South Africa has a road network spanning around 750 000 km, making it the longest among all African nations and ranking it as the seventh longest globally. The South African National Roads Agency Limited (SANRAL) is responsible for the management of the country’s national road infrastructure, which encompasses a vast network spanning approximately 22 197 km of paved highways (Sanral 2024). Provincial governments are responsible for over 220 000 km, while the municipal network is approximately 275 000 km. The CoJ Metropolitan Municipality has about 13 599 km of roads, with over 91% which are surfaced and 9% unsurfaced (Hancock 2022). According to the 2022 CoJ Metropolitan Municipality annual report, 45% of the road network is in excellent condition, 23% is poor and 32% is extremely poor (City of Johannesburg 2022). Potholes, structural cracks, faulty streetlights, clogged storm water drains, neglected walkways and traffic islands are signs of poor road conditions. As a result of storm-water drain obstruction, water flows along heavily used streets, making them unusable by vehicles and forming new potholes and expanding existing ones (CSIR 2023). The CoJ Metropolitan Municipality has fixed up to 100 000 potholes in 9 months in 2023 and road users report at least 1000 potholes every day (Peters-Scheepers 2023).

The constant and essential process of maintaining road infrastructure is crucial to retaining the quality, safety and operation of road surfaces. Afridi, Erlingsson and Sjögren (2023) suggest that to ensure road assets are preserved and properly maintained, government departments should carry out regular road condition assessments. Chen et al. (2022) further recommend adopting proactive strategies to mitigate the deterioration of road infrastructure caused by various factors such as ageing and usage over time. Road condition assessments are used to perform a comprehensive evaluation of the road surface to detect the presence of cracks, potholes, rutting and signs of degradation (Pavel, Tan & Abdullah 2022). The assessment of road conditions using traditional methodologies is carried out by the utilisation of manual tools, accurate measuring devices and visual inspections (Pavel et al. 2022). According to Pavel et al. (2022), the process of visual inspection requires experts to evaluate the state of road infrastructure components by using data obtained via visual inspections. This process is fundamentally subjective, as it mostly relies on the evaluative opinions and expertise of the inspectors (Ren et al. 2022). Various inspectors with diverse backgrounds and perspectives may potentially perceive occurrences in different ways, thus resulting in disparities or variations in their assessments (Jia et al. 2023). According to Ahmad et al. (2024), relying just on visual inspections may be inadequate in detecting structural or subsurface issues that might potentially compromise the overall integrity of the highway. Visual inspections are occasionally limited to superficial analysis and may not sufficiently detect concealed flaws (Afridi et al. 2023). The visual inspection evaluation of road conditions encompasses several methodologies such as the pavement condition index (PCI), the international roughness index (IRI), functional categorisation and network-level assessment (Ren et al. 2022).

Although the PCI is a valuable tool for evaluating pavement conditions, Hancock (2022) remarks it does have several drawbacks. Hancock (2022) states that PCI is not suitable as an engineering tool for local governments when making choices regarding pavement management because of its limitations. More precisely, when a PCI is created using data from condition surveys, Lendra et al. (2023) outline that a significant amount of crucial engineering information is omitted, specifically data related to cracking. One significant drawback of PCI is that roadway segments might exhibit comparable or identical PCI values while having distinct types of distress (Lendra et al. 2023). Lin et al. (2022) are of the view that performing PCI surveys necessitates a substantial amount of time and effort, particularly for extensive networks.

While IRI is a widely used index for indicating road surface roughness, there are some limitations (Lin et al. 2022). The measurement of IRI values has traditionally relied on a manual method, which presents several drawbacks such as time-consuming procedures, traffic disruption and relatively high costs (Wintruff & Fernandes 2023). Pinatt et al. (2020) present a view that manual methods may be cost-effective but can be time-consuming and subject to user bias. Moolla and Tetley (2021) state that IRI tools provide objective deflection data, but require specialised equipment and may cause disruptions in traffic flow. Moolla and Tetley (2021) also state that during the IRI assessment, inaccuracy in the number of vehicles’ wheels can affect the accuracy of IRI values. The relationship between IRI and other measures of roughness and riding quality is not always clear (Lin et al. 2022). Lin et al. (2022) highlight that the inspection of surface roughness may not detect underlying structural problems. While passenger comfort is important, the methodology lacks a comprehensive assessment of other aspects related to road conditions, such as road surface degradation, safety issues and structural soundness (Afridi et al. 2023). Ahmed, Kays & Sadri (2023) say that IRI offers a level of stability that guarantees that trends and variations in pavement roughness may be precisely monitored and compared on an annual basis. As a result of its inherent stability, Ahmed et al. (2023) say that the IRI is a dependable tool for forecasting future pavement conditions. Jia et al. (2023) add that it facilitates the planning of maintenance and rehabilitation operations. International roughness index is a universally acknowledged and globally utilised standardised approach for quantifying road roughness (Wintruff & Fernandes 2023).

