Abstract
Debris accumulation is a primary driver of reduced hydraulic capacity and urban flooding in sewer networks, yet it remains underexamined in predictive modelling. This study develops an integrated framework to predict debris criticality (cross-sectional area obstruction ≥ 25%) using a comprehensive municipal CCTV inspection database from Hong Kong’s urban drainage network, comprising 5470 drainage folders and 30,241 inspection files collected over 14 years (2007–2021). From this network-scale database, 8165 debris-affected pipe segments were extracted for analysis, including an isolated-debris subset for Weibull Accelerated Failure Time (WAFT) modelling and the full debris cohort for machine learning prediction. The WAFT model identified pipe diameter, age, and Vitrified Clay (Time Ratio = 0.889 vs. Concrete) as dominant physical risk determinants, compounded by traffic and industrial land use. Interpreted via SHAP, the machine learning confirmed diameter and age as primary predictors. The tuned LightGBM achieved the strongest discrimination (ROC-AUC = 0.871), and threshold optimisation recovered 76% of high-criticality segments, supporting proactive intervention. This convergence of survival and ML findings provides complementary evidence and directional consistency across inferential and predictive modelling frameworks, supporting a transition toward risk-stratified maintenance to enhance the hydrological resilience of urban catchments.