Proposal of an Integrated Method of Unmanned Aerial Vehicle and Artificial Intelligence for Crack Detection, Classification, and PCI Calculation of Airport Pavements

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 photogrammetry 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.

1. Introduction

The integration of automation into pavement management systems, particularly in airport infrastructure, marks a transformative step forward in civil engineering. Traditional methods for inspecting airport pavements, which rely heavily on manual labor and visual assessments, are often inefficient, costly, and prone to human error. In contrast, modern technologies such as Artificial Intelligence (AI) and Unmanned Aerial Vehicles (UAVs) offer more precise, efficient, and scalable solutions for pavement monitoring and assessment. Moreover, this approach contributes to sustainability by reducing the environmental impact of traditional pavement evaluation methods by minimizing the need for on-ground equipment and reducing carbon emissions associated with.

This research introduces a novel, fully automated system for calculating the Pavement Condition Index (PCI) of airport pavements by utilizing computer vision and machine learning techniques. Specifically, a custom-trained YOLO11 segmentation model developed by the authors is adopted to detect and classify various types of pavement cracks on 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 consistency 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 operations, 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 maintenance 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 [1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35].

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 measurements 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] developed 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 vehicles to minimize operational costs, and Malekloo et al. [12], who developed a cost-efficient AI-enabled system using dashcam imagery for distress classification. These studies highlight the potential of automation to revolutionize pavement management but often face challenges related to dataset diversity, environmental conditions, and scalability.

To address the gaps in these prior studies, this research proposes an accessible and computationally efficient solution. A YOLO11 Custom Model is deployed on a widely used free platform to infer orthophotos of individual pavement sections. The model detects and visualizes longitudinal and transverse cracks through semantic masks, generating reports with detailed information on distress types and quantification. These data are then used to calculate the Pavement Condition Index (PCI) by evaluating parameters such as distress severity, density, and deduct values. This approach simplifies implementation, minimizes computational requirements, and makes advanced pavement evaluation accessible to non-professional users, overcoming key limitations of existing systems.

3. Methodology

This study aims to develop an integrated AI and UAV system for quickly and automatically assessing airport pavement conditions, and the key steps of the proposed methodology are presented in the following flowchart (see Figure 1).

As illustrated in Figure 1, the methodology starts with the development of a custom AI model capable of detecting and segmenting different types of pavement distresses. To build this model, a large dataset of high-resolution images was collected from various airport infrastructures using a lightweight UAV. These images, capturing various types of distresses, form the foundation for training the AI algorithm.

Then, each image was manually labeled by a qualified engineer to ensure accuracy in identifying and classifying pavement defects. This labeling step is crucial for the AI model to learn how to detect and classify different types of distresses effectively.

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 segmenting 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 automated 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 photogrammetry 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.

3.1.2. Dataset Preparation and AI Algorithm Training

Creating the dataset was a crucial step in training the AI models. To facilitate the process of labeling and preparation, a new project was initiated on a widely used computer vision platform (i.e., Roboflow). The raw images were uploaded, and each image underwent detailed manual labeling. This involved annotating every image at the pixel level to precisely identify and classify pavement distresses.

Focus was given to two primary types of cracks: longitudinal and transverse (see Figure 2). The labeling process was conducted with exceptional attention to detail, ensuring even the smallest cracks were accurately identified and annotated. This level of precision is essential for creating a high-quality dataset capable of effectively training the AI models for accurate crack detection and classification.

Red semantic masks in Figure 2 represent longitudinal cracks, and the purple ones represent the transverse cracks.

Once labeling was complete, the dataset was formatted according to the YOLO standard, widely used for object detection and instance segmentation tasks. This format requires each image to be accompanied by a corresponding annotation file in .txt format. The annotation file specifies the object class and the bounding box coordinates for each object in the image, normalized to the image dimensions. Each line in the file represents an object, and the data is structured as follows: <object-class>, <x-center>, <y-center>, <width>, <height>, where the coordinates and dimensions are relative to the image size.

A preprocessing technique was employed to artificially increase the dataset by an order of magnitude. Specifically, each photogram was tiled to train the algorithm without resizing, thereby preserving very thin cracks. This methodology allowed training the algorithm on an artificial big dataset without spending many hours to manually label thousands of pictures.

