Machine learning approach to root intrusion prediction in urban sewers using CCTV and environmental features

Abstract

Sewer blockages from root intrusion pose significant economic and environmental challenges for water utilities worldwide. While machine learning (ML) offers promising solutions for infrastructure management, its application to this specific failure mode remains largely unexplored. This study develops a comprehensive ML framework for predicting root intrusion risk in sewer systems by integrating physical pipe attributes, environmental features, and demographic data from Hong Kong. Three classification algorithms; Logistic Regression (LR), Support Vector Machine (SVM), and Random Forest (RF) were systematically evaluated and compared. Results showed that LR and RF achieved statistically equivalent and superior performance with an AUC-ROC of 0.933, while SVM performed marginally lower at 0.914. The comparable performance of the simpler linear model (LR) with the complex ensemble method (RF) indicates that the predictive relationships are predominantly linear in nature. Feature importance analysis revealed that geographic (District) and demographic (Total Population) contextual factors were more influential predictors than specific pipe characteristics. These findings provide water utilities with a highly interpretable and effective tool for proactive asset management, enabling targeted inspections, optimized maintenance scheduling, and improved resource allocation.