Road networks are critical to economic activity and public safety, yet traditional pavement inspection remains largely manual and reactive. Visual surveys are slow, inconsistent, and costly, and many structural failures begin below the surface, invisible to the naked eye. In my survey, I synthesize recent machine-learning advances that enable faster, more objective, and predictive pavement management, providing the foundation for data-driven, proactive road maintenance.

What is reviewed

I systematically examine four interconnected areas. First, supervised detection methods, ranging from handcrafted features to modern convolutional neural networks (CNNs) and hybrid CNN–RNN architectures, which extract patterns from visual and sensor data. Second, edge and federated deployment strategies that allow in-field, privacy-preserving monitoring without centralizing sensitive data. Third, reinforcement learning and metaheuristic approaches for maintenance and rehabilitation scheduling, translating detection outputs into cost-effective action plans. Finally, I explore multimodal sensing, combining RGB imagery, thermal cameras, LiDAR, IMU, and ground-penetrating radar (GPR), with a particular focus on subsurface integration to detect latent structural issues.

Key technical takeaways

I found that multimodal fusion is essential for early detection of latent structural problems. Surface-level imagery alone may miss moisture infiltration or base-layer failures, while thermal and GPR signals provide critical depth and diagnostic insight. Edge and federated learning approaches enable inspection fleets to collaborate on model improvement without centralizing raw imagery, effectively balancing privacy, bandwidth, and scalability. Hybrid temporal models, such as CNN combined with LSTM or BiLSTM, reduce false positives on moving platforms by integrating visual and inertial data. Importantly, optimization algorithms close the loop: detected distresses, including crack density, rutting depth, or International Roughness Index (IRI) forecasts, can feed reinforcement learning or metaheuristic schedulers that optimize long-term maintenance and rehabilitation under budgetary, traffic, and environmental constraints. Emerging tools like AutoML and explainable AI further accelerate deployment and improve operator trust, offering neural architecture search for edge models and SHAP/LIME interpretability for predictive maintenance decisions.

Practical recommendations for implementers

I recommend starting with a pragmatic sensor mix. For continuous mapping, RGB cameras and IMU sensors provide robust baseline data, while thermal cameras and GPR are valuable for targeted campaigns where subsurface risk is likely. Benchmarking on the target hardware is critical, evaluating both inference latency and model accuracy on representative embedded GPUs or smartphones. Federated updates allow fleet-scale learning without centralizing raw data, improving scalability and privacy. Domain adaptation is necessary to handle geographic or material shifts, either by collecting a small local labeled set or applying transfer learning techniques. Finally, I emphasize that detection outputs should be directly integrated into maintenance and rehabilitation pipelines, ensuring that maps and predictive insights translate into actionable schedules that account for traffic, cost, and operational constraints.

Research gaps and open challenges

Despite progress, several challenges remain. Public, well-annotated, geographically diverse datasets are still scarce, limiting generalizability. Standardized runtime and on-device benchmarks are needed to evaluate edge AI solutions fairly. Developing robust and reproducible multimodal fusion pipelines, particularly involving GPR, is an active research frontier. Ethical and environmental considerations are also important, including the privacy of geotagged imagery and the carbon footprint of large-scale model training, which require careful mitigation strategies.

Impact beyond pavements

The methodologies I surveyed extend beyond asphalt roads. Similar approaches can be applied to other infrastructure domains, such as bridges and rail tracks, and even non-transport areas including agriculture and medical imaging. Establishing reproducible benchmarks and open datasets will accelerate cross-domain adoption and facilitate more resilient, data-driven infrastructure management practices.

For the full literature synthesis, unified taxonomy, benchmark proposals, and deployment checklists, see:

Asadi, Y. Paving the Future of Intelligent Pavement Defect Detection with Machine Learning: A Comprehensive Survey of Techniques and Applications. Int. J. Pavement Res. Technol. (2026). https://doi.org/10.1007/s42947-026-00711-y