The dataset was divided into training, validation, and testing subsets to ensure proper generalization during the model training process; 70% of the images were allocated 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 careful dataset formatting was essential for developing a robust and accurate model capable of autonomously identifying pavement distresses. At the end of the dataset configuration, the training process was performed.

3.1.3. Validation of Custom AI Model

To validate the model’s performance, the results from the custom-trained model were compared with on-site visual inspections (see Figure 3). This comparison ensured that the model’s crack detection accuracy could be assessed against real-world observations.

The validation of the AI model proposed in this study to detect cracks was carried out by comparing the lengths of detected longitudinal and transverse cracks with their corresponding on-site measurements, followed by calculating the error. For illustrative purposes, the validation results of two random sample units (see Figure 4) are presented.

The length confrontation of measured on-site and AI-driven detected cracks clarifies that the proposed model detected most cracks accurately on both orthophotos, as proved in

Table 1. Length confrontation of measured on-site and AI-driven detected cracks.

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

.

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

3.2. Pavement Condition Assessment

3.2.1. Image Acquisition

The first step in assessing pavement condition using the proposed methodology is image acquisition. For this study, the survey was carried out on an 800-m-long and 15-m-wide asphalt taxiway at an Italian airport. Approximately 900 high-resolution images were captured from an altitude of 14 m. Each image, with a resolution of 5472 × 3648 pixels, provided detailed and clear views of the pavement, capturing various types of distresses, including longitudinal and transverse cracks.

The high-resolution camera of the adopted UAV captured detailed photographs, essential for generating accurate orthomosaics, which serve as the foundation for subsequent pavement condition analyses. Additionally, the UAV’s extended flight time of up to 30 min per battery allowed extensive area coverage without frequent interruptions, making it a practical and cost-effective solution for airport pavement inspections.

To further enhance accuracy, target patches were strategically placed on the pavement (see Figure 5a) and georeferenced using GPS (see Figure 5b). These targets ensured uniform image scaling and accurate spatial referencing, facilitating precise analysis of pavement conditions.

White and red targets, measuring 30 × 30 cm, were positioned every 40 m along both sides of the taxiway. GPS data was recorded for each high-resolution image, enabling precise localization and mapping of cracks and other pavement distresses. This approach ensured that detected issues were accurately referenced within the orthomosaics. The UAV’s efficiency and ability to cover large areas quickly enabled the survey to be completed in just a few hours, underscoring its reliability and practicality for airport pavement inspections.

3.2.2. Orthophoto Creation

The aerial images captured during the survey were processed using advanced photogrammetry software (i.e., Metashape, developed by Agisoft LLC, a company based in St. Petersburg, Russia) to generate high-quality georeferenced orthomosaics. This processing involved converting raw image data into accurate 3D surface reconstructions using Structure-from-Motion (SfM) algorithms. These algorithms enabled the creation of detailed and precise models, ensuring a reliable representation of the pavement’s condition.

The software’s functionality included generating dense point clouds, creating detailed meshes (see Figure 6a), and applying texture mapping—steps essential for producing orthophotos of exceptional quality.

Geotagging was integrated into the process, with each geotag positioned every 40 m along the taxiway, ensuring accurate spatial referencing for all identified defects, as illustrated in Figure 6b. According to the figure, each blue flag represents a geotag positioned every 40 m along the taxiway.

The resulting orthophotos, with a resolution of 6666 × 6666 pixels, represented real-world areas of 20 × 20 m (see Figure 7). These high-resolution images provided a detailed and scalable representation of the taxiway surface, facilitating the detection and classification of pavement cracks with precision. For this study, based on the length and width of the understudy taxiway, 40 orthophotos were captured.

3.2.3. Local Inference on Taxiway Orthophotos

In this study, due to the computational demands of processing such large images and the inherent efficiency of YOLO models with smaller image patches, a tiling approach was adopted. This method ensured compatibility with the model while preserving the high resolution required for precise crack detection and annotation. In this regard, each orthophoto was divided into smaller tiles (as a computational process), each with a maximum dimension of 1280 × 1280 pixels.

Subsequently to the tiling process, a slight overlap between adjacent tiles is applied to maintain continuity during reconstruction and to minimize information loss or misalignment at the edges. The image was segmented into smaller sections based on a grid system, using predefined coordinates to crop the original orthophoto into tiles suitable for analysis.

