step in implementing the proposed methodology is image acquisition, a
crucial process for creating a reliable dataset that captures a wide variety of potential
distresses present in airport pavements. This survey was conducted at Italian airports,
where hundreds of high-resolution images were captured using a lightweight UAV flying
at an altitude of approximately 14 m. This specific flight altitude was selected to ensure the
detection of thin cracks, with each image achieving a resolution of 5472 × 3648 pixels.
A lightweight (approx. 250 g) and compact UAV was used for the aerial photogram-
metry survey, ensuring efficient and precise data collection. With GPS-assisted flight
capabilities, the UAV enabled stable and autonomous waypoint navigation, facilitating
consistent and repeatable image acquisition.
Sustainability 2025, 2025,
Sustainability 17, x 17,
FOR PEER REVIEW
3180 5 of 19
5 of 19

3.1.2. Datasetthe
Creating Preparation and AI
dataset was Algorithm
a crucial stepTraining
in training the AI models. To facilitate the
Creating the dataset was a crucial
process of labeling and preparation, a new projectstep in training
wastheinitiated
AI models.
on aTowidely
facilitate the com-
used
process
puter of labeling
vision platformand(i.e.,
preparation, a newThe
Roboflow). project
rawwas initiated
images on auploaded,
were widely usedandcomputer
each image
vision platform (i.e., Roboflow). The raw images were uploaded, and each image underwent
underwent detailed manual labeling. This involved annotating every image at the pixel
detailed manual labeling. This involved annotating every image at the pixel level to
level to precisely identify and classify pavement distresses.
precisely identify and classify pavement distresses.
Focus was given to two primary types of cracks: longitudinal and transverse (see
Focus was given to two primary types of cracks: longitudinal and transverse (see
Figure
Figure2).2).The
Thelabeling
labeling process wasconducted
process was conducted withwith exceptional
exceptional a ention
attention to ensuring
to detail, detail, ensur-
ingeven
even the
the smallest
smallest cracks
cracks werewere accurately
accurately identified
identified and annotated.
and annotated. This levelThis level ofispreci-
of precision
sion is essential
essential for creating
for creating a high-quality
a high-quality datasetofcapable
dataset capable of training
effectively effectively
the training
AI modelsthe AI
models for accurate
for accurate crack detection
crack detection and classification.
and classification.

Figure 2. 2.
Figure Example
Exampleofofaccurate
accurate image annotation
image annotation toto identify/classify
identify/classify longitudinal/transverse
longitudinal/transverse cracks.
cracks.

RedRed semanticmasks
semantic masks in in Figure
Figure 22represent
represent longitudinal
longitudinal cracks,
cracks,and and
the purple ones ones
the purple
represent the transverse cracks.
represent the transverse cracks.
Once labeling was complete, the dataset was formatted according to the YOLO stan-
Once labeling was complete, the dataset was forma ed according to the YOLO stand-
dard, widely used for object detection and instance segmentation tasks. This format
ard, widely used for object detection and instance segmentation tasks. This format re-
requires each image to be accompanied by a corresponding annotation file in .txt format.
quires each imagefiletospecifies
The annotation be accompanied by a and
the object class corresponding
the boundingannotation file infor.txt
box coordinates format.
each
Theobject in the image, normalized to the image dimensions. Each line in the file represents each
annotation file specifies the object class and the bounding box coordinates for
object in theand
an object, image, normalized
the data to the
is structured asimage
follows:dimensions.
<object-class>,Each line in the<y-center>,
<x-center>, file represents
an<width>,
object, and the data
<height>, where is the
structured
coordinatesasand
follows: <object-class>,
dimensions are relative<x-center>,
to the image <y-center>,
size.
<width>, A preprocessing
<height>, where technique was employed
the coordinates to artificiallyare
and dimensions increase
relative theto
dataset by ansize.
the image
order of magnitude. Specifically, each photogram was tiled to train the
A preprocessing technique was employed to artificially increase the dataset by an algorithm without
resizing, thereby preserving
order of magnitude. Specifically, veryeach
thinphotogram
cracks. Thiswasmethodology
tiled to train allowed training the
the algorithm without
algorithm on an artificial big dataset without spending many hours to manually label
resizing, thereby preserving very thin cracks. This methodology allowed training the al-
thousands of pictures.
gorithm on an artificial big dataset without spending many hours to manually label thou-
The dataset was divided into training, validation, and testing subsets to ensure proper
sands of pictures.during the model training process; 70% of the images were allocated for
generalization
The dataset
training, 20% forwas divided
validation, andinto
10% training,
for testing.validation,
This division and testing
ensured subsets
effective to ensure
training,
proper generalization
providing during
sufficient data the model
to validate training
and test process;
the model’s 70% of the images were allo-
performance.
cated for training, 20% for validation, and 10% for testing. This division ensured effective
training, providing sufficient data to validate and test the model’s performance.
The combination of high-resolution UAV imagery, precise manual labeling, and care-
ful dataset forma ing was essential for developing a robust and accurate model capable
of autonomously identifying pavement distresses. At the end of the dataset configuration,
 Sustainability 2025, 17, 3180 6 of 19
Sustainability 2025, 17, x FOR PEER REVIEW 6 of 19

The combination of high-resolution UAV imagery, precise manual labeling, and careful
Sustainability 2025, 17, x FOR PEER REVIEW
dataset formatting was essential for developing a robust and accurate model capable 6 ofof
19
3.1.3. Validation of Custom AI Model
autonomously identifying pavement distresses. At the end of the dataset configuration, the
To validate
training process the
was model’s
performed. performance, the results from the custom-trained model
were3.1.3. Validation
compared of Custom
with on-siteAI Model
visual inspections (see Figure 3). This comparison ensured
3.1.3. Validation of Custom AI Model
that theTo validate
model’s the model’s
crack detection performance,
accuracy couldthe results from the
be assessed custom-trained
against real-worldmodel
observa-
were Tocompared
validate the model’s
with performance,
on-site visual the results
inspections (see from the3).
Figure custom-trained
This model
comparison were
ensured
tions.
compared with on-site
that the model’s crackvisual inspections
detection (see
accuracy Figure
could 3). This comparison
be assessed ensured observa-
against real-world that the
model’s
tions. crack detection accuracy could be assessed against real-world observations.

Figure 3. Validation process; on-site cracks classification and quantification.

Figure3.3.Validation
Figure Validationprocess;
process;on-site
on-sitecracks
cracksclassification
classificationand
andquantification.
quantification.

The
Thevalidation ofofthe AI model
model proposedinin this study to detect cracks carried
was carried
The validation
validation of the
the AI
AI model proposed
proposed in this
thisstudy
studytotodetect
detectcracks
crackswas was carried
outout
byby
out comparing
by comparingthe
comparing thelengths
the lengths of
lengths of detected
ofdetected longitudinal
detectedlongitudinal
longitudinal andand
and transverse
transverse
transverse cracks
crackscracks
withwith
with their their
their
corresponding
corresponding
corresponding on-site
on-sitemeasurements,
on-site measurements, followedby
measurements, followed
followed by
by calculating
calculating
calculating thethe
the error.
error.
error. For For illustrative
Forillustrative
illustrative
purposes,
purposes, the
purposes, thevalidation
the validationresults
validation results of
resultsof two
oftwo random
tworandom
random sample
sample
sample units
units
units(see
(see(see Figure
Figure
Figure 4)4)are4)presented.
are are presented.
presented.

