An Integrated System for Fine-Grained Crack Identification and Dynamic PCI Assessment of Asphalt Pavements with Geometric Features

Abstract

Asphalt pavement maintenance is critical for road service life and traffic safety, yet conventional crack detection and Pavement Condition Index (PCI) assessment methods suffer from inefficiency and subjectivity. This paper presents an integrated system for intelligent crack recognition and automated PCI evaluation, aiming to bridge the gap between automated identification and intelligent assessment. The system employs an optimized YOLOv11l-seg model for precise crack segmentation and geometric parameter extraction, and introduces a refined PCI model incorporating geometry-based adjustment factors for differentiated scoring. Using unmanned aerial vehicle (UAV) data, a fully automated workflow is established—from image acquisition and stitching to crack detection, PCI calculation, and result visualization. Experimental results demonstrate the accuracy of extracted crack parameters and the superior discriminative capability and engineering rationality of the proposed PCI model over conventional approaches. The generated panoramic condition maps provide intuitive visual support for maintenance decision-making. This research validates the feasibility of a fully auto-mated closed-loop system from detection to evaluation, offering a practical solution for intelligent pavement maintenance.

1. Introduction

Asphalt pavements, being a core component of transportation infrastructure, directly influence road longevity and traffic safety through their service performance. With China’s transport development strategy shifting from large-scale new construction to a balanced approach of construction and maintenance, this places heightened demands on the precision and timeliness of asphalt pavement maintenance. Among various pavement defects, cracking represents the most prevalent and initial form of deterioration. Research indicates that cracks are not only a visible manifestation of structural degradation but also a key factor in inducing secondary defects, such as water ingress [1] and loss of base bearing capacity, which significantly accelerates pavement performance deterioration. Consequently, achieving early, precise identification and scientifically quantifiable assessment of pavement cracks is a decisive step in implementing preventive maintenance, optimizing maintenance resource allocation, and extending asset service life, carrying significant engineering and economic value.

However, traditional road condition surveys rely heavily on manual visual inspections and handwritten records, exhibiting inherent drawbacks such as inefficiency, high subjectivity, significant safety risks, and poor data standardization and traceability. Faced with an ever-expanding road network and demands for more refined management, this traditional approach has become unsustainable. Developing automated and intelligent road inspection and assessment technologies has thus become both an industry consensus and an urgent necessity.

Presently, pavement crack detection technology is undergoing a profound transformation from traditional manual methods towards intelligent automation [2]. Automated evaluation systems integrating deep learning with multi-source sensing have emerged as a recent research focus, demonstrating the potential for high-precision automated systems [3]. Within this process, the selection and application of technical parameters [4] and performance metrics prove crucial for effective pavement management [5]. Conventional PCI assessments rely heavily on experienced engineers conducting field surveys and manual documentation, exhibiting numerous drawbacks including inefficiency, high subjectivity, significant safety risks, and poor data traceability. Although existing research has attempted automated assessment through integrated multi-sensor systems, such as the REMS system developed by Setyawan et al. [6], which combines accelerometers, cameras, and laser sensors to synchronously capture IRI and damage imagery—these systems have yet to establish automated computational logic linking damage identification to PCI scoring. Consequently, they fail to form a complete inspection-evaluation closed loop. This reflects a common deficiency in current automated assessment systems regarding the deep integration of evaluation models, rendering them ill-suited to meet the demands of modern, large-scale road network maintenance management.

Computer vision technologies, exemplified by deep learning, provide robust support for automated crack detection. In recent years, automated pavement distress detection based on vision technology has become a research focus, with numerous reviews and practices on deep learning-based methods emerging [7,8]. Single-stage object detection and instance segmentation models such as the YOLO series have gained widespread application in this field due to their favorable balance of speed and accuracy [9]. To accommodate deployment requirements on mobile platforms like drones, lightweight model design has also emerged as a research focus. For instance, Meng et al. [10] employed knowledge distillation to achieve lightweight adaptation of the YOLOv8 model, enhancing inference speed while maintaining high detection accuracy. Nevertheless, such research remains focused on model compression and acceleration, failing to deeply integrate lightweight detection outcomes with pavement maintenance evaluation systems. Consequently, a complete closed-loop system encompassing “lightweight detection–detailed assessment” has yet to be established. Regarding evaluation models, some studies have attempted to simplify PCI computation using machine learning methods. For instance, Kheirati and Golroo [11] grouped 20 pavement defects into four categories via cluster analysis and constructed an ANN model to fuse multi-dimensional pavement data for evaluation, markedly improving assessment efficiency. Furthermore, Elhadidy et al. [12] established a high-precision regression model between IRI and PCI based on the LTPP database, offering a novel approach for indirectly assessing pavement condition using automatically collected smoothness data. However, such methods still rely on traditional inspection techniques to obtain IRI data, failing to achieve automated, refined PCI calculation directly from crack image recognition results.

With advances in computer technology, computer vision techniques represented by deep learning have provided powerful solutions for automated crack detection [13]. Researchers have also begun attempting to establish preliminary closed-loop systems spanning detection to assessment [14]. Recently, Ranjbar et al. [15] developed an integrated system for identifying, segmenting, and classifying the severity of oil bleeding defects in road surfaces. Nevertheless, such approaches either remain confined to single defect types or fail to integrate deeply with standardized PCI evaluation models covering multiple crack types, nor do they achieve automated PCI calculation based on geometric parameters.

It is noteworthy that despite the rapid advancement of automated detection technologies, their accuracy in identifying low-contrast cracks—such as light-colored, minute fissures and fatigue cracks—remains suboptimal. This represents not only an algorithmic challenge but is also intrinsically linked to the microstructural evolution resulting from asphalt material ageing [16]. Furthermore, effectively translating detection outcomes into maintenance decision support—achieving the transition from “data visualization” to “intelligent decision-making”—represents an urgent research priority. At the micro-mechanism level, asphalt ageing involves changes in micelle size and surface activity, with these microstructural evolutions directly influencing macroscopic cracking behavior [17]. Consequently, establishing a closed-loop system integrating micro-mechanism understanding with macro-level intelligent assessment will be pivotal to advancing pavement maintenance from “perception” to “cognition”.