According to Starke and Geiger (2022), network-level assessment involves a certain degree of subjectivity when estimating distress amounts. Ren et al. (2022) say that if the inspectors are not well qualified, there is a higher chance of having more variability in the results compared to if distress quantities had been directly measured. At the network level, the significance lies in matters pertaining to the amount and quality of data (Afridi et al. 2023). The measurement of data quantity, in terms of both the type of data being measured and the amount being measured, has consequences for both time and cost (Afridi et al. 2023). Typically, the cost of data collecting increases as the volume or level of detail of the acquired data increases (Pereira & Vieira 2022). Simultaneously, Wintruff and Fernandes (2023) are of the view that the availability of additional or more comprehensive data for analysis would lead to improved decision-making.

Although road network-level assessment offers numerous benefits, it also presents certain drawbacks and challenges (Afridi et al. 2023). Chen et al. (2022) state that performing thorough evaluations of extensive road networks can incur high costs; Afridi et al. (2023) add that it necessitates the use of substantial financial resources for the procurement of advanced technology, equipment and human resources. Jia et al. (2023) add that processing and evaluating the vast quantities of data gathered during assessments can be intricate and demanding in terms of resources. According to Lin et al. (2022), network-level assessment requires advanced data management systems and experienced workers. Wintruff and Fernandes (2023) state that maintaining the precision and uniformity of data throughout a vast network might pose difficulties. Erroneous or contradictory data can result in incorrect decisions and unsuccessful solutions (Lin et al. 2022). Pinatt et al. (2020) found that the evaluation procedure, particularly when it entails physical inspections, might result in temporary interruptions to traffic flow and local populations. The efficiency of road network-level assessments can be hindered by the scarcity of competent persons and innovative technology in certain places (Deni 2020). Lin et al. (2022) add that the process of incorporating new evaluation data into current transport management systems and databases can be difficult and may necessitate extra resources and time.

Machine vision is an automated system that uses optical devices and non-contact sensors to capture and analyse images of real objects (Schattler, Wolters & Zimmerman 2020). Machine vision can perform a diverse range of tasks that human vision is unable to handle (Moolla & Tetley 2021). Schattler et al. (2020) outline that performing capturing of images is achieved through the utilisation of advanced vision sensors, cleverly designed optical transmission systems and powerful image processing algorithms (Figure 1). In the domain of machine vision, images function as the conveyors of information. Image processing and evaluation are crucial technologies used in vision detection systems to automatically comprehend received images (Moolla & Tetley 2021). The use of computer vision technology allows for the creation of machine vision, which is used to detect surface objects.

Within industrial applications, there are three distinct defect detection tasks that rely on machine vision: classification, localisation and segmentation (Kitsios & Kamariotou 2021). Kitsios and Kamariotou (2021) argue that in most cases, defect classification is utilised to determine whether a particular fault is present in a picture (Cazzato et al. 2020). Cazzato et al. (2020) state that information that specifies the features of the target is extracted from the image pixels through the process of feature extraction. After that, the differences between the various targets are mapped to a feature space with a reduced dimension. For determining the categories of defects that are present in an image, it is critical that the features that are picked to describe the image distinguish between different types of images (Kitsios & Kamariotou 2021). The fundamental objective of defect classification is to train the classifier in accordance with the extracted feature set and then to ensure that it correctly identifies the classification of surface defects for each surface defect using either supervised or unsupervised pattern recognition techniques. Defect localisation involves the use of imaging techniques and algorithms to accurately locate and detect defects in objects. In the specific context of a picture, fault localisation necessitates accurately identifying the defect’s position and categorising it accordingly (Cazzato et al. 2020). Defect localisation is the process of identifying all the specific locations within the product where defects are present. Furthermore, significant advancements have been achieved in defect localisation by utilising Convolutional Neural Networks (CNN)-based image classification techniques and deep learning-based object detection technology (Abdollahi et al. 2020). Image segmentation is a method that divides images into different sections that are distinct from one another for the purpose of identifying objects of interest with the main objective of predicting categories that each pixel in an image belongs to (Cazzato et al. 2020). Regional-based, edge-based and specialised theory-based approaches are used for the process of segmentation.