(a) (b)

(a) 4. Examples of cracks detected by the proposed AI(b)

Figure model (a) Annotated sample unit n°2; (b)
Annotated sample unit n°10.
Figure
Figure 4. Examples
4. Examples ofof cracksdetected
cracks detectedby
bythe
theproposed
proposed AI
AI model
model(a)
(a)Annotated
Annotatedsample unit
sample n◦ 2;
unit n°2; (b)
(b) Annotated sample unit ◦
n 10.
Annotated
Thesample
lengthunit n°10.
confrontation of measured on-site and AI-driven detected cracks clarifies
that The
the proposed model detected most cracks
length confrontation of measured accurately
on-site on bothdetected
and AI-driven orthophotos,
cracksasclarifies
proved
The length
in Table
that 1. confrontation
the proposed
of measured on-site and AI-driven detected cracks
model detected most cracks accurately on both orthophotos, as proved
clarifies
that
inthe proposed
Table 1. model detected most cracks accurately on both orthophotos, as proved
in Table 1.
 Sustainability 2025, 17, x FOR PEER REVIEW 7 of 19
Sustainability 2025, 17, 3180 7 of 19

Table1.1.Length
Table Lengthconfrontation
confrontationofofmeasured
measuredon-site
on-siteand
andAI-driven
AI-drivendetected
detectedcracks.
cracks.

Ortho- Cracks AI Detected

AI Detected On-Site
On-SiteMeasured
Measured Undetected
Undetected Error
Error
Orthophoto Cracks
photo Direction
Direction (m)(m) (m)(m) (m)
(m) (%)
(%)
Longitudinal 113.34 116.22 2.88 2.48
n°2 ◦ Longitudinal 113.34 116.22 2.88 2.48
n 2 Transverse
Transverse 58.91
58.91 62.04
62.04 3.13
3.13 5.04
5.04
Longitudinal
Longitudinal 120.34
120.34 123.56
123.56 3.22
3.22 2.61
2.61
n°10
n◦ 10 Transverse
Transverse 54.23
54.23 58.40
58.40 7.14
7.14 2.87
2.87

Overall, these results highlight the high accuracy of the custom-trained model in de-
Overall, these results highlight the high accuracy of the custom-trained model in
tecting both longitudinal and transverse cracks, with minimal missed detections that align
detecting both longitudinal and transverse cracks, with minimal missed detections that
closely with on-site observations.
align closely with on-site observations.

3.2.Pavement
3.2. PavementCondition
ConditionAssessment
Assessment
3.2.1.Image
3.2.1. ImageAcquisition
Acquisition
Thefirst
The firststep
stepininassessing
assessingpavement
pavementcondition
conditionusing
usingthe theproposed
proposedmethodology
methodologyisis
image acquisition. For this study, the survey was carried out on
image acquisition. For this study, the survey was carried out on an 800-m-long and an 800-m-long and 15-m-
15-m-
wideasphalt
wide asphalttaxiway
taxiway at an
at an Italian
Italian airport.
airport. Approximately
Approximately 900 high-resolution
900 high-resolution imagesimages
were
were captured
captured from an from an altitude
altitude of 14
of 14 m. m. Each
Each image,
image, with with a resolution
a resolution of 5472 × 3648
of 5472 × 3648 pix-
pixels,
els, provided
provided detailed
detailed and views
and clear clear views
of the of the pavement,
pavement, capturingcapturing
variousvarious
types oftypes of dis-
distresses,
tresses, including
including longitudinal
longitudinal and transverse
and transverse cracks. cracks.
Thehigh-resolution
The high-resolutioncameracameraof ofthe
theadopted
adopted UAVUAVcaptured
captureddetailed
detailedphotographs,
photographs, es- es-
sentialfor
sential forgenerating
generating accurate
accurate orthomosaics,
orthomosaics, which
which serve
serve as theas foundation
the foundation for subse-
for subsequent
pavement condition
quent pavement analyses.analyses.
condition Additionally, the UAV’s
Additionally, extended
the flight timeflight
UAV’s extended of uptime
to 30of
min
up
per battery
to 30 min perallowed
ba eryextensive
allowed area coverage
extensive without
area frequent
coverage withoutinterruptions, making it a
frequent interruptions,
practical
making it and cost-effective
a practical solution for airport
and cost-effective pavement
solution for airportinspections.
pavement inspections.
To
Tofurther
furtherenhance
enhanceaccuracy,
accuracy,target
targetpatches
patcheswere
werestrategically
strategically placed
placed on
on the
the pave-
pave-
ment
ment(see
(seeFigure
Figure5a)
5a)and
andgeoreferenced
georeferencedusing usingGPS
GPS(see
(seeFigure
Figure5b). 5b).These
Thesetargets
targetsensured
ensured
uniform
uniform image
image scaling
scaling and
andaccurate
accuratespatial
spatialreferencing,
referencing, facilitating
facilitating precise
precise analysis
analysis of
of
pavement
pavementconditions.
conditions.

(a) (b)

Figure5.5.(a)
Figure (a)Target
Targetpatches
patchespositioning;
positioning;(b)
(b)Target
Targetpatches
patcheslocalized
localizedbybya aGPS
GPSstation.
station.

White
White and
and red targets, measuring
red targets, measuring30 30×× 3030
cm,cm,
werewere positioned
positioned every
every 40 m40along
m along
both
both sides of the taxiway. GPS data was recorded for each high-resolution image,
sides of the taxiway. GPS data was recorded for each high-resolution image, enabling pre- en-
abling precise localization and mapping of cracks and other pavement distresses.
cise localization and mapping of cracks and other pavement distresses. This approach en- This
approach
sured thatensured
detectedthat detected
issues issues were
were accurately accurately
referenced referenced
within within theThe
the orthomosaics. orthomo-
UAV’s
saics.
efficiency and ability to cover large areas quickly enabled the survey to be completedto
The UAV’s efficiency and ability to cover large areas quickly enabled the survey in
Sustainability 2025, 17, x FOR PEER REVIEW 8 of 19
Sustainability 2025, 17, 3180 8 of 19

justcompleted
be a few hours, underscoring
in just its reliability
a few hours, underscoringand its
practicality
reliabilityfor
andairport pavement
practicality inspec-
for airport
tions.
pavement inspections.

3.2.2.
3.2.2. Orthophoto
Orthophoto Creation
Creation
The
The aerial
aerial images
images captured
captured during
during the survey were processed using advanced pho-
togrammetry
togrammetry software (i.e., Metashape,
Metashape, developed by Agisoft LLC, aa company company based
based in
in
St.
St. Petersburg,
Petersburg,Russia)
Russia)totogenerate
generate high-quality
high-quality georeferenced
georeferencedorthomosaics.
orthomosaics.ThisThis
process-
pro-
ing involved
cessing converting
involved rawraw
converting image datadata
image intointo
accurate 3D 3D
accurate surface
surfacereconstructions
reconstructionsusing
us-
Structure-from-Motion (SfM) algorithms. These algorithms enabled the creation
ing Structure-from-Motion (SfM) algorithms. These algorithms enabled the creation of de- of detailed
and
tailedprecise models,
and precise ensuring
models, a reliable
ensuring representation
a reliable of the pavement’s
representation condition.
of the pavement’s condition.
The software’s functionality included generating dense point clouds,
The software’s functionality included generating dense point clouds, creating creating detailed
de-
meshes (see Figure
tailed meshes 6a), and
(see Figure 6a), applying
and applyingtexture mapping—steps
texture mapping—steps essential for for
essential producing
produc-
orthophotos
ing orthophotosof exceptional quality.
of exceptional quality.

