A Data-Driven Framework for Proactive Urban Pothole Management: Integrating Explainable AI and GIS Spatial Clustering

Urban pothole management conventionally relies on reactive, complaint-driven maintenance, which inevitably suffers from operational delays and inefficient resource distribution. To catalyze a paradigm shift toward proactive maintenance, this study proposes a comprehensive data-driven framework that integrates machine learning, explainable AI (XAI), and Geographic Information System (GIS) spatial clustering to predict localized pothole risks. Utilizing a dataset of 21,397 spatial points (13,104 occurrences and 8293 non-occurrences) collected from Gwangju Metropolitan City (2022–2024), 19 diverse spatial, environmental, and geometric variables were extracted. To ensure strict statistical rigor and mitigate data leakage, dual-track encoding and Phi-K correlation analyses were applied to capture complex non-linear dependencies. Among the six predictive models evaluated, the Random Forest algorithm demonstrated superior performance, achieving an accuracy of 90.47% and an ROC-AUC of 0.957. Subsequent XAI interpretation using Shapley Additive exPlanations (SHAP) empirically revealed that physical scale and spatial vulnerabilities synergistically accelerate pavement fatigue. Furthermore, Density-Based Spatial Clustering of Applications with Noise (DBSCAN) successfully translated these predictive probabilities into 14 highly localized hotspots. Ultimately, the findings of this study provide a highly actionable, evidence-based guideline for local governments to abandon uniform budget distribution and transition to a preemptive, concentrated allocation of municipal resources directed exclusively at high-risk zones.