Metaheuristic-optimised gradient boosting models for predicting pavement quality and functional-structural performance indices

The increasing demand for road maintenance has intensified the need for effective pavement management, where preventive strategies rely on accurate, interpretable predictions. This study developed XGBoost and CatBoost models optimised using Particle Swarm Optimization (PSO) and Jellyfish Search Optimization (JSO) to predict PQI (PCI, RQI, RDI, SRI, and PWI) and PSSI. The models were trained on four years of highway survey data incorporating categorical variables (lane position, direction, maintenance history), climatic conditions, traffic volume, section length, and observed distress. JSO-optimised models demonstrated greater test stability than PSO-based counterparts. Validation results showed JSO-CatBoost R2 values of 0.8016 (PQI), 0.9344 (PCI), 0.6946 (RQI), 0.6027 (RDI), 0.6027 (SRI), 0.5924 (PWI), and 0.7366 (PSSI). SHAP analysis revealed physically consistent mechanisms, including non-additive crack interactions, moisture-accelerated ageing, and load-induced deformation. These findings indicate that JSO-CatBoost models provide quantifiable predictions of pavement condition indices, while SHAP provides interpretable insights, supporting condition assessment and index-based maintenance prioritisation.