(a)

(b)

Figure 6. (a)
Figure (a)Example
Exampleofofdetailed
detailedmeshes processed
meshes by by
processed advanced
advancedphotogrammetry software;
photogrammetry (b)
software;
Example
(b) of geotagging
Example processed
of geotagging processedby by
advanced photogrammetry
advanced photogrammetrysoftware.
software.

Geotagging
Geotagging was was integrated
integrated into
into the
the process,
process, with
with each
each geotag
geotag positioned
positioned every
every 40
40 mm
along the taxiway, ensuring accurate spatial referencing
along the taxiway, ensuring accurate spatial referencing for all identified
identified defects, as illus-
trated
trated in
in Figure
Figure 6b.
6b. According to the figure,
figure, each
each blue
blue flag
flag represents
represents aa geotag
geotag positioned
positioned
every 40 m along the taxiway.
every 40 m along the taxiway.
The
The resulting
resulting orthophotos,
orthophotos, with
with aa resolution
resolution of 6666 ×
of 6666 × 6666 pixels, represented real-
world
world areas
areasof 20×
of20 × 20 m (see Figure 7). These high-resolution images provided a detailed detailed
and
and scalable
scalable representation
representationofofthe
thetaxiway
taxiwaysurface,
surface,facilitating
facilitatingthethedetection
detectionand classifica-
and classifi-
tion of pavement cracks with precision. For this study, based on the length
cation of pavement cracks with precision. For this study, based on the length and width and width of
the understudy taxiway, 40 orthophotos were captured.
of the understudy taxiway, 40 orthophotos were captured.
 Sustainability2025,
Sustainability 2025,17,
17,3180
x FOR PEER REVIEW 99 of 19
of 19

Figure7.
Figure 7. Examples
Examples of
of captured
captured orthophotos.
orthophotos.

3.2.3.
3.2.3. Local
Local Inference
Inference on on Taxiway
TaxiwayOrthophotos
Orthophotos
In
In this study, due to the computationaldemands
this study, due to the computational demandsof of processing
processingsuchsuch large
large images
images and
and
the inherent efficiency of YOLO models with smaller image patches,
the inherent efficiency of YOLO models with smaller image patches, a tiling approach was a tiling approach
was adopted.
adopted. This This method
method ensured
ensured compatibility
compatibility with with the model
the model whilewhile preserving
preserving the
the high
high resolution
resolution required
required for precise
for precise crackcrack detection
detection andand annotation.
annotation. In this
In this regard,
regard, eacheach or-
ortho-
thophoto was divided into smaller tiles (as a computational process),
photo was divided into smaller tiles (as a computational process), each with a maximum each with a maximum
dimension
dimension of 1280 ×
of 1280 1280 pixels.
× 1280 pixels.
Subsequently
Subsequently to the tilingprocess,
to the tiling process,a aslight
slightoverlap
overlap between
between adjacent tiles
adjacent is applied
tiles to
is applied
maintain
to maintain continuity
continuityduring reconstruction
during reconstruction andand
to minimize
to minimizeinformation
informationlossloss
or misalign-
or misa-
ment
lignment at the edges. The image was segmented into smaller sections based onsystem,
at the edges. The image was segmented into smaller sections based on a grid a grid
using predefined coordinates to crop the original orthophoto into tiles suitable
system, using predefined coordinates to crop the original orthophoto into tiles suitable for for analysis.
Once divided, each tile was processed by the YOLO11 Custom Model, which employs
analysis.
segmentation
Once divided,algorithms trained
each tile was on manually
processed byannotated
the YOLO11 datasets
Custom to detect
Model,and classify
which em-
pavement cracks.
ploys segmentation algorithms trained on manually annotated datasets to detect and clas-
To ensure cracks.
sify pavement reliable results, a confidence threshold was applied during the detection
process. This threshold, which can
To ensure reliable results, range from
a confidence 0% (where
threshold wasthe modelduring
applied considersthe excessive
detection
details, potentially leading to over-detection of cracks) to 100% (where
process. This threshold, which can range from 0% (where the model considers excessive the model only
identifies details it isleading
details, potentially fully confident about, potentially
to over-detection of cracks) resulting
to 100%in(where
under-detection),
the model onlywas
carefully calibrated. For this study, a confidence threshold of 50% was chosen,
identifies details it is fully confident about, potentially resulting in under-detection), was striking a
balance
carefullybycalibrated.
allowing For onlythis
highly certain
study, predictions
a confidence to be annotated.
threshold of 50% wasThis reduced
chosen, falsea
striking
positives and improved the accuracy of defect identification.
balance by allowing only highly certain predictions to be annotated. This reduced false
After processing, the annotated tiles were reassembled to reconstruct the original
positives and improved the accuracy of defect identification.
orthophoto, now fully annotated with detected cracks. The reassembly involved aligning
After processing, the annotated tiles were reassembled to reconstruct the original or-
the tiles based on their original grid coordinates, with the overlapping sections ensuring
thophoto, now fully annotated with detected cracks. The reassembly involved aligning
seamless continuity of cracks that spanned multiple tiles.
the tiles based on their original grid coordinates, with the overlapping sections ensuring
After processing, the annotated tiles were reassembled to reconstruct the original
seamless continuity of cracks that spanned multiple tiles.
orthophoto (see Figure 8), now fully annotated with detected cracks. The reassembly
After processing, the annotated tiles were reassembled to reconstruct the original or-
involved aligning the tiles based on their original grid coordinates, with the overlapping
thophoto (see Figure 8), now fully annotated with detected cracks. The reassembly in-
sections ensuring seamless continuity of cracks that spanned multiple tiles.
volved aligning the tiles based on their original grid coordinates, with the overlapping
As demonstrated, to facilitate interpretation, different colors were used to represent
sections ensuring seamless continuity of cracks that spanned multiple tiles.
each crack type—red for longitudinal cracks and purple for transverse cracks. In fact,
longitudinal cracks were aligned parallel to the taxiway’s direction, while transverse
cracks intersected perpendicularly. The distinct color-coded segmentation masks effectively
highlighted these patterns, demonstrating the model’s capability to accurately and reliably
identify pavement distresses.
Sustainability 2025,
Sustainability 17,17,
2025, 3180
x FOR PEER REVIEW 10 of 19 10 of 19

Figure Examplesofof
8. Examples
Figure 8. final
final annotated
annotated orthophoto
orthophoto byproposed
by the the proposed AI model.
AI model.