Ethical approval to conduct this study was obtained from the Tshwane University of Technology Research Ethics Committee (No. REC2024=04=039(MS)). Strict ethical considerations were adhered to throughout all stages of the research. Ethical clearance was obtained from the CoJ before data collection commenced. All participants were informed of the purpose of the study, their voluntary participation and their right to withdraw at any time without consequence. Informed consent was obtained prior to administering questionnaires, interviews and focus groups. To protect confidentiality, participants’ identities and responses were anonymised, and all collected data were stored securely and used solely for academic purposes.

Table 1 represents the demographics of research participants within the seven regions of the CoJ Metropolitan Municipality.

The demographic profile presented in Table 1 indicates distinct distributions within gender and racial categories for a specified population. Gender distribution is characterised by a predominant proportion of males, constituting 64.2%, while females represent 34.6% and an additional 1.2% of individuals identify as other, suggesting recognition of diverse gender identities. In terms of racial composition, most of the population is black, accounting for 87.7%, followed by coloured individuals at 8.6% and Indian/Asian at 3.7%. These data highlight a demographic landscape where black individuals significantly outnumber other racial groups, with meaningful representation from other racial backgrounds.

The occupational profile in Table 2 presented a distribution across different organisational levels. Executive managers, comprising chief executive officer (CEOs), chief financial officers (CFOs), chief operating officers (COOs) and heads of departments (HODs), represent 0.85% of the sample, reflecting a small, but strategically significant segment of leadership. Senior managers and directors account for 2.48%, while managers and deputy directors, predominantly registered professional engineers, constitute 17.56% of the participants. The largest proportion of the sample consists of skilled employees, including administrators, artisans and technicians, who make up 79.11% of the population.

Figure 3 illustrates the geographic demarcation of the seven administrative regions of the CoJ, labelled A through G. These regions represent spatial zones within the metropolitan boundary, each encompassing a range of suburbs, townships and industrial areas with differing socio-economic characteristics.

Notably, potholes predominate in Regions A, E and G, constituting 80.0%, 60.0% and 75.0% of the reported damages, respectively, which suggests significant road infrastructure distress. Surface cracks are predominantly observed in Regions B and C, accounting for 83.3% and 54.5% of damages, respectively, indicative of potential underlying issues such as material failure or subgrade instability. Other categories of damage, including worn lane markings and water flooding, appear less frequently yet are evident in Regions C through E, highlighting concerns related to visibility and drainage efficiency. In addition, uneven surfaces are specifically prevalent in Region F, representing 13.3% of the reported damages, potentially because of inadequate construction practices or the impact of heavy vehicular traffic. These observations underscore the necessity for region-specific road maintenance and rehabilitation strategies that directly address the predominant forms of damage in each area, thereby enhancing the overall safety and durability of road infrastructure.

Figure 4 represents common types of road damage in seven regions of CoJ indicating the distribution of various types of road damage across regions labelled A through G, indicating substantial regional discrepancies.

Statistical analyses were used to evaluate the validity, reliability and overall robustness of the study’s findings. The Chi-square analysis validated that the observed associations among variables were statistically significant and improbable to have arisen by chance. The Kaiser–Meyer–Olkin (KMO) measure and Bartlett’s Test of Sphericity were used to confirm the sample adequacy and inter-variable correlations that justified advanced statistical analysis. The analysis of variance (ANOVA) inside the regression model demonstrated that the implementation of artificial intelligence (AI) technologies considerably and consistently influenced the quality of road condition evaluations. Table 3 represents the results of a Chi-square Test that was conducted to analyse the frequency distribution of road damage across different regions. The test aims to determine if there are statistically significant differences in the types of road damage reported in the various regions of the CoJ.

The analytical findings from the Chi-square Test reveal significant variations in the types of road damage across different regions, with a p-value of 0.004. This statistical significance highlights distinct regional challenges, suggesting that environmental, administrative or material quality differences could be influencing the prevalence and types of road damage observed. Table 4 represents KMO test for suitability factor analysis.

The application of the KMO and Bartlett’s Tests to evaluate the dataset’s suitability for factor analysis yielded a KMO value of 0.627. This result not only confirms the adequacy and acceptability of the sample size but also indicates that the data are appropriate for detecting underlying patterns between variables. These analyses collectively provide a robust statistical framework to further explore and understand the dynamics of road damage, potentially guiding targeted interventions and resource allocation for infrastructure maintenance and improvement.