The evolution from bounding-box-based detection to pixel-level crack segmentation is a critical step in improving parameter extraction accuracy, an approach validated in related studies [18]. The YOLOv11l-seg model optimized in the preliminary phase of this study incorporates the DynamicHead detection head [19] within its original framework to enhance multi-scale feature fusion capabilities. It further replaces the original C2f and Bottleneck structures with dilation-wise residual (DWR) modules [20] to improve feature extraction for small targets and complex-shaped cracks. This optimized model demonstrated outstanding performance in crack segmentation tasks, achieving an mAP50-95 of 0.63. It significantly improved crack recognition and contour extraction accuracy under complex lighting and noisy conditions while maintaining high inference speed.

A significant issue across existing research is that most studies terminate at the endpoint of automated crack detection—that is, they conclude after locating and classifying cracks within images, failing to integrate with maintenance evaluation systems such as the Pavement Condition Index [21,22]. These precisely identified crack datasets remain unintegrated with the Pavement Condition Index—a critical maintenance assessment framework in engineering practice. Consequently, an automated closed-loop process spanning detection to evaluation remains unattainable.

Moreover, existing research predominantly focuses on the macroscopic identification and classification of cracks, with insufficient attention paid to the micro-mechanisms of crack formation, material degradation pathways, and the energy evolution process of structural defects. For instance, at the micro-level, the topological structure of the aggregate skeleton within asphalt mixtures and the force chain transmission mechanism directly influence its crack resistance and rutting resistance [23]; while the degradation of asphalt-resin adhesion capacity is closely linked to the diminished stability of micelle structures during ageing [24]. The crack-resistance performance of pavement base materials is similarly rooted in their microstructure at the macroscopic level. For instance, microstructural characterization studies of crack-resistant base materials such as cement-stabilized crushed stone have revealed intrinsic relationships between material composition, interfacial transition zones, and micro-mechanical properties. This provides a fundamental understanding of mechanisms for preventing reflective cracking at the material design level [25]. Regarding structural cracks, the formation mechanism of reflective cracks can be rationally explained from an energy evolution perspective [26], and the development of embedded strain sensors has provided new avenues for synchronous monitoring of dynamic strain and modulus in asphalt layers [27]. Although these studies do not directly address automated identification systems, they provide crucial microstructural and mechanistic foundations for understanding crack genesis, progression, and their impact on pavement performance. They also point the way towards further optimization of intelligent assessment systems—such as incorporating material state parameters and coupling micro-damage models. Consequently, establishing a closed-loop system that integrates micro-level mechanistic understanding with macro-level intelligent assessment will be pivotal to advancing pavement maintenance from mere “sensing” to true “cognition”.

To address the aforementioned issues and bridge the gap between automated recognition and intelligent evaluation, this paper endeavors to construct an integrated system for intelligent crack recognition and automated PCI assessment. The contributions of this research are as follows: (1) This paper focuses on building an integrated system, utilizing an optimized, high-performance recognition model as the core of the entire system to achieve automated extraction of crack parameters. (2) An enhanced, more refined PCI evaluation model is proposed, incorporating geometry-based adjustment factors (derived from multiple linear regression) and differentiated deduction rules with severity-dependent deduction ceilings, enabling more scientifically weighted quantification of crack impacts on pavement performance. (3) Through field experiments using unmanned aerial vehicles (UAVs), the feasibility of the entire workflow—from data acquisition to result visualization—is validated, providing intuitive support for maintenance decision-making.

The subsequent sections of this paper are organized as follows: Section 2 clarifies the objectives and scope of this research; Section 3 details the system’s overall framework, crack recognition and parameter extraction modules, accuracy validation protocols, optimized PCI model design, and data acquisition and visualization workflows; Section 4 describes image acquisition methods and image library construction; Section 5 presents and discusses results, including recognition accuracy validation, PCI evaluation comparisons, statistical analysis of recognition outcomes, and visualization effects; Section 6 summarizes the paper and outlines future research directions.

2. Objective and Scope

To address the disconnect between automated identification and intelligent assessment in current pavement maintenance research, alongside the inefficiency and high subjectivity of traditional PCI evaluation methods, this study aims to develop and validate an integrated automated pavement assessment system. Specific research objectives are as follows:

• Construct and validate an automated pavement assessment system integrating advanced crack detection with an optimized PCI evaluation model;
• Design an optimized PCI model incorporating crack type and severity to enhance the scientific rigor and engineering applicability of the assessment.
• Generate visualized panoramic road condition maps through drone-based comparative trials to demonstrate and validate the effectiveness of the evaluation results.

This research scope builds upon the previously optimized YOLOv11l-seg model, focusing on the identification and evaluation of four typical crack types: transverse cracks, longitudinal cracks, network cracks, and irregular cracks. It aims to demonstrate the core functionality and feasibility of the proposed integrated system.

3. Methodology

3.1. Crack Identification and Parameter Extraction Scheme Design

3.1.1. Model Overview

This system employs an optimized YOLOv11l-seg model as its core recognition engine. The model enhances multi-scale feature fusion capabilities by incorporating a DynamicHead detection head, whilst replacing the original core components with a Dilation-Wise Residual (DWR) module to improve feature extraction for small target cracks. In the crack segmentation task, this optimized model demonstrates outstanding performance, achieving an mAP50-95 of 0.63. It outperforms mainstream comparative models across key metrics, including Precision and Recall, providing a reliable foundation for subsequent automated extraction of geometric parameters.

3.1.2. Parameter Extraction

Building upon the instance segmentation outputs of the YOLOv11l-seg model, this system utilizes the pixel-level mask boundary point coordinates it generates to automatically compute key geometric parameters of cracks. For different crack types, we employ differentiated extraction strategies, the computational principles of which are outlined below:

(1) Linear Cracks (Transverse, Longitudinal):

For linear cracks such as transverse and longitudinal fissures, the system rapidly estimates their length and average width based on their bounding rectangles. This method ensures sufficient precision while maintaining engineering efficiency. As illustrated in Figure 1, the extraction rules are defined as follows:

Length: The dimension of the longer side of the bounding rectangle is taken, i.e.,

$$L=\max (a,b)$$

(1)

Average width: Take the shorter side of the enclosing rectangle, i.e.,

$$W_{avg}=\min (a,b)$$

(2)

(2) Surface cracks (reticulated, vertical):

For network cracks and irregular cracks, the failure pattern is predominantly regional, making the area the core evaluation metric. The system directly utilizes the mask boundary point coordinates predicted by the model, employing the shoelace formula to calculate their projected area.