4. PCI CalculationtoMethod
As demonstrated, facilitate interpretation, different colors were used to represent
each crack type—red for longitudinal cracks and purple for transverse cracks. In fact, lon-
The calculation of the Pavement Condition Index (PCI) is a vital component in evaluat-
gitudinal cracks were aligned parallel to the taxiway’s direction, while transverse cracks
ing the condition of airport pavements. This study introduces an automated methodology
intersected perpendicularly. The distinct color-coded segmentation masks effectively
that analyzes annotated orthophotos to quantify crack characteristics, such as length, width,
highlighted these pa erns, demonstrating the model’s capability to accurately and relia-
density, andpavement
bly identify severity.distresses.
The model is then used to compute the PCI following the ASTM D5340 standard [36],
4. PCIprovides
which Calculation Method
a systematic framework for assessing pavement conditions. This approach
ensures that inspections are consistent
The calculation of the Pavement Condition and reliable, enabling
Index (PCI) airport
is a vital authorities
component to compare
in eval-
pavement conditionofdata
uating the condition across
airport variousThis
pavements. timestudy
periods and locations.
introduces an automated method-
ologyThe
thatASTM
analyzesD5340 outlines
annotated a detailedto
orthophotos process forcrack
quantify identifying and assessing
characteristics, such asdifferent
length, width, density, and severity.
types of pavement distresses, including cracking. In accordance with standard procedures,
The step
the first modelisistothen usedthe
divide to compute the PCI following
taxiway—classified as athe ASTMof
section D5340 standard [36], smaller
the airport—into
which provides
sample a systematic
units. According framework for [36],
to specifications assessing pavement conditions.
it is recommended Thisa section
to divide ap- into
proach ensures that inspections are2 consistent and 2 reliable, enabling airport authorities to
sample units of between 185 m and 454 m for asphalt-surfaced airfields. For this study,
compare pavement condition data across various time periods and locations.
the taxiway was divided into 40 sample units, each covering 400 m2 .
The ASTM D5340 outlines a detailed process for identifying and assessing different
For each distress detected within a sample unit, its metrics must be quantified using
types of pavement distresses, including cracking. In accordance with standard proce-
appropriate
dures, the first measurement
step is to divideunits. In this study, longitudinal
the taxiway—classified as a sectionand transverse
of the cracks were
airport—into
quantified
smaller sample units. According to specifications [36], it is recommended to divide a sec- of each
by their length, as specified by the standards. By dividing the total length
distress
tion intotype
samplebyunits
the area of the sample
of between unit,
185 m2 and the
454 m2density level of the distress
for asphalt-surfaced airfields.was determined,
For this
study,
as the taxiway in
demonstrated was divided into
Equations (1) 40
andsample
(2). units, each covering 400 m2.
For each distress detected within a sample unit, its metrics must be quantified using
Total lenght
appropriate measurement units. In this study, o f longitudianl
longitudinal cracks
and transverse [m] were
cracks
Density longitudinal cracks [%] = × 100
2 ] total length of
(1)
Sample By
quantified by their length, as specified by the standards. unitdividing
Area [mthe
each distress type by the area of the sample unit, the density level of the distress was
determined, as demonstrated in Equations Total lenght o f transverse cracks[m]
Density [%] = (1) and (2).
tranverse cracks × 100 (2)
ℎ
Sample unit Area [m[2 ] ]
[%] = × 100 (1)
[ ]level of each crack, which can
Another parameter to calculate the PCI is the severity
be categorized as low, medium, or high. The standard [36] defines the severity for linear
ℎ [ ]
[%] =pavement based on their average width
cracks in asphalt concrete airfield × 100as follows.
(2)
[ ]
• Another
Low severity withtoaverage
parameter calculatecrack width
the PCI less
is the than 6.4
severity mm;
level of each crack, which can
•be categorized
Medium severity with average crack width between
as low, medium, or high. The standard [36] defines 6.4 mm and 25.4
the severity formm;
linear
•cracks
High severity
in asphalt with airfield
concrete averagepavement
crack width
basedgreater
on theirthan 25.4width
average mm. as follows.
 Low
Onceseverity with average
the density of each crack width
distress andless
itsthan 6.4 mm;
corresponding severity level are determined,
the associated Deduct Value (DV) for each crack type can be calculated. This is performed
according to the Figure B-33. Longitudinal/Transverse Cracking, proposed by M. Y. Shahin [37]
for asphalt concrete airfield pavement, which provides a standardized approach to quanti-
fying the impact of each distress on pavement condition.
Sustainability 2025, 17, 3180 11 of 19

To automate the process, a mathematical expression was derived for each severity
level curve by analyzing the trend of each curve, as presented in Equations (3)–(5).

DV H = 0.006352 × DD3 − 0.40525 × DD2 + 9.04947 × DD + 8.19191 (3)

DV M = 0.003702 × DD3 − 0.2258 × DD2 + 4.8119 × DD + 3.1753 (4)

DV L = 0.002362 × DD3 − 0.13465 × DD2 + 2.8948 × DD − 0.3497 (5)

In these equations, DVH represents the deduct value for cracks with high severity;
DVM represents the deduct value for cracks with medium severity; DVL represents the
deduct value for cracks with low severity; DD represents the distress density expressed
in percentages.
Once the deduct value for each sample unit is determined, the pavement condition
index can be calculated by subtracting the DV from 100, which represents the ideal condition
of a pavement with no distresses (see Equation (6)).

PCI i = 100 − DV i (i = 1, . . . , n) (6)

In this equation, i represents a random sample unit; n represents total number of

sample units.
Additionally, the overall PCI for the entire section (e.g., the entire taxiway) is calculated
as the average of the PCI values from all the sample units. The final score dedicated to the
PCI ranges from 0 (fully deteriorated pavement) to 100 (pavement in excellent condition).
This score serves as a critical metric for prioritizing maintenance and rehabilitation activities.
For instance, a low PCI score indicates an urgent need for intervention, while a high PCI
score suggests the pavement is in good condition.
The adoption of ASTM D5340 provides stakeholders with an objective and standard-
ized method for evaluating pavement health. This minimizes subjectivity and ensures that
different personnel or organizations using the same protocol can derive comparable results.
In the proposed methodology, annotated orthophotos with color-coded segmentation
masks (see Figure 8) are processed to extract specific features of longitudinal (red) and
transverse (purple) cracks. These crack masks are converted to grayscale and skeletonized
to create one-pixel-wide representations of the cracks using an open-source Python library,
such as Skimage (scikit-image 0.21.0, released in 2023), see Figure 9, which is necessary
for17,measuring
Sustainability 2025, the average crack width and length with high precision, providing detailed12
x FOR PEER REVIEW of 19
input for PCI computation.

(a) (b)

Figure 9. Examples of (a)

Figure Longitudinal
9. Examples of (a)cracks skeleton;
Longitudinal (b)skeleton;
cracks Transverse cracks skeleton.
(b) Transverse cracks skeleton.

The skeletonization process enhances precision by simplifying the geometry of the

cracks while preserving their structural integrity. This step, which calculates the length of
the cracks, reduces the complexity of the data, making it easier to analyze the crack fea-
tures accurately.
Using a distance transform algorithm, the crack metrics are calculated to determine
Sustainability 2025, 17, 3180 (a) (b)
12 of 19
Figure 9. Examples of (a) Longitudinal cracks skeleton; (b) Transverse cracks skeleton.