Once divided, each tile was processed by the YOLO11 Custom Model, which employs segmentation algorithms trained on manually annotated datasets to detect and classify pavement cracks.

To ensure reliable results, a confidence threshold was applied during the detection process. This threshold, which can range from 0% (where the model considers excessive details, potentially leading to over-detection of cracks) to 100% (where the model only identifies details it is fully confident about, potentially resulting in under-detection), was carefully calibrated. For this study, a confidence threshold of 50% was chosen, striking a balance by allowing only highly certain predictions to be annotated. This reduced false positives and improved the accuracy of defect identification.

After processing, the annotated tiles were reassembled to reconstruct the original orthophoto, now fully annotated with detected cracks. The reassembly involved aligning the tiles based on their original grid coordinates, with the overlapping sections ensuring seamless continuity of cracks that spanned multiple tiles.

After processing, the annotated tiles were reassembled to reconstruct the original orthophoto (see Figure 8), now fully annotated with detected cracks. The reassembly involved aligning the tiles based on their original grid coordinates, with the overlapping sections ensuring seamless continuity of cracks that spanned multiple tiles.

As demonstrated, to facilitate interpretation, different colors were used to represent 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.

4. PCI Calculation Method

The calculation of the Pavement Condition Index (PCI) is a vital component in evaluating the condition of airport pavements. This study introduces an automated methodology that analyzes annotated orthophotos to quantify crack characteristics, such as length, width, density, and severity.

The model is then used to compute the PCI following the ASTM D5340 standard [36], which provides a systematic framework for assessing pavement conditions. This approach ensures that inspections are consistent and reliable, enabling airport authorities to compare pavement condition data across various time periods and locations.

The ASTM D5340 outlines a detailed process for identifying and assessing different types of pavement distresses, including cracking. In accordance with standard procedures, the first step is to divide the taxiway—classified as a section of the airport—into smaller sample units. According to specifications [36], it is recommended to divide a section into sample units of between 185 m² and 454 m² for asphalt-surfaced airfields. For this study, the taxiway was divided into 40 sample units, each covering 400 m².

For each distress detected within a sample unit, its metrics must be quantified using appropriate measurement units. In this study, longitudinal and transverse cracks were quantified by their length, as specified by the standards. By dividing the total length of each distress type by the area of the sample unit, the density level of the distress was determined, as demonstrated in Equations (1) and (2).

$$Densit{y}_{ longitudinal cracks}\left[\%\right]={\displaystyle \frac{Total lenght of longitudianl cracks\left[\mathrm{m}\right]}{Sample unit Area \left[{\mathrm{m}}^2\right]}}\times 100$$

(1)

$$Densit{y}_{ tranverse cracks}\left[\%\right]={\displaystyle \frac{Total lenght of transverse cracks\left[\mathrm{m}\right]}{Sample unit Area \left[{\mathrm{m}}^2\right]}}\times 100$$

(2)

Another parameter to calculate the PCI is the severity level of each crack, which can be categorized as low, medium, or high. The standard [36] defines the severity for linear cracks in asphalt concrete airfield pavement based on their average width as follows.

• Low severity with average crack width less than 6.4 mm;
• Medium severity with average crack width between 6.4 mm and 25.4 mm;
• High severity with average crack width greater than 25.4 mm.

Once the density of each distress and its 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 quantifying the impact of each distress on pavement condition.

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).

$${DV}_H=0.006352\times {DD}^3-0.40525\times {DD}^2+9.04947\times DD+8.19191$$

(3)

$${DV}_M=0.003702\times {DD}^3-0.2258\times {DD}^2+4.8119\times DD+3.1753$$

(4)

$${DV}_L=0.002362\times {DD}^3-0.13465\times {DD}^2+2.8948\times DD-0.3497$$

(5)

In these equations, DV_H represents the deduct value for cracks with high severity; DV_M represents the deduct value for cracks with medium severity; DV_L 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)).

$${PCI}_i=100-{DV}_i (i=1,\dots ,n)$$

(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 standardized 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 for measuring the average crack width and length with high precision, providing detailed input for PCI computation.

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 features accurately.

Using a distance transform algorithm, the crack metrics are calculated to determine the average crack width. The total crack length and average width are converted from pixels to real-world units using a predefined scale factor. Crack length is reported in meters, while the average width is expressed in centimeters. These metrics form the basis for evaluating crack severity and consequent DV and PCI calculations. Figure 10 provides an example of a PCI report, regarding a few sample units, generated by the proposed model.