Assuming a crack mask comprises an ordered sequence of boundary points $\left(x_1,y_1\right),\left(x_2,y_2\right),\dots ,(x_n,y_n$), the formula for its area A is:

$$A={\displaystyle \frac{1}{2}}|{\displaystyle \sum _{i=1}^{n-1}(x_i{y}_{i+1}-x_{i+1}{y}_i)+(x_n{y}_1-x_1{y}_n)}|$$

(3)

As illustrated in Figure 2, this method achieves precise fitting of irregular crack shapes by decomposing polygonal regions into trapezoids for computation and accumulation, thereby yielding accurate pixel areas. Subsequent coordinate system conversion then provides the actual physical area.

Through the aforementioned approach, the system enables automated, batch extraction of key geometric parameters for four crack types, furnishing precise data inputs for subsequent PCI evaluation models.

3.2. Accuracy Verification Scheme Design

To quantitatively assess the accuracy of crack geometric parameters extracted using the YOLOv11l-seg model, thereby providing data support for the reliability of subsequent PCI evaluation models, this study designed a dedicated precision verification experiment. This approach uses manual field measurement results as the benchmark, comparing them with parameters automatically identified and extracted by the model. The methodological workflow is as follows:

• Sample Selection: Within the Section 4, 20 representative cracks were selected as samples, chosen for their accessibility and ease of measurement, ensuring coverage of both transverse and longitudinal linear crack types.
• Manual Measurement: Using high-precision tape measures, crack width gauges, and other tools, professionals conducted multiple measurements of the length and average width of the sample cracks, recording the results. The mean value was taken as the true geometric parameter for each crack.
• Model extraction: Within the panoramic images captured and stitched by the UAV, the aforementioned sample cracks were located. The YOLOv11l-seg model was then employed to automatically identify and extract their corresponding lengths and average widths.
• Accuracy calculation: Mean Absolute Error (MAE) was adopted as the core evaluation metric, calculated using the following formula:

$$MAE=({\displaystyle \frac{1}{n}}\times {\displaystyle \sum |\mathrm{Predicted} \mathrm{value}-\mathrm{Actual} \mathrm{value}|})$$

(4)

This metric provides a straightforward indication of the average magnitude of error between the model’s predicted values and the manually measured values [28,29].

3.3. Optimized PCI Evaluation Model Design

As the core of this research, this chapter aims to construct a PCI evaluation model that integrates seamlessly with automated identification systems while offering greater precision. This model will undergo systematic optimization, encompassing the classification of crack severity, the formulation of quantitative evaluation rules, and the development of the final calculation formula.

3.3.1. Severity Grading

Scientific grading is the prerequisite for accurate assessment. Drawing upon mainstream domestic and international standards alongside engineering practice experience, this study categorizes the severity of various cracks into three uniform grades—minor, moderate, and severe—primarily based on crack width, edge damage condition, and degree of material loss. The specific definitions are outlined in

Table 1. Classification Criteria for Crack Severity Levels.

Severity Level	Example of A Linear Crack	Classification Criteria
Mild		Crack width ≤ 3 mm, intact edges, no or minor spalling, no material loss, basically clean cracks
Moderate		Crack width > 3 mm and ≤6 mm, minor spalling or loose materials at edges, possible minor material loss, debris in cracks
Severe		Crack width > 6 mm, severe spalling or peeling at edges, significant material loss, cracks filled with debris, or obvious faulting (>3 mm) on both sides

.

Concurrently, to account for the distinct morphological characteristics of different crack types and their varying impacts on pavement performance, this study has established specific quantitative evaluation metrics and measurement units for each, ensuring targeted and accurate assessments. Details are presented in

Table 2. Evaluation Metrics for Different Crack Types.

Crack Type	Measurement Unit	Evaluation Criteria
Transverse Crack	Effective count (number/200 m)	Cracks with a length ≥ 0.3 m are counted as one effective crack
Longitudinal crack	Effective length (m/200 m)	Cracks with a length < 0.3 m are not counted
Irregular Crack	Damaged area ratio (%)	More severe damage should result in higher deduction scores
Map Cracking	Damaged area ratio (%)	More severe damage should result in higher deduction scores

.

As shown in Table 2, for linear cracks (transverse and vertical), the primary consideration is their extent of propagation. These are quantified using the effective number of cracks and effective length, respectively, with a measurement threshold of 0.3 m established to disregard insignificant minor cracks. For planar cracks (irregular, networked), the primary consideration is their affected area, with the damage area ratio serving as the core metric. This differentiated evaluation framework establishes the foundation for subsequent differentiated deduction models.

3.3.2. Crack Assessment Form for Evaluation Models

Based on the aforementioned grading system and evaluation criteria, this study has formulated detailed crack deduction rules to achieve automatic mapping from crack parameters to PCI deduction values. These rules are designed separately for linear cracks and planar cracks, as shown in

Table 3. Specification for Deduction Points for transverse crack.

Metric Indicators	Severity	Valid Range of Values	Penalty Points	Max Cumulative Penalty Points
Number of effective strands (strands/200 m)	mild	0–1	1.0	10
		2–3	2.0
		≥4	3.0
	moderate	0–1	3.0	20
		2–3	5.0
		≥4	8.0
	severe	0–1	6.0	30
		≥2	10.0

,

Table 4. Specification for Deduction Points for longitudinal crack.

Metric Indicators	Severity	Valid Range of Values	Penalty Points	Max Cumulative Penalty Points
Effective length (meter/200 m)	mild	0–1	3.0	15
		2–3	5.0
		4–6	8.0
		≥7	12.0
	moderate	0–0.5	5.0	25
		1–2	8.0
		3–4	12.0
		≥5	18.0
	severe	0–0.3	10.0	35
		0.6–1.0	15.0
		≥1.5	20.0

,

Table 5. Detailed Penalty Criteria for map cracking.

Metric Indicators	Severity	Valid Range of Values	Penalty Points	Max Cumulative Penalty Points
Percentage of damaged area (%)	mild	0–5	2.0	15
		6–10	3.0
		11–20	4.0
		≥21	5.0
	moderate	0–3	4.0	25
		4–7	5.0
		8–15	6.0
		≥16	7.0
	severe	0–2	7.0	35
		3–5	9.0
		≥6	12.0

and

Table 6. Detailed Penalty Criteria for Irregular crack.

Metric Indicators	Severity	Valid Range of Values	Penalty Points	Max Cumulative Penalty Points
Percentage of damaged area (%)	mild	0–3	3.0	20
		4–7	5.0
		8–15	7.0
		≥16	10.0
	moderate	0–2	6.0	30
		3–5	9.0
		6–10	12.0
		≥11	15.0
	severe	0–1	12.0	40
		2–4	18.0
		≥5	25.0

.