The
The skeletonization
skeletonization process
process enhances
enhances precision
precision by
by simplifying
simplifying the the geometry
geometry ofof the
the
cracks while preserving their structural integrity. This step, which calculates
cracks while preserving their structural integrity. This step, which calculates the length of the length
of
thethe cracks,
cracks, reduces
reduces thethe complexity
complexity of the
of the data,data, making
making it easier
it easier to analyze
to analyze the the crack
crack fea-
features accurately.
tures accurately.
Using
Usingaadistance
distancetransform
transform algorithm,
algorithm,thethe
crack metrics
crack metricsareare
calculated to determine
calculated the
to determine
average crack width. The total crack length and average width are converted
the average crack width. The total crack length and average width are converted from from pixels to
real-world units using
pixels to real-world a predefined
units scale factor.
using a predefined Crack
scale length
factor. Crackis reported
length isinreported
meters, in
while
me-
the average width is expressed in centimeters. These metrics form the basis
ters, while the average width is expressed in centimeters. These metrics form the basis for for evaluating
crack severity
evaluating andseverity
crack consequent DV and PCIDV
and consequent calculations. Figure 10 provides
and PCI calculations. an provides
Figure 10 example of an
aexample
PCI report, regarding a few sample units, generated by the proposed model.
of a PCI report, regarding a few sample units, generated by the proposed model.

Figure 10.Example
Figure10. Exampleof
ofPCI
PCIreport
reportgenerated
generatedby
bythe
theproposed
proposedmodel
modelfor
forseveral
severalsample
sampleunits.
units.

5.
5. Validation
Validationof
ofPCI
PCICalculation
CalculationMethod
Method
The
Theproposed
proposed model
Sustainability 2025, 17, x FOR PEER REVIEW
model was
was validated
validated by
by comparing
comparing the
the PCI
PCI calculated
calculated using
using the
the auto-
auto-
13 of 19
mated
mated model with the PCI determined through on-site visual inspections (see Figure11).
model with the PCI determined through on-site visual inspections (see Figure 11).

Figure 11. Validation

Figure11. Validationprocess;
process;on-site
on-sitevisual
visualinspection
inspectionand
andmeasurement.
measurement.

The
Thevalidation
validationofofthe
thePCI
PCIvalues calculated
values by by
calculated thethe
AI model proposed
AI model in this
proposed in study was
this study
conducted by comparing
was conducted by comparingthe crack
the characteristics detected
crack characteristics by the by
detected model—such as length
the model—such as
length (L), width (W), density (D), severity (S), and deduct values (DV)—with their corre-
sponding on-site measurements. The error was then calculated to assess the model’s ac-
curacy.
Among the analyzed sample units, the validation results of four random examples
are presented below. The first illustrative sample unit is n°7 (see Figure 12) and the cracks
 Figure 11. Validation process; on-site visual inspection and measurement.

Sustainability 2025, 17, 3180 The validation of the PCI values calculated by the AI model proposed in this13study of 19
was conducted by comparing the crack characteristics detected by the model—such as
length (L), width (W), density (D), severity (S), and deduct values (DV)—with their corre-
(L), width (W),
sponding density
on-site (D), severityThe
measurements. (S),error
and deduct values
was then (DV)—with
calculated theirthe
to assess corresponding
model’s ac-
on-site
curacy.measurements. The error was then calculated to assess the model’s accuracy.
Among
Among thethe analyzed
analyzed sample
sample units,
units, the
the validation
validation results
results ofoffour
fourrandom
randomexamples
examples
are presented below. The first illustrative sample unit is n ◦ 7 (see Figure 12) and the cracks
are presented below. The first illustrative sample unit is n°7 (see Figure 12) and the cracks
metrics
metricsconfrontation
confrontationisisprovided
providedin inTable
Table2.2.

Figure12.
Figure 12.Sample
Sampleunit ◦ 7.
unitnn°7.

Table2.2.Confrontation
Table Confrontationofofcalculated
calculatedPCI
PCIvia
viaAI
AImodel
modeland
andon-site
on-siteinspection;
inspection;sample
sampleunit ◦ 7.
unitnn°7.

Cracks L L W W D D S S DV
DV PCI
PCI Error
Error
Detection
Detection Method Cracks
Method Direction
Direction (m) (m) (cm) (cm) (%) (%) (-) (-) (-)
(-) (-)
(-) (%)
(%)
Longitudinal 96.65 2.56 24.16 High 79.86
AI-model Longitudinal 96.65 2.56 24.16 High 79.86 22
AI-model Transverse 22
Transverse 71.90 71.90 2.55 2.55 17.9817.98 High
High 76.81
76.81 2.00
On-site Longitudinal 100.52 2.44 25.13 High 82.00 2.00
On-site Longitudinal 100.52 2.44 25.13 High 82.00 20
inspection Transverse 73.48 2.42 18.37 High 78.00 20
inspection Transverse 73.48 2.42 18.37 High 78.00

The model demonstrated an error of −3.85% in estimating the length of longitudinal

The model
cracks and demonstrated
−2.15% an error
for transverse of −
cracks. 3.85%evaluating
When in estimatingthethe lengthcrack
average of longitudinal
width, the
cracks
error was −2.15%for
and+4.92% forlongitudinal
transverse cracks. When
cracks and evaluating
+5.37% the average
for transverse crack
cracks. width,these
Despite the
error was +4.92% for longitudinal cracks and +5.37% for transverse cracks.
differences, the model exhibited a highly accurate performance in calculating the
Sustainability 2025, 17, x FOR PEER REVIEW Despite these
14 PCI,
of 19
differences,
with only athe
2%model exhibited atohighly
error compared on-siteaccurate performance
inspection results. in calculating the PCI, with
only a 2% error compared to on-site inspection results.
The
The second illustrative sample
second illustrative sampleunit n◦ 12
unitisisn°12 (see
(see Figure
Figure 13)13)
andand
the the cracks
cracks metrics
metrics con-
confrontation is provided in Table
frontation is provided in Table 3. 3.

Figure13.
Figure 13.Sample
Sampleunit ◦ 12.
unitnn°12.

Table 3. Confrontation of calculated PCI via AI model and on-site inspection; sample unit n°12.

Cracks L W D S DV PCI Error

Detection Method
Direction (m) (cm) (%) (-) (-) (-) (%)
Sustainability 2025, 17, 3180 14 of 19

Figure 13. Sample unit n°12.