5. Validation of PCI Calculation Method

The proposed model was validated by comparing the PCI calculated using the automated model with the PCI determined through on-site visual inspections (see Figure 11).

The validation of the PCI values calculated by the AI model proposed in this study 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 corresponding on-site measurements. The error was then calculated to assess the model’s accuracy.

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 metrics confrontation is provided in

Table 2. Confrontation of calculated PCI via AI model and on-site inspection; sample unit n°7.

Detection Method	Cracks Direction	L	W	D	S	DV	PCI	Error
		(m)	(cm)	(%)	(-)	(-)	(-)	(%)
AI-model	Longitudinal	96.65	2.56	24.16	High	79.86	22	2.00
	Transverse	71.90	2.55	17.98	High	76.81
On-site inspection	Longitudinal	100.52	2.44	25.13	High	82.00	20
	Transverse	73.48	2.42	18.37	High	78.00

.

The model demonstrated an error of −3.85% in estimating the length of longitudinal cracks and −2.15% for transverse cracks. When evaluating the average crack width, the error was +4.92% for longitudinal cracks and +5.37% for transverse cracks. Despite these differences, the model exhibited a highly accurate performance in calculating the PCI, with only a 2% error compared to on-site inspection results.

The second illustrative sample unit is n°12 (see Figure 13) and the cracks metrics confrontation is provided in

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

Detection Method	Cracks Direction	L	W	D	S	DV	PCI	Error
		(m)	(cm)	(%)	(-)	(-)	(-)	(%)
AI-model	Longitudinal	138.69	2.69	34.67	High	86.00	19	1.00
	Transverse	62.97	3.15	15.74	High	75.00
On-site inspection	Longitudinal	143.57	2.60	35.89	High	86.00	20
	Transverse	65.61	3.10	16.40	High	74.00

.

The model exhibited a −3.40% error in estimating the length of longitudinal cracks and a −4.02% error for transverse cracks. When assessing the average crack width, the model showed a +3.46% error for longitudinal cracks and a +1.61% error for transverse cracks. But again, the model maintained a high level of accuracy in calculating the PCI, with only a 1% deviation compared to on-site inspection results.

The third illustrative sample unit is n°17 (see Figure 14) and the cracks metrics confrontation is provided in

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

Detection Method	Cracks Direction	L	W	D	S	DV	PCI	Error
		(m)	(cm)	(%)	(-)	(-)	(-)	(%)
AI-model	Longitudinal	125.30	2.42	31.32	High	86.00	20	1.00
	Transverse	62.49	2.98	15.62	High	74.88
On-site inspection	Longitudinal	130.30	2.53	32.57	High	86.00	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 confrontation is provided in

Table 5. Confrontation of calculated PCI via AI model and on-site inspection; sample unit n°18.

Detection Method	Cracks Direction	L	W	D	S	DV	PCI	Error
		(m)	(cm)	(%)	(-)	(-)	(-)	(%)
AI-model	Longitudinal	98.64	2.42	24.66	High	80.17	28	1.00
	Transverse	42.34	2.78	10.58	High	63.50
On-site inspection	Longitudinal	102.59	2.52	25.65	High	83.00	27
	Transverse	44.03	2.85	11.01	High	62.00

.

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 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 calculating the PCI, with only a 1% deviation compared to on-site inspection results.

6. Results and Discussion

According to ASTM D5340-24 [36], pavement condition can be assessed using a condition rating system that provides a descriptive evaluation of the pavement’s state 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. 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

.

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

Sample Unit	PCI Value	Score Rating	Sample Unit	PCI Value	Score Rating	Sample Unit	PCI Value	Score Rating	Sample Unit	PCI Value	Score 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

, where each unit’s condition is color-coded according to the adopted scale.

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:

• 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;
• 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, enabling 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 demonstrates 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 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 potential 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 comprehensive coverage of the pavement surface. To achieve high-resolution imagery with minimal gaps, an 80% overlap in both the transverse and longitudinal directions was applied. 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 sustainability by reducing the environmental impact of manual inspections. UAV-based surveys minimize the need for on-ground equipment and reduce carbon emissions associated 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.