The design of the deduction criteria embodies the following core principles:

Graded deductions: For the same severity level, progressively increasing deduction values are applied across different thresholds of effective quantity or area proportion, enabling evaluation outcomes to more sensitively reflect the scale of defects.

Differentiated Weighting: A comparison of Table 3 and Table 4 with Table 5 and Table 6 reveals that, for comparable severity and scale, surface cracks (particularly severe irregular cracks) generally incur higher deduction values and upper limits than linear cracks. This reflects the differing impacts of crack types on pavement structural integrity and performance, assigning implicitly higher weightings to networked and irregular cracks.

Max deduction control: Cumulative deduction caps are established for each crack type, adhering to fundamental PCI evaluation principles. This prevents excessive overall scoring reductions due to localized severe damage, thereby preserving differentiation and ensuring engineering rationality in assessment outcomes.

3.3.3. Calibration of Geometric Adjustment Factors Based on Regression Analysis

To objectively determine the geometric adjustment factors for linear cracks, this study collected linear crack data and corresponding PCI values from multiple keyframe image units. For each unit, longitudinal cracks were statistically categorized by width severity levels, recording the effective lengths (L_mild, L_mod, L_sev). Transverse cracks were statistically categorized by the proportion of lane width they cover, recording the effective counts (T_mild, T_mod, T_sev).

A multiple linear regression model was established as follows:

$$PCI={\beta }_0-({\beta }_1{L}_{\mathrm{mild}}+{\beta }_2{L}_{\mathrm{mod}}+{\beta }_3{L}_{\mathrm{sev}}+{\beta }_4{T}_{\mathrm{mild}}+{\beta }_5{T}_{\mathrm{mod}}+{\beta }_6{T}_{\mathrm{sev}})+\epsilon$$

(5)

where β₀ = 100, PCI is the dependent variable, β₁–β₆ are the regression coefficients for the independent variables, and ε is the random error term.

To validate the rationality of the multiple linear regression model, partial leverage plots and residual plots were used. In the partial leverage plots (Figure 3), the horizontal coordinate of each point represents the leverage value of the variable, and the vertical coordinate represents the residual after removing the influence of that variable. The slope of the fitted line is the partial regression coefficient. The results (Figure 3) show that the slopes for transverse cracks at mild, moderate, and severe levels follow an approximate ratio of 1:1.4:1.8; for longitudinal cracks, the ratio is approximately 1:1.5:2.0.

Residual plots (Figure 4) were used to test the model assumptions. As shown in Figure 4, the residuals are randomly distributed around the zero line without any obvious trend, indicating linearity and homoscedasticity. This supports the validity of the regression coefficient estimates.

The regression coefficients obtained were: L-mild = 0.198, L-mod = 0.297, L-sev = 0.401, T-mild = 0.490, T-mod = 0.689, T-sev = 0.892. All coefficients were significant at the p < 0.001 level, with R² = 0.92.

Taking the mild severity level as the baseline (coefficient set to 1.0), the adjustment factors for other severity levels were calculated as:

Longitudinal cracks: moderate = β₂/β₁ = 1.5, severe = β₃/β₁ = 2.03 → rounded to 1.0, 1.5, 2.0

Transverse cracks: moderate = β₅/β₄ = 1.4, severe = β₆/β₄ = 1.82 → rounded to 1.0, 1.4, 1.8

These rounded values, presented in

Table 7. Adjustment Factors for Geometric Characteristics of Linear Cracks.

Crack Types	Range of Geometric Parameters	Grade	Degree of Impact	Adjustment Factor	Effective Value Calculation
transverse crack	width < 3 mm	mild	Low	1.0	Actual length × 1.0
	3 mm ≤ width ≤ 6 mm	moderate	Moderate	1.5	Actual length × 1.5
	width > 6 mm	severe	high	2.0	Actual length × 2.0
longitudinal crack	Length < 1/3 Road width	mild	low	1.0	Actual quantity × 0.9
	1/3 Road width ≤ Length ≤ 2/3 Road width	moderate	moderate	1.4	Actual quantity × 1.4
	Length > 2/3 Road width	severe	high	1.8	Actual quantity × 1.8

Note: The adjustment factors are derived from multiple linear regression analysis (see Section 3.3.3) and rounded to one decimal place. The ratios (1.0:1.5:2.0 for longitudinal cracks; 1.0:1.4:1.8 for transverse cracks) reflect the relative impact of different severity levels on pavement condition.

, are consistent with engineering knowledge that wider or longer cracks have a greater impact on pavement performance. The adjustment factors are applied to convert the raw metric (length or count) into an adjusted effective value, which is then used to query the corresponding deduction points in the penalty tables (Table 3, Table 4, Table 5 and Table 6). This ensures that the deduction points simultaneously account for both the “scale” and the “severity” of each crack. To further refine the assessment, this study innovatively introduces a geometric characteristic adjustment factor. This factor aims to modify single-dimensional metrics (e.g., length, number of cracks) based on another key geometric characteristic (e.g., width, length-to-width ratio), thereby more comprehensively reflecting the actual hazard posed by cracks. The adjustment rules are detailed in Table 7.

Taking transverse cracks as an example, a crack with an actual length of 2 m and a width of 5 mm (moderate) has an adjusted effective length of 2 m × 1.5 = 3 m. The system will use this “3 m” as the effective length to query deductions according to the detailed rules in Table 4. This mechanism ensures the final deduction value incorporates both the scale and severity of the crack, rendering the PCI assessment outcome more scientifically sound and accurate.

Sensitivity Analysis: To evaluate how sensitive the final PCI score is to variations in the adjustment factors, we performed a simple sensitivity test on test section A. The adjustment factors for both longitudinal and transverse cracks were separately multiplied by 0.8 and 1.2 (i.e., ±20% variation), while keeping all other calculation procedures unchanged. The resulting PCI scores varied by less than ±3.2% compared to the original score of 77.8. This indicates that the PCI evaluation is reasonably robust to moderate variations in the adjustment factors, and the specific rounded values derived from regression are reliable for engineering practice.

3.3.4. PCI Model Optimization Approach

This study has implemented three core optimizations to the traditional PCI model, aiming to enhance the evaluative results’ discriminatory power and better align them with engineering practice.

