Multi-tier framework for automated sewer defect classification and inspection prioritization

Aging assets and climate stressors threaten sewer reliability, while most closed-circuit television (CCTV) inspection evidence remains locked in semi-structured PDF reports, constraining utilities to reactive maintenance. This paper develops an explainable framework that digitizes 12,806 Hong Kong CCTV defect records into a geodatabase, fuses them with pipe, climate, traffic and population covariates, and trains Bayesian-optimized categorical boosting (CatBoost) models in a three-tier hierarchy to predict structural/operational defects, defect families and specific manifestations. The best models achieve macro-F1 above 0.80 at Tier 1 and retain strong performance in deeper tiers. SHapley Additive exPlanations (SHAP) analysis reveals rainfall, pipe age, diameter and traffic as dominant drivers and identifies thresholds that are consistent with engineering judgement. Model outputs are converted into severity-aware inspection rankings, enabling a shift from reactive responses to predictive scheduling. Contributions are threefold: data fusion of digitized CCTV records with contextual covariates; multi-tier CatBoost+SHAP classification; and an automated prioritization workflow.