<p>The automated detection of pavement distress has undergone a paradigm shift, transitioning from manual inspections to data-driven machine learning (ML) frameworks that enable proactive infrastructure management. This review synthesizes recent advancements in ML paradigms applied to pavement defect detection and maintenance planning, structured around four interconnected objectives: analyzing supervised methodologies from classical feature-based classifiers to deep convolutional neural networks (CNNs) and hybrid CNN–recurrent neural network (RNN) architectures; evaluating edge-computing and federated learning deployment strategies for real-time, distributed monitoring; surveying reinforcement learning and metaheuristic optimization techniques for predictive maintenance scheduling; and identifying critical research gaps and future directions. We present a unified taxonomy spanning sensing modalities (RGB, thermal, LiDAR, IMU, GPR) and ML model families, emphasizing multimodal fusion and domain adaptation challenges. Particular focus is given to subsurface sensing integration such as ground-penetrating radar and infrared thermography which enables early detection of latent structural degradation often invisible to surface-level imaging alone. Additionally, we highlight emerging trends in AutoML, explainable artificial intelligence (AI), and neural architecture search, which enhance model interpretability and deployment efficiency. Despite significant progress, challenges remain, including dataset scarcity, limited generalizability across geographic domains, and computational constraints in real-time edge deployment. We conclude with practical recommendations for benchmarking, open dataset development, and the integration of end-to-end pipelines that link defect detection with maintenance optimization. This survey provides a comprehensive reference for advancing intelligent pavement management systems through machine learning.</p>