(1)

Weighting Adjustment: Traditional PCI models exhibit insufficient discrimination between different crack types. This study implements implicit weighting through differentiated deduction criteria and upper limits. As shown in Table 3, Table 4, Table 5 and Table 6, under comparable severity and scale, the deduction values and upper limits for planar cracks are significantly higher than those for linear cracks. This differentiated weighting allocation based on crack type and geometric characteristics aligns more closely with the multi-scale mechanisms governing pavement structural performance degradation. Research indicates that the propagation of planar cracks is often directly correlated with fatigue damage in the base course material, warranting a significantly higher weighting than linear cracks [30]. In contrast, longitudinal cracks’ impact is largely localized. This weighting differential enables the final assessment to more accurately reflect the pavement’s overall health status.

(2)

Maximum deduction limit: To prevent an excessively low PCI score for a survey unit arising from extreme damage in a localized section (such as a severe networked crack), which could obscure the overall road condition and mislead maintenance decisions, this study has established cumulative deduction limits for each crack category. This design ensures the system identifies sections with multiple, widely distributed defects rather than overreacting to a single severe defect. It guarantees that the scoring results reflect major issues while still providing a balanced representation of the overall deterioration of the section, thereby providing a basis for the scientific allocation of maintenance resources.

(3)

Optimized PCI Formula: Traditional PCI calculation simply subtracts the total penalty points for various cracks from the maximum score. This study has refined the approach, resulting in the following final PCI formula:

$$PCI=100-({\displaystyle \sum _{i=1}^n\min (C{D}_i,C{L}_i)})$$

(6)

In which:

CD_i represents the deduction value obtained by consulting the deduction criteria (Table 3, Table 4, Table 5 and Table 6) for the i-th type of crack based on its adjusted effective value (see Table 7).

CL_i represents the maximum cumulative deduction for this category of cracks;

min(CD_i, CL_i) Ensure the upper limit of the final deduction system.

The core input of the formula is derived through a refined model that integrates geometric feature adjustment factors: this reflects the “severity” of these cracks (width, length proportion), achieving a transition from “quantitative” to “qualitative + quantitative” assessment. This significantly enhances the scientific rigor and precision of the evaluation.

4. Image Acquisition Methods and Image Library Construction

4.1. Data Acquisition and Visualization Process

To ensure the quality of system input data and the feasibility of subsequent processing, this study designed a standardized workflow for unmanned aerial vehicle data acquisition and processing, ultimately presenting evaluation results through intuitive visualization.

(1) Data Acquisition Experiment

This study employed the DJI Mini 3 drone (Shenzhen, China) as the data acquisition platform (Figure 5), conducting experiments on two sections of asphalt pavement exhibiting distinct road surface characteristics (Figure 6). The flight mission was meticulously designed with the following key parameters: Flight altitude was set at 7 m, with a flight speed of 0.75 m/s. Video resolution was maximized at 3840 × 2160 pixels, featuring an aperture of f/1.7, an equivalent focal length of 24 mm, and a field of view of 82.1°. These parameters were chosen to achieve sufficient ground resolution for identifying minute cracks while ensuring image clarity and freedom from motion blur, and to provide adequate overlap for subsequent image stitching.

Figure 6 presents the satellite aerial image of the road section under investigation (In Figure 6a, segment AB represents test section A; in Figure 6b, segment CD represents test section B), with an original resolution of 3840 × 2160 pixels, corresponding to actual physical dimensions of 1.35 m (length) × 0.76 m (width). The assessment area for this study is at single-lane scale, measuring 200 m in length and 7.5 m in width, encompassing the typical distribution range of road defects within this section.

(2) Image stitching and crack detection

Upon acquiring road surface video footage, key frames are first extracted at fixed intervals (e.g., every 10 frames) to eliminate redundant frames while covering the entire assessment section. Subsequently, an image stitching algorithm based on feature point matching synthesizes these key frames into a complete orthophoto of the road section (Figure 7). This panoramic image forms the foundation for subsequent automated analysis. Finally, the stitched panoramic image is input into the pre-trained DH-DWR-YOLOv11l-seg model to complete the identification, classification, and contour segmentation of all cracks, automatically extracting their geometric parameters.

(3) Visualization of Results

To achieve an at-a-glance evaluation, this study combined the calculated PCI scores with spatial location data to generate a color-coded evaluation mosaic panorama (Figure 8). Following industry convention, the system maps PCI score ranges to five color categories: 90–100 (Excellent) corresponds to dark green, 80–90 (Good) to light green, 70–80 (Average) to yellow, 60–70 (Poor) to orange, and 0–60 (Very Poor) to red.

This visualization approach has revolutionized traditional decision-making models reliant on paper reports. Maintenance managers can now directly identify deteriorating road sections (red/orange segments) from the map without interpreting complex data, enabling swift formulation of precise maintenance plans. For instance, red sections receive priority repairs while green sections undergo routine inspections, significantly enhancing the efficiency and scientific rigor of maintenance decisions.

4.2. Asphalt Section Comparative Test Setup

To validate the effectiveness and applicability of the proposed system, this study selected two asphalt pavement sections with distinct road condition characteristics as test sections (as shown in Figure 6) for comparative analysis.

Test Section A (117°8′30.145 E, 34°13′1.502 N): This section exhibits prolonged service life with multiple crack types visible, including pronounced transverse cracks, longitudinal cracks, and localized network cracking. Its relatively poor condition provides an ideal sample for verifying the discriminatory capability of the PCI evaluation model.

Test Section B (117°12′2.704 E, 34°13′8.443 N): This section comprises newly constructed or recently maintained pavement in overall good condition, exhibiting only minor fine cracks. It serves to test the system’s identification accuracy and evaluation stability under favorable road conditions.

The assessment areas for both sections comprise standard lanes measuring 200 m in length and 7.5 m in width. Data collection utilized the DJI Mini 3 unmanned aerial vehicle platform, with specific parameters as detailed previously. The flight altitude was set at 7 m to balance image resolution with capture efficiency, ensuring clear identification of cracks exceeding 1 mm in width. The flight speed of 0.75 m per second, combined with high-resolution video recording (3840 × 2160 pixels), ensured sufficient overlap between consecutive frames, laying the foundation for subsequent high-quality image stitching.