Table 3. Confrontation of calculated PCI via AI model and on-site inspection; sample unit n◦ 12.
Table 3. Confrontation of calculated PCI via AI model and on-site inspection; sample unit n°12.
Detection Cracks L W D S DV PCI Error
Method Cracks
Direction (m) L (cm) W (%) D (-) S DV
(-) PCI
(-) Error
(%)
Detection Method
Direction
Longitudinal 138.69 (m) 2.69 (cm) 34.67 (%) High(-) (-)
86.00 (-) (%)
AI-model Longitudinal 19
AI-model Transverse 62.97138.69 3.15 2.69 15.7434.67 High
High 86.00
75.00 19
Transverse 62.97 3.15 15.74 High 75.00 1.00
On-site Longitudinal 143.57 2.60 35.89 High 86.00 1.00
On-site Longitudinal 143.57 20
inspection Transverse 65.61 3.10 2.60 16.4035.89 High
High 86.00
74.00 20
inspection Transverse 65.61 3.10 16.40 High 74.00

The
The model exhibited aa−−3.40%
model exhibited 3.40% error
error inin estimating
estimating the
the length
length of
of longitudinal
longitudinal cracks
cracks
and −
and a −4.02% error for transverse cracks. When assessing the average crack width, the
a 4.02% error for transverse cracks. When assessing the average crack width, the
model
model showed
showed aa +3.46%
+3.46% error
error for
for longitudinal
longitudinal cracks
cracks and
and aa+1.61%
+1.61% error
error for
fortransverse
transverse
cracks.
cracks. But
But again,
again, the
the model
model maintained
maintained aa high high level
level of
of accuracy
accuracy in
in calculating
calculating the
the PCI,
PCI,
with only a 1% deviation compared to on-site inspection
with only a 1% deviation compared to on-site inspection results. results.
The ◦ 17 (see Figure 14) and the cracks metrics con-
The third
third illustrative
illustrative sample
sample unit
unit is
is nn°17 (see Figure 14) and the cracks metrics con-
frontation
frontationisisprovided
providedin inTable
Table4.4.

Figure14.
Figure 14. Sample
Sampleunit ◦ 17.
unitnn°17.

Table 4. Confrontation of calculated PCI via AI model and on-site inspection; sample unit n◦ 17.

Detection Cracks L W D S DV PCI Error

Method Direction (m) (cm) (%) (-) (-) (-) (%)
Longitudinal 125.30 2.42 31.32 High 86.00
AI-model 20
Transverse 62.49 2.98 15.62 High 74.88
1.00
On-site Longitudinal 130.30 2.53 32.57 High 86.00
inspection 21
Transverse 65.62 2.83 16.40 High 72.00

The model demonstrated a −3.84% error in estimating the length of longitudinal

cracks and a −4.77% error for transverse cracks. For the average crack width, the model
exhibited a −4.34% error for longitudinal cracks and a +5.30% error for transverse cracks.
Despite these variations, the model once again achieved remarkable accuracy in calculating
the PCI, with only a 1% deviation compared to on-site inspection results.
The last illustrative sample unit is n◦ 18 (see Figure 15) and the cracks metrics con-
frontation is provided in Table 5.
 exhibited a −4.34% error for longitudinal cracks and a +5.30% error for transverse cracks.
Despite these variations, the model once again achieved remarkable accuracy in calculat-
ing the PCI, with only a 1% deviation compared to on-site inspection results.
Sustainability 2025, 17, 3180 15 of 19
The last illustrative sample unit is n°18 (see Figure 15) and the cracks metrics con-
frontation is provided in Table 5.

◦ 18.
Figure 15.Sample
Figure15. Sampleunit
unitnn°18.

◦ 18.
Table5.5.Confrontation
Table Confrontationofofcalculated
calculatedPCI
PCIvia
viaAI
AImodel
modeland
andon-site
on-siteinspection;
inspection;sample
sampleunit
unitnn°18.

Detection Cracks
Cracks L L W W D D S S DV
DV PCI Error
Error
Detection Method
Method Direction
Direction (m) (m) (cm) (cm) (%) (%) (-) (-) (-)
(-) (-) (%)
(%)
Longitudinal
Longitudinal 98.64 98.64 2.42 2.42 24.6624.66 High
High 80.17
80.17
AI-model
AI-model 28
Transverse
Transverse 42.34 42.34 2.78 2.78 10.5810.58 High
High 63.50
63.50
1.00
1.00
On-site
On-site Longitudinal
Longitudinal 102.59102.59 2.52 2.52 25.6525.65 High
High 83.00
83.00 27
27
inspection
inspection Transverse
Transverse 44.03 44.03 2.85 2.85 11.0111.01 High
High 62.00
62.00

The model demonstrated a −3.85% error in estimating the length of longitudinal

The model demonstrated a −3.85% error in estimating the length of longitudinal
cracks and a −3.84% error for transverse cracks. For the average crack width, the model
cracks and a −3.84% error for transverse cracks. For the average crack width, the model
exhibited a −3.97% error for longitudinal cracks and a −2.45% error for transverse cracks.
exhibited a −3.97% error for longitudinal cracks and a −2.45% error for transverse cracks.
Despite these variations, the model once again achieved remarkable accuracy in calculat-
Despite these variations, the model once again achieved remarkable accuracy in calculating
ing the PCI, with only a 1% deviation compared to on-site inspection results.
the PCI, with only a 1% deviation compared to on-site inspection results.

6. Results
6. Results and
and Discussion
Discussion
According to ASTM D5340-24 [36], pavement condition can be assessed using a con-
According to ASTM D5340-24 [36], pavement condition can be assessed using a
dition rating system that provides a descriptive evaluation of the pavement’s state based
condition rating system that provides a descriptive evaluation of the pavement’s state
on its PCI value. This rating system helps categorize the condition of each sample unit,
based on its PCI value. This rating system helps categorize the condition of each sample
unit, ranging from Good to Failed, depending on the PCI value. Each condition rating is
often represented using a color-coded scale for easier visualization and interpretation. For
this study, one of the color-coded rating scales suggested by ASTM D5340 was applied, as
explained in Table 6.

Table 6. Adopted Pavement Condition Index (PCI) Rating Scales suggested by ref. [36].

PCI Value Score Rating Color Code

85 to 100 Good Dark Green
70 to 84 Satisfactory Light Green
55 to 69 Fair Yellow
40 to 54 Poor Light Red
25 to 39 Very Poor Medium Red
10 to 24 Serious Dark Red
0 to 9 Failed Dark Grey
Sustainability 2025, 17, 3180 16 of 19

The taxiway pavement at the selected airport has a service life of 20 years. This taxiway
was selected because it has not undergone any maintenance or rehabilitation interventions
in the past 10 years, making it a reliable reference for this study. It is in the northeast of
Italy where it was subject to temperatures lower than −10 ◦ C during the winter and higher
than +37 ◦ C during the summer. The PCI values, calculated by the proposed model, for all
sample units of the taxiway under investigation are summarized in Table 7, where each
unit’s condition is color-coded according to the adopted scale.

Table 7. PCI values and Rating Scales related to each sample unit of understudy taxiway.