Table 5 (ANOVA) evaluates the effectiveness of a regression model in explaining variations in road condition assessments, using predictors such as Robots, ESNLP, machine learning and machine vision.

Table 5 shows a Regression Sum of Squares at 37.504, indicating the variability explained by the model and a Residual Sum of Squares at 21.466, representing unexplained variability. The total variation in the dataset is 58.970. With five predictors, the model has an F-value of 107.971, suggesting a highly significant fit (p < 0.001), confirming that the predictors substantially and significantly contribute to predicting road conditions assessment, with the mean square values (7.501 for regression and 0.069 for residual) further supporting the model’s explanatory power.

The multiple regression analysis was carried out using Statistical Package for the Social Sciences (SPSS) Version 24. According to the final model, machine vision, ESNLP, and robots with beta coefficients of 0.193, 0.165 and 0.269 for X₁ (machine vision), X₂ (ESNLP) and X₃ (robots), respectively, are critical factors for road condition assessment.

The developed mathematical equation of the model is as follows (Equation 1):

Where:

RCA = Road condition assessment;

X₁ = Machine Vision (MV) with β = 0.193 and p-value < 0.05;

X₂ = Expert System and Natural Language Processing (ESNLP) with β = 0.165 and p-value < 0.05;

X₃ = Robots (R) with β = 0.269 and p-value < 0.05.

The qualitative phase highlighted critical deficiencies in the current system, including reliance on manual and subjective assessment methods, outdated tools and techniques, inconsistent record keeping, slow response times and insufficient integration of data into planning processes (Table 6). Across all qualitative data sources, potholes consistently emerged as the most significant and recurrent road damages, reported as the primary cause of poor road user experience and as a challenge that overwhelms existing maintenance capacity. These shortcomings were compounded by resource constraints such as inadequate funding, limited technical skills and shortages of skilled personnel, which hinder systematic and timely interventions. Triangulation of findings ensured validity, and key themes that emerged emphasised the need for modern, technology-driven approaches. Collectively, the study highlighted that while practitioners possess valuable experiential knowledge, there is an urgent need to move towards AI-enabled solutions to improve accuracy, efficiency and evidence-based decision-making in road condition monitoring and maintenance planning.

To enhance road maintenance operations, it is recommended that the CoJ Metropolitan Municipality should adopt advanced machine vision technologies to improve the precision, efficiency and cost-effectiveness of road inspections. Qualitative findings from interviews and focus groups strongly highlighted the over-reliance on manual, subjective assessments and highlighted the need for modern, automated systems that can provide real-time and consistent data on road conditions. In addition, the study recommends a broader implementation of preventative maintenance strategies across all regions to reduce the recurrence of severe road damage. These strategies should be supported by region-specific maintenance plans that consider the unique demographic, geographical and usage patterns of each region, as highlighted by participants who reported that certain high traffic areas deteriorate faster than others. Furthermore, the findings indicate the need for capacity building through training and awareness programmes, to equip municipal staff with digital skills and to promote understanding and acceptance of new technologies. Addressing resource constraints and improving coordination between teams were also emphasised in the qualitative results as critical steps to ensure that the adoption of advanced technologies and proactive strategies leads to sustained improvements in road maintenance operations.

Future research should broaden the geographic scope, employ larger datasets and incorporate longitudinal analyses to examine the long-term impact and cost-effectiveness of AI-driven frameworks in different municipal contexts.

Despite its potential, this methodology faces limitations such as the need for high-quality imaging infrastructure, dependence on skilled personnel for model training and interpretation and challenges in integrating data sources. Data privacy, cybersecurity risks and resistance to adopting new technologies in municipalities can also constrain the effectiveness of this framework.

Implementing a machine vision-based Road Condition Assessment framework in South African municipalities involves a substantial initial capital investment estimated between R23 million and R79 million, driven by the acquisition of high-resolution cameras, LiDAR-equipped drones, autonomous robots, computing infrastructure and AI software (Visser et al. 2022). Annual recurring costs are estimated at R8 million to R18 million, covering system maintenance, AI model retraining, software licensing, specialist staff and cybersecurity. Although the upfront costs are significant, the framework enables optimised maintenance planning, reduces emergency repairs and extends road lifespans, leading to long-term cost efficiencies and a typical return on investment within 5 years – 7 years for a large metropolitan municipality.