4.3. System Experimental Procedure

This study developed and implemented an automated pavement evaluation system spanning data acquisition to maintenance decision support. The system’s overall workflow constitutes a tightly integrated closed-loop process, designed to achieve full automation from asphalt pavement crack identification and analysis through to evaluation and visualization. The experimental steps correspond to system modules as follows:

• System Initialization and Data Acquisition: The experiment commenced with system initialization and data collection. Two test sections exhibiting distinct road condition characteristics were selected. UAV imaging parameters were calibrated to ensure acquisition of high-quality pavement video data suitable for subsequent stitching and recognition. Operations utilized the DJI Mini 3 UAV platform during clear weather with favorable lighting conditions. The drone was manually controlled to maintain a constant altitude of 7 m directly above the test section, flying at a uniform speed of 0.75 m/s along the centerline of the carriageway. This captured high-definition video covering the entire 200-metre assessment area.
• Data Processing and Crack Analysis Core Module: The captured video data enters the data processing and crack analysis core module. First, key frames are extracted at fixed intervals to eliminate redundant frames while covering the entire assessment section. Subsequently, these key frames are automatically stitched into a complete orthophoto panorama of the road section using feature point matching algorithms. We tested both SIFT and ORB; ORB was ultimately adopted due to its higher computational efficiency and sufficient accuracy for UAV-based image sequences, whereas SIFT produced comparable stitching quality but required approximately 30% longer processing time. This panorama forms the foundation for subsequent automated analysis. The stitched panorama is then input into an optimally trained YOLOv11l-seg model, which automatically identifies, classifies, and performs instance segmentation for all cracks while extracting their geometric parameters.
• Upon obtaining reliable crack parameters, the process proceeds to the automated PCI calculation and evaluation stage. The system automatically assigns base penalty values to cracks based on their type and severity according to a predefined evaluation table, applying geometry-based adjustment factors for refinement. Finally, an optimized PCI calculation formula synthesizes all crack information to derive the final PCI score for the study section.
• Result Visualization and Decision Support: The system’s final output comprises result visualization and decision support. Firstly, it stitches key frame images of identified cracks into a panoramic view of the road section. This is overlaid with a visualized evaluation map color-coding PCI scores and crack distribution, providing an intuitive overview of overall road condition and localized deterioration zones. Concurrently, it generates detailed assessment reports for the tested road sections, providing maintenance departments with intuitive, scientifically grounded evidence for formulating precise and efficient maintenance decisions.

Through this standardized system workflow and experimental methodology, this study has successfully established a comprehensive, reproducible technical framework, laying an empirical foundation for a closed-loop system spanning from data acquisition to intelligent evaluation.

5. Results and Discussion

5.1. Verification of Recognition and Parameter Extraction Accuracy

To quantitatively assess the reliability of crack geometric parameters extracted using the YOLOv11l-seg model, this study conducted a detailed comparison between manual field measurements of 20 linear cracks and the model’s automated identification results. Sample data for selected cases is presented in

Table 8. Comparison of Sample Crack Accuracy Verification.

Number	Severity	Crack Type	Max Length Measured (m)	Max Length Predicted (m)	Average Width Measured (m)	Average Width Predicted (m)
01	moderate	longitudinal crack	12.50	12.6345	0.0185	0.0203
02	Moderate	longitudinal crack	5.50	5.7107	0.0214	0.0196
03	mild	longitudinal crack	4.00	4.0533	0.0164	0.0143
04	mild	longitudinal crack	4.00	4.0827	0.0224	0.0245
05	moderate	longitudinal crack	5.50	5.5590	0.0226	0.0205
06	moderate	longitudinal crack	4.50	4.4656	0.0263	0.0285
07	moderate	longitudinal crack	6.00	5.9048	0.0339	0.0360
08	moderate	longitudinal crack	7.00	6.8935	0.0291	0.0270
09	moderate	longitudinal crack	6.00	5.7365	0.0227	0.0248
10	moderate	longitudinal crack	4.50	4.2512	0.0446	0.0425
11	severe	transverse crack	8.50	8.3078	0.0269	0.0246
12	severe	transverse crack	7.50	7.6094	0.0223	0.0245
13	moderate	transverse crack	8.00	8.0365	0.0245	0.0222
14	moderate	transverse crack	10.50	10.5483	0.0199	0.0224
15	severe	transverse crack	8.50	8.4411	0.0518	0.0496
16	moderate	transverse crack	2.50	2.6465	0.0172	0.0194
17	severe	transverse crack	6.50	6.4865	0.0314	0.0289
18	severe	transverse crack	8.00	7.9802	0.0287	0.0312
19	mild	transverse crack	2.00	1.9094	0.0498	0.0473
20	mild	transverse crack	2.00	1.9587	0.0358	0.0381
21	mild	map cracking	External rectangle: 35 mm × 35 mm		Model-predicted area: 1340.05 mm2
22	mild	map cracking	External rectangle: 85 mm × 50 mm		Model-predicted area: 4164.43 mm2
23	mild	Irregular crack	External rectangle: 30 mm × 30 mm		Model-predicted area: 897.68 mm2
24	mild	Irregular crack	External rectangle: 70 mm × 40 mm		Model-predicted area: 2765.32 mm2

MAE/Note: (1) Longitudinal crack max length relative error range: 0.764–5.529%, mostly between 1% and 2%. (2) Transverse crack max length relative error range: 0.208–5.86%, mostly below 2%. (3) Longitudinal crack average width MAE: 2.1 mm. (4) Transverse crack average width MAE: 2.6 mm. Note for Samples 21–24: The bounding rectangle area is the area of the minimum enclosing rectangle of the crack. The ratio of model-predicted area to bounding rectangle area serves as a morphological fitting index, which is not used as ground-truth validation for area extraction accuracy.

. For the four planar cracks (Samples 21–24), due to the absence of manually measured physical areas as ground truth, this comparison is not intended as an accuracy validation. Instead, it is redefined as a morphological fitting index, calculated as the ratio of the model-predicted area to the bounding rectangle area, which reflects the regularity of the crack shape rather than the absolute accuracy of area extraction.

To provide a more intuitive representation of the model’s error fluctuations in length and width extraction, the mean absolute error (MAE) was employed as the core metric for precision calculation. An error curve analysis diagram was plotted:

Length Extraction Accuracy: The relative error ranges for longitudinal and transverse crack lengths are concentrated within 1–2% and below 2%, respectively. As observed in Figure 9 (Linear Crack Length Error Trend Analysis Chart), the model’s predicted crack lengths closely align with actual measurements. Errors are consistently maintained at low levels with minimal fluctuation, demonstrating the model’s high stability and accuracy in crack length extraction.

