Standardization and Benchmarking of Datasets and Evaluation Protocols for Machine Learning-Based Water Leak Detection: A Systematic Review

Despite the increasing use of machine learning (ML) and deep learning (DL) for leak detection and localization in water distribution networks (WDNs), standardized datasets, common evaluation protocols, and consistent reporting practices remain limited, thereby restricting meaningful comparison among studies. In this systematic review, PRISMA 2020 was followed to examine the benchmarking practices, data sources, sensing modalities, and evaluation strategies adopted in recent ML/DL-based leak-detection studies. Scopus was searched in April 2026, and 90 studies were included after screening against predefined eligibility criteria. Substantial variation was identified in dataset generation, model validation, and performance assessment. Thirty studies were assigned to the public-benchmark category, but only 23 used one of the three named water-leak resources tracked in this review; the remainder included topology-only reuse or other heterogeneous public resources. Pressure measurements were the dominant sensing modality, whereas flow, transient, and smart-meter data remained comparatively underexplored. Considerable inconsistency was also observed in performance metrics, validation procedures, and leak-scenario definitions, preventing fair comparison of detection and localization performance. Based on these findings, a practical reporting framework is proposed to improve transparency, reproducibility, and comparability. The principal recommendations include standardized dataset documentation, consistent evaluation protocols, expanded community benchmarks, uncertainty quantification, and model interpretability to support operational deployment.