Sample PCI Score Sample PCI Score Sample PCI Score Sample PCI Score
Unit Value Rating Unit Value Rating Unit Value Rating Unit Value Rating
n◦ 1 23 Serious n◦ 11 21 Serious n◦ 21 40 Poor n◦ 31 22 Serious
n◦ 2 21 Serious n◦ 12 19 Serious n◦ 22 38 Very Poor n◦ 32 22 Serious
n◦ 3 21 Serious n◦ 13 21 Serious n◦ 23 36 Very Poor n◦ 33 24 Serious
n◦ 4 22 Serious n◦ 14 22 Serious n◦ 24 27 Very Poor n◦ 34 25 Very Poor
n◦ 5 18 Serious n◦ 15 27 Very Poor n◦ 25 20 Serious n◦ 35 24 Serious
n◦ 6 19 Serious n◦ 16 24 Serious n◦ 26 25 Very Poor n◦ 36 30 Very Poor
n◦ 7 22 Serious n◦ 17 20 Serious n◦ 27 30 Very Poor n◦ 37 21 Serious
n◦ 8 30 Very Poor n◦ 18 28 Very Poor n◦ 28 32 Very Poor n◦ 38 22 Serious
n◦ 9 21 Serious n◦ 19 25 Very Poor n◦ 29 22 Serious n◦ 39 23 Serious
n◦ 10 21 Serious n◦ 20 35 Very Poor n◦ 30 23 Serious n◦ 40 21 Serious

As shown in the table, among the 40 sample units of the runway under study, 26 units
are rated as being in Serious condition, 13 units in Very Poor condition, and 1 unit in
Poor condition.
This visualization not only highlights the state of individual pavement sections but
also allows for a quick overview of the overall condition of the taxiway. The color-coded
PCI output serves a dual purpose:
1. Local Maintenance Prioritization: Sections with lower PCI values (e.g., those rated as
poor) can be flagged for immediate maintenance activities, such as crack sealing or
patching, to prevent further deterioration;
2. Long-term Maintenance Planning: The PCI data also aids in designing comprehensive
maintenance and rehabilitation plans, such as resurfacing or reconstruction, for areas
that require more extensive intervention.
This systematic approach provides airport authorities with actionable insights, en-
abling them to allocate resources effectively, prioritize critical repairs, and develop long-
term strategies to ensure the safety and reliability of airport pavements.

7. Conclusions
This study represents a significant advancement in automating airport pavement
management by integrating UAV-based aerial photogrammetry with advanced AI models,
specifically employing YOLO architectures. By utilizing high-resolution UAV imagery in
combination with a custom-trained YOLO11 segmentation model, the research demon-
strates an efficient, scalable, and automated system for detecting and classifying pavement
cracks. The system delivers high precision in one of the most challenging tasks in computer
vision, which is the accurate detection and segmentation of extremely thin elements.
The results highlight an average error rate of ±3.70% for missed longitudinal cracks
and ±3.65% for missed transverse cracks in length, with average errors of ±4.24% and
4.09% for longitudinal and transverse crack widths, respectively. Notably, the automatically
Sustainability 2025, 17, 3180 17 of 19

calculated Pavement Condition Index (PCI) deviates by just ±1.33% from values obtained
through on-site inspection.
Beyond cracks detection and segmentation, the findings highlight the broader poten-
tial of integrating automated crack detection with UAV technology in airport pavement
management. This approach introduces a transformative shift in maintenance practices,
enabling faster, more accurate, and data-driven decision-making. The synergy between
UAVs and advanced AI models significantly reduces inspection times while maintaining
high accuracy in assessments.
In this study, the UAV followed a predefined flight route designed to ensure com-
prehensive coverage of the pavement surface. To achieve high-resolution imagery with
minimal gaps, an 80% overlap in both the transverse and longitudinal directions was ap-
plied. This overlapping strategy improves the accuracy of the orthomosaic reconstruction
and ensures no areas are missed during the survey. The image scale was determined
automatically based on the flight altitude and the specified overlap parameters, ensuring
consistent resolution across all captured images, to achieve a detailed and uniform dataset
suitable for precise crack detection and PCI calculation.
This approach not only improves operational efficiency but also contributes to sus-
tainability by reducing the environmental impact of manual inspections. UAV-based
surveys minimize the need for on-ground equipment and reduce carbon emissions asso-
ciated with traditional pavement evaluation methods. Furthermore, the automation of
PCI calculation ensures timely maintenance, extending pavement lifespan and reducing
resource consumption.
Moreover, the system’s scalable and adaptable design offers potential applications
in other areas of infrastructure management, paving the way for future advancements in
automated inspection technologies.

Author Contributions: Conceptualization, E.T., M.C. and S.C.; methodology, E.T., S.C. and V.P.;
software, V.P.; validation, E.T., and V.P.; formal analysis, V.P.; investigation, V.P.; resources, S.C.; data
curation, E.T. and V.P.; writing—original draft preparation, V.P.; writing—review and editing, M.K.;
visualization, M.K.; supervision, M.K.; project administration, E.T., M.C. and S.C. All authors have
read and agreed to the published version of the manuscript.

Funding: This research received no external funding.

Institutional Review Board Statement: Not applicable.

Informed Consent Statement: Not applicable.

Data Availability Statement: Data is contained within the article.

Conflicts of Interest: The authors declare no conflicts of interest.

References
1. Li, Q.; Zou, Q.; Zhang, D.; Mao, Q. FoSA: F* seed-growing approach for crack-line detection from pavement images. Image Vis.
Comput. 2011, 29, 861–872. [CrossRef]
2. Chen, W.; Yuan, J.; Li, M. Application of GIS/GPS in Shanghai Airport pavement management system. Procedia Eng. 2012, 29,
2322–2326. [CrossRef]
3. Lima, D.; Santos, B.; Almeida, P. Methodology to assess airport pavement condition using GPS, laser, video image and GIS. In
Pavement and Asset Management; CRC Press: Boca Raton, FL, USA, 2019; pp. 301–307.
4. Habib, A.; Lin, Y.J.; Ravi, R.; Shamseldin, T.; Elbahnasawy, M. LiDAR-Based Mobile Mapping System for Lane Width Estimation in
Work Zones; Purdue University: West Lafayette, IN, USA, 2018.
5. Kong, F.; Liu, K.; Cao, C.; Du, X.; Jiao, H. Application of digital construction whole process monitoring and management
technology for airport asphalt pavement. In IOP Conference Series: Earth and Environmental Science; IOP Publishing: Bristol, UK,
2021; Volume 643, p. 012190.
Sustainability 2025, 17, 3180 18 of 19