Width Extraction Accuracy: As depicted in Figure 10 (Linear Crack Width Error Curve Analysis), the width extraction error is marginally greater than that for length. However, its Mean Absolute Error (MAE) values (2.1 mm for longitudinal cracks and 2.6 mm for transverse cracks) remain within the permissible range for engineering applications. The fluctuation in error primarily stems from pixel-level deviations inherent to image resolution limitations and the subjective nature of measurement point selection during manual surveys.

Error analysis indicates that minor discrepancies primarily stem from: (1) pixel-level errors due to image resolution limitations; (2) minor pixel deviations at model segmentation boundaries; (3) measurement inaccuracies when manually tracing cracks that cannot perfectly follow their natural curvature; (4) inherent width variations in cracks and subjective measurement point selection.

In summary, despite these unavoidable errors, the Mean Absolute Error (MAE) for all key parameters remains within an acceptable range for engineering applications. Accuracy validation confirms that the crack geometric parameters extracted using the YOLOv11l-seg optimized model are reliable and precise, establishing a robust data foundation for subsequent automated PCI calculation and evaluation. Verification of the accuracy in extracting the area of planar cracks will be analyzed in conjunction with specific data within the subsequent Section 5. Findings indicate the model demonstrates high accuracy in extracting crack geometric parameters.

5.2. PCI Evaluation Results and Comparative Analysis

To verify the scientificity and superiority of the optimized PCI model, this study took the road section A (with poor road conditions) as an example and conducted a detailed comparison of the evaluation results between the optimized model and the traditional PCI model.

(1) Refinement and rationalization of evaluation results

Figure 11 visually demonstrates the significant differences in crack statistics for the same road section between the two models. In the traditional model, all 197 transverse cracks and 148 longitudinal cracks were simply classified as “minor”. However, the optimized model, by introducing geometric feature adjustment factors based on width and length, achieved a refined distinction of crack severity. The results show that only 12% of the transverse cracks were rated as minor, while 35% and 54% were classified as moderate and severe, respectively; for longitudinal cracks, 35% and 4% were classified as moderate and severe, respectively. This distribution more accurately reflects the actual severity of cracks on this road section, proving a significant improvement in qualitative accuracy of the optimized model. This refined classification directly leads to a total deduction value closer to the actual road conditions, making the final PCI score more distinguishable and valuable for engineering reference.

(2) Scientific Adjustment of Explicit Weights

The optimized model’s refinement of explicit weights is clearly illustrated in the radar chart of Figure 12. In the traditional model, all crack types were assigned a weight of 1, failing to reflect the differential impact of various crack types on pavement performance. Following optimization, the model assigns differentiated weight ranges to distinct crack types: network cracks (1.2–1.4) and irregular cracks (1.3–1.5) receive the highest weights, longitudinal cracks (1.1–1.2) follow, while transverse cracks (1.0–1.2) carry the lowest relative weight. This weighting distribution aligns entirely with pavement engineering principles: network and irregular cracks typically signify reduced structural bearing capacity and severe material deterioration, posing the greatest threat to pavement service life and thus warranting the highest deduction weights.

(3) Comprehensive Incorporation of Implicit Weights

Beyond explicit weights, the optimized model introduces multiple implicit weights through its computational logic:

Size Weighting: Adjustable factors enable the system to automatically assign higher effective values to longer and wider cracks, achieving a nuanced assessment where same type and grade, but differing sizes yield differing impacts;

Morphological weighting: By adopting the “damage area ratio” as the core metric for planar cracks and setting higher deduction ceilings, this inherently assigns greater morphological weighting than linear cracks; severity weighting: within the same crack type, the deduction values and ceilings for severe grades far exceed those for minor grades, constituting an exponential weighting increase.

The optimized PCI model integrates explicit and implicit weights to construct a multidimensional, multi-tiered comprehensive evaluation system. This framework not only quantifies crack “quantity” but also deepens into a holistic assessment of crack quality. Ultimately, the conventional PCI model assigned a score of 92.6 to the test section A, whereas the optimized model yielded a score of 77.8. It should be noted that no independent PCI survey following ASTM D6433 was conducted for this test section. Therefore, the superiority of the optimized model is argued based on: (i) its ability to differentiate crack severities (Figure 13), (ii) the more reasonable score distribution (Figure 14), and (iii) the visual plausibility that a section with numerous severe cracks (Figure 15a) should not receive an “excellent” rating of 92.6. The optimized score of 77.8 (“moderate”) is far more consistent with the observed visual condition. The latter more accurately revealed the presence of significant structural damage in this section, closely aligning with the actual pavement condition. This provides precise and scientific grounds for formulating maintenance decisions prioritizing repairs.

5.3. Recognition and Parameter Extraction Results

The crack identification and parameter extraction module of this system demonstrated outstanding effectiveness and stability across both test road sections. Figure 13 presents a panoramic image featuring image stitching and overlaid model prediction boxes, providing the most intuitive visual evidence for road condition assessment.

Figure 13a clearly shows that the road section A under examination is covered with numerous successfully identified cracks, with densely clustered prediction boxes. Among these are several long and wide severe cracks, visually indicating a concerning road condition. In stark contrast, Figure 13b depicts road section B under examination, where the road surface appears largely intact. The identified cracks are fewer in number and exhibit minor forms, intuitively reflecting its favorable road condition.

To enable quantitative comparison of road conditions, the system automatically counted and categorized all cracks across both sections. The results are presented in Figure 14.

The statistical findings for test section A (Figure 14a) corroborate its poor road condition: not only is the total number of cracks high, but more significantly, the proportion of cracks classified as severe is exceptionally high. Among the 197 transverse cracks, medium and severe grades collectively accounted for 89% (35% + 54%); similarly, medium and severe grades constituted 39% of longitudinal cracks. This indicates that the damage to this section extends beyond superficial phenomena, deeply affecting the pavement structure and necessitating prompt maintenance measures.

Statistical results for Section B (Figure 14b) reveal markedly different characteristics: while cracks are present, their severity predominantly falls within mild and moderate categories. No severe cracks were identified in either transverse or longitudinal cracks, with moderate cracks accounting for 70% and 69% respectively. This distribution of defects indicates that this section is in the early to mid-stages of performance degradation, making it suitable for preventive maintenance planning rather than requiring immediate large-scale repairs.

In summary, the identification and statistical results not only visually demonstrate the differences in road condition between the two sections but, more importantly, provide detailed and reliable data inputs for the optimized PCI evaluation model developed in this study. These automatically extracted, refined parameters form a robust foundation for generating subsequent scientific evaluation conclusions and visualization outputs.

