Hydraulically Motivated Feature Engineering and Spatial Ensemble Machine Learning: Smart-Water Leakage Detection and Spatial Prioritization

Abstract

Non-revenue water from undetected leaks is a persistent inefficiency of urban infrastructure, yet public machine-learning (ML) leak-detection benchmarks are reused without scrutiny of their labels. This study audited a widely circulated smart-water benchmark and found its binary label to be an exact function of two predictors, pressure and Flow_Rate. An XGBoost classifier on hydraulically motivated features was evaluated under Leave-One-Spatial-Component-Out (LOSCO) cross-validation with fold-safe preprocessing. It reached 99.88% accuracy (ROC-AUC = 0.9998), but ablation showed that this performance is carried entirely by the three features derived from the two rule variables; the other five features scored ROC-AUC = 0.494. A 200-permutation label test (null ROC-AUC = 0.498 ± 0.019) and an all-row preprocessing counterfactual detected no test-fold information entering training. Isolation Forest fusion gave no significant gain (DeLong p = 0.23). Across 125 pipe segments, the risk score showed no spatial autocorrelation (Moran’s I = 0.020, p = 0.26) and no false-discovery-rate-significant Getis–Ord Gi* hotspot, and Gi* smoothing did not improve segment ranking. On the independently labelled LeakDB benchmark, the features reached ROC-AUC = 0.72–0.79 and ranked the true leak node above chance (exact p = 0.015). The framework supports transparent spatial prioritization and benchmark auditing, not leak localization.