Prediction of pavement maintenance quality and performance indicators using particle swarm optimized gradient boosting decision trees

The rapid global expansion of highway networks has significantly increased maintenance demand. Pavement preventive maintenance is an effective strategy for prolonging pavement lifespan and optimizing budgets. This study developed tree-based models, specifically XGBoost, LightGBM, and CatBoost, with and without Particle Swarm Optimization (PSO), to predict the Pavement Maintenance Quality Index (PQI) and four key performance indicators: Pavement Surface Condition (PCI), Riding Quality Index (RQI), Rutting Depth (RDI), and Skid Resistance (SRI). The models were trained on four years of measured data, incorporating features such as age, lane, direction, section length, average annual climate, and distress and repair area. The Shapley Additive Explanations method was used to evaluate feature importance. The results demonstrated that PSO enhanced the performance of all models, with PSO_CatBoost achieving the highest predictive accuracy. Validation on unseen data from a subsequent year and a separate highway confirmed strong model performance, yielding R² values of 0.7984 and 0.6575 for PQI and 0.9379 and 0.9 for PCI, while the mean absolute percentage error remained below 1%. The findings suggest that the PSO_CatBoost framework could offer a reliable, data-driven method for lane-specific pavement quality prediction, potentially providing maintenance authorities with a practical tool for screening and prioritising preventive interventions. Tree-based methods enhanced with particle swarm optimization are proposed for pavement maintenance quality index, surface condition, roughness, rut depth, and skid resistance.The proposed models demonstrate generalization capability across all lanes and directions of the highway.SHAP analysis was extensively employed to characterise the influence of features.Pavement maintenance quality index prediction accuracy was validated for the optimized CatBoost model on both the same and a separate highway segment. Tree-based methods enhanced with particle swarm optimization are proposed for pavement maintenance quality index, surface condition, roughness, rut depth, and skid resistance. The proposed models demonstrate generalization capability across all lanes and directions of the highway. SHAP analysis was extensively employed to characterise the influence of features. Pavement maintenance quality index prediction accuracy was validated for the optimized CatBoost model on both the same and a separate highway segment.