An explainable AI framework for multi-stressor climate risk screening of road networks using XGBoost and SHAP

Road networks face increasing risks from compound climate and traffic stressors, yet practical tools for interpretable, operational assessment remain limited. This study develops a reproducible explainable artificial intelligence framework for network-scale climate risk screening of road infrastructure using Extreme Gradient Boosting and SHapley Additive exPlanations. Models were developed for cracking, roughness, and rutting across 7419 Victorian road links. Performance varied across distress types and classes. Stage-specific diagnostics showed that the rutting model did not reliably discriminate distressed and minority severity classes. Rutting was therefore retained as a reported predictive outcome but excluded from the adopted Unified SHAP aggregation and Climate Risk Formula. The adopted formula integrates the cracking and roughness models within a reduced-form hazard–exposure–vulnerability structure. Global attribution identified Exposure as the largest component (41.1%), followed by Vulnerability (33.1%) and Hazard (25.8%). At the local Climate Risk Score level, mean Aumann–Shapley contributions were highest for Vulnerability (40.1%), followed by Hazard (31.5%) and Exposure (28.4%). Bootstrap and Random Forest sensitivity analyses showed generally stable dominant predictor rankings, although exact component shares remained model-dependent. The resulting scores provide relative, spatially explicit indicators for network screening and prioritisation rather than precise predictions of rare severe distress. The framework offers a transparent and scalable basis for cross-sectional infrastructure climate risk assessment while aligning explainability claims with demonstrated model reliability.