5.4. Visualization of Results and Decision Support

The system’s final output—a color-coded PCI assessment panorama—integrates all preceding analytical findings into a single, readily comprehensible “road condition diagnostic map”. Figure 15 presents the final visualization results for the two test road sections.

Each colored unit corresponds to a keyframe image (with approximately 30% overlap between adjacent units)

This visualization approach divides continuous road sections into multiple contiguous assessment units. The system calculates an independent PCI score for each unit based on the total penalty points from all cracks within it. These scores are then colored according to predefined mapping rules (Excellent—dark green, Good—light green, Fair—yellow, Poor—orange, Very Poor—red), ultimately assembling into a comprehensive full-color visual spectrum of the entire road section.

The results for Figure 15a (Road Section A under assessment) align closely with the identification statistics in Figure 13a and Figure 14a, clearly visualizing the road condition: numerous orange and yellow units indicate areas where the pavement has deteriorated to a “Poor” or “Fair” level, signifying critical zones with severe structural damage requiring immediate priority repair. Concurrently, the presence of light green units in certain sections indicates that these segments warrant heightened attention.

Conversely, Figure 15b (Section B under assessment) displays an overall healthy green coloration, with only scattered yellow units in localized areas. This aligns perfectly with the statistical findings in Figure 13b (indicating sparse cracks) and Figure 14b (showing cracks predominantly of light to moderate severity). This visualization provides an intuitive maintenance recommendation: “No immediate intervention required; maintain routine inspections”.

This visualization approach fundamentally transforms traditional decision-making reliant on numerical reports. Maintenance managers can now grasp the macro-level health status of entire road sections within seconds, precisely pinpointing specific degraded segments, without needing to interpret complex crack statistics tables or PCI score lists. It converts scientific data into intuitive actionable guidance, significantly enhancing the efficiency and scientific rigor of maintenance decisions, thereby achieving a true closed-loop transition from “intelligent identification” to “intelligent decision-making”.

Time efficiency comparison: To quantify the time savings of the proposed automated system, we compared the total data acquisition and processing time with that of conventional manual crack measurement for the same 200 m test section. Manual inspection (including on-site walking, crack identification, length/width measurement, and paper recording) typically takes about 90 min by an experienced engineer. In contrast, our UAV-based system requires approximately 12 min for drone flight (including takeoff, landing, and data transfer) and 8 min for automated post-processing (keyframe extraction, stitching, crack segmentation, PCI calculation, and visualization), totaling 20 min. This corresponds to a time reduction of about 78% (70 min saved per 200 m section). For longer road networks, the efficiency gain would be even more substantial due to the automated batch processing capability.

A rough cost estimate indicates that the one-time hardware investment (UAV + workstation) is about 1550U SD, while the per-kilometer operational cost is approximately 5, compared to $22.5 for manual inspection, representing a cost reduction of over 75%.

6. Conclusions

This study focuses on establishing a complete technological closed-loop system, progressing from automated identification to intelligent evaluation. It successfully constructs and validates an integrated system for intelligent crack detection and automated PCI assessment in asphalt pavement maintenance. By utilizing drone-based field data acquisition, the research implemented a fully automated workflow encompassing video frame extraction, image stitching, crack identification and parameter extraction, automated PCI calculation, and result visualization. This effectively addresses the core issues of inefficient traditional methods and the disconnect between “identification” and “evaluation” in existing research.

The primary contributions of this research are threefold:

(1)

Construction and validation of an end-to-end integrated system: Overcoming the limitation that current research often stops at “identification”, this study developed an end-to-end system Centered on an optimized YOLOv11l-seg model. This system achieves unmanned processing from raw UAV video to final PCI scores and visualized panoramic maps, providing an efficient, engineered solution for pavement maintenance management.

(2)

Introduction of a refined PCI evaluation model: Addressing the inadequate differentiation of crack types and severity in traditional PCI models, this study proposes an improved, more granular evaluation framework. By incorporating geometry-based adjustment factors and differentiated deduction weights with upper limits, the model scientifically quantifies the actual impact of cracks of varying types, dimensions, and severity on pavement performance. This significantly enhances the engineering rationality and decision-making value of evaluation outcomes.

(3)

Innovative Visualization and Decision Support: The system generates color-coded PCI panoramic maps, transforming complex multidimensional data into intuitive “road condition diagnostic diagrams”. This output format significantly enhances result interpretability, enabling maintenance departments to pinpoint deteriorated sections at a glance and formulate precise maintenance strategies, achieving an intelligent transition from data to decision-making.

Although this study has yielded anticipated outcomes, the system retains scope for optimization.

Sample size limitation: The accuracy verification of crack geometric parameters was based on 20 crack samples. While this sample size is acceptable for a proof-of-concept study, it limits the statistical generalizability of the results. Broader validation across diverse pavement textures, lighting conditions (e.g., low light, strong shadows, glare), and crack morphologies (e.g., finer or more complex patterns) is necessary before the system can be deployed at scale. Future work should therefore include larger and more varied datasets to further substantiate the robustness of the proposed approach.

Current models require improved accuracy in identifying low-contrast cracks (e.g., light-colored, minute fissures, fatigue cracks), this represents not only an algorithmic challenge but is intrinsically linked to microstructural evolution in ageing asphalt materials, such as changes in micelle size and surface activity. A limitation of this study is the absence of an independent ground-truth PCI measurement obtained by trained engineers following ASTM D6433. The validation of the optimized PCI model therefore relies on internal comparison and visual plausibility. Future research should incorporate formal field surveys to further substantiate the model’s accuracy. Future research will focus on:

(1)

Incorporating multi-source data (e.g., vehicle-mounted LiDAR point clouds) to construct richer multi-modal datasets, enhancing model generalization capabilities;

(2)

Expanding defect types and integrating mechanisms: Future systems must extend to identifying and assessing structural defects like reflective cracking. Drawing from research on crack-resistant base course material design and microstructural characterization to understand crack-resistance mechanisms will enhance the predictive capability of assessment models regarding crack progression;

(3)

Deepening multi-scale assessment models: The ultimate goal is to construct an intelligent platform encompassing “perception-cognition-decision-making”. To this end, assessment models must be refined by integrating macro-level crack data with micro-level material models, such as asphalt ageing states and crack-resistant properties of base materials. Through multi-scale information linkage, this approach achieves a cognitive leap from identifying “where cracks occur” to understanding “why and how cracks develop”, thereby providing scientific grounds for lifecycle-based preventive maintenance decisions.