6. Kovačič, B.; Doler, D.; Sever, D. The innovative model of runway sustainable management on smaller regional airports. Sustain-
ability 2021, 13, 652. [CrossRef]
7. Ford, L.R. Network Flow Theory; RAND Corporation: Santa Monica, CA, USA, 1956; Available online: https://www.rand.org/
pubs/papers/P923.html (accessed on 1 November 2024).
8. Pietersen, R.A.; Beauregard, M.S.; Einstein, H.H. Automated method for airfield pavement condition index evaluations. Autom.
Constr. 2022, 141, 104408.
9. Zhao, H.; Qin, G.; Wang, X. Improvement of canny algorithm based on pavement edge detection. In Proceedings of the 3rd
International Congress on Image and Signal Processing, Yantai, China, 16–18 October 2010; IEEE: Piscataway, NJ, USA, 2010;
Volume 2, pp. 964–967.
10. Zhang, H.; Dong, J.; Gao, Z. Automatic segmentation of airport pavement damage by AM-Mask R-CNN algorithm. Eng. Rep.
2023, 5, e12628.
11. Shon, H.; Cho, C.S.; Byon, Y.J.; Lee, J. Autonomous condition monitoring-based pavement management system. Autom. Constr.
2022, 138, 104222. [CrossRef]
12. Malekloo, A.; Liu, X.C.; Sacharny, D. AI-enabled airport runway pavement distress detection using dashcam imagery. Comput.-
Aided Civ. Infrastruct. Eng. 2024, 39, 2481–2499. [CrossRef]
13. Liu, F.; Xu, G.; Yang, Y.; Niu, X.; Pan, Y. Novel approach to pavement cracking automatic detection based on segment extending.
In Proceedings of the International Symposium on Knowledge Acquisition and Modeling, Wuhan, China, 21–22 December 2008;
IEEE: Piscataway, NJ, USA, 2008; pp. 610–614.
14. Mei, Q.; Gül, M.; Azim, M.R. Densely connected deep neural network considering connectivity of pixels for automatic crack
detection. Autom. Constr. 2020, 110, 103018.
15. Zhu, J.; Zhong, J.; Ma, T.; Huang, X.; Zhang, W.; Zhou, Y. Pavement distress detection using convolutional neural networks with
images captured via UAV. Autom. Constr. 2022, 133, 103991. [CrossRef]
16. Du, Y.; Pan, N.; Xu, Z.; Deng, F.; Shen, Y.; Kang, H. Pavement distress detection and classification based on YOLO network. Int. J.
Pavement Eng. 2021, 22, 1659–1672. [CrossRef]
17. Yang, X.; Li, H.; Yu, Y.; Luo, X.; Huang, T.; Yang, X. Automatic pixel-level crack detection and measurement using fully
convolutional network. Comput.-Aided Civ. Infrastruct. Eng. 2018, 33, 1090–1109.
18. Long, J.; Shelhamer, E.; Darrell, T. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE Conference
on Computer Vision and Pattern Recognition, Boston, MA, USA, 7–12 June 2015; pp. 3431–3440.
19. Ronneberger, O.; Fischer, P.; Brox, T. U-net: Convolutional networks for biomedical image segmentation. In Medical Image
Computing and Computer-Assisted Intervention–MICCAI 2015: 18th International Conference, Munich, Germany, 5–9 October 2015,
Proceedings, Part III 18; Springer International Publishing: Berlin/Heidelberg, Germany, 2015; pp. 234–241.
20. Badrinarayanan, V.; Kendall, A.; Cipolla, R. Segnet: A deep convolutional encoder-decoder architecture for image segmentation.
IEEE Trans. Pattern Anal. Mach. Intell. 2017, 39, 2481–2495.
21. Chen, L.C.; Papandreou, G.; Kokkinos, I.; Murphy, K.; Yuille, A.L. Deeplab: Semantic image segmentation with deep convolutional
nets, atrous convolution, and fully connected crfs. IEEE Trans. Pattern Anal. Mach. Intell. 2017, 40, 834–848. [CrossRef]
22. Redmon, J.; Divvala, S.; Girshick, R.; Farhadi, A. You only look once: Unified, real-time object detection. In Proceedings of the
IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, NV, USA, 27–30 June 2016; pp. 779–788.
23. Mandal, V.; Uong, L.; Adu-Gyamfi, Y. Automated Road crack detection using deep convolutional neural networks. In Proceedings
of the 2018 IEEE International Conference on Big Data (Big Data), Seattle, WA, USA, 10–13 December 2018; pp. 5212–5215.
24. Wu, P.; Liu, A.; Fu, J.; Ye, X.; Zhao, Y. Autonomous surface crack identification of concrete structures based on an improved
one-stage object detection algorithm. Eng. Struct. 2022, 272, 114962.
25. Terven, J.; Córdova-Esparza, D.M.; Romero-González, J.A. A comprehensive review of yolo architectures in computer vision:
From yolov1 to yolov8 and yolo-nas. Mach. Learn. Knowl. Extr. 2023, 5, 1680–1716. [CrossRef]
26. Zhang, C.; Chen, X.; Liu, P.; He, B.; Li, W.; Song, T. Automated detection and segmentation of tunnel defects and objects using
YOLOv8-CM. Tunn. Undergr. Space Technol. 2024, 150, 105857.
27. Brooks, C.; Dobson, R.; Banach, D.; Oommen, T.; Zhang, K.; Mukherjee, A.; Havens, T.; Ahborn, T.; Escobar-Wolf, R.; Bhat, C.;
et al. Implementation of Unmanned Aerial Vehicles (UAVs) for Assessment of Transportation Infrastructure: Phase II; No. SPR-1674;
Michigan Technological University: Houghton, MI, USA, 2018.
28. Knodler, M. The Application of Unmanned Aerial Systems in Surface Transportation; Massachusetts Department of Transportation,
Office of Transportation Planning: Boston, MA, USA, 2020.
29. Jiang, L.; Xie, Y.; Ren, T. A deep neural networks approach for pixel-level runway pavement crack segmentation using drone
captured images. In Proceedings of the Transportation Research Board 99th Annual Meeting, Washington, DC, USA, 12–16
January 2020.
30. Liu, P.; Huang, Y.; Lai, J.; Wu, T.; Tsai, M. A review of rotorcraft unmanned aerial vehicle (UAV) developments and applications in
civil engineering. Smart Struct. Syst. 2014, 13, 1065–1094. [CrossRef]
Sustainability 2025, 17, 3180 19 of 19

31. Aldea, E.; Le Hégarat-Mascle, S. Robust crack detection for unmanned aerial vehicles inspection in an a-contrario decision
framework. J. Electron. Imaging 2015, 24, 061119. [CrossRef]
32. Kumar, P.; Sharma, M. Functional condition evaluation of airfield pavements using automated road survey system—A case study
of a small sized airport. In Road and Airfield Pavement Technology, Proceedings of the 12th International Conference on Road and Airfield
Pavement Technology, Colombo, Sri Lanka, 14–16 July 2021; Springer International Publishing: Cham, Switzerland, 2022; pp. 185–196.
33. Zhai, S.; Xu, Y. The automated segmentation and enhancement of cracks on airport pavements using three-dimensional imaging
techniques. In Proceedings of the International Conference on Algorithm, Imaging Processing, and Machine Vision (AIPMV
2023), Qingdao, China, 15–17 September 2023; SPIE: Bellingham, WA, USA, 2024; Volume 12969, pp. 343–350.
34. Wu, Z.; Zhang, Y.; Zhang, Z.; Lu, N.; Zhang, Z.; Zhang, F.; He, X. A Pavement Distress Detection Method Based on Yolov5 Model.
In Proceedings of the 2022 41st Chinese Control Conference (CCC), Hefei, China, 25–27 July 2022; IEEE: Piscataway, NJ, USA,
2022; pp. 6564–6569.
35. Chen, X.; Liu, C.; Chen, L.; Zhu, X.; Zhang, Y.; Wang, C. A Pavement Crack Detection and Evaluation Framework for a UAV
Inspection System Based on Deep Learning. Appl. Sci. 2024, 14, 1157. [CrossRef]
36. ASTM D5340-24; Standard Test Method for Airport Pavement Condition Index Surveys. ASTM: West Conshohocken, PA,
USA, 2024.
37. Shahin, M.Y. Pavement Management for Airports, Roads, and Parking Lots, 2nd ed.; Springer Science + Business Media, LLC:
New York, NY, USA, 2005.

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.