Explainable machine learning for water quality prediction across freshwater systems and water infrastructure: A comprehensive review

Water quality prediction in freshwater systems and engineered water infrastructure is increasingly supported by machine learning (ML), yet limited model interpretability constrains its scientific and operational utility. Explainable artificial intelligence (XAI) has emerged as a critical framework for addressing this limitation by providing insights into model behavior, dominant predictors, nonlinear responses, and context-dependent interactions. However, existing reviews have largely treated XAI as a complementary component, offering descriptive summaries without critically examining methodological assumptions, interpretive boundaries, or cross-system applicability. In contrast, this review provides a comprehensive and critical synthesis that positions explainability as a central analytical framework and integrates applications across both freshwater systems and engineered water infrastructure. More than 70 studies published since 2018 were analyzed to identify recurring explanatory patterns across key water quality parameters, including nutrients, algal indicators, oxygen- and organic matter–related variables, composite indices, and operational parameters. Across these domains, XAI reveals that model-learned water quality patterns are often structured by nonlinear, threshold-dependent, and context-specific interactions among temperature, nutrient availability, hydrological conditions, light availability, and operational variables. The synthesis highlights the complementary roles of global and local explanations in capturing system heterogeneity and supporting both process understanding and operational decision-making. Key limitations include the non-causal nature of feature attribution and sensitivity to data structure, predictor collinearity, model configuration, and explanation stability. Future work should prioritize integration with process-based and causal inference frameworks, uncertainty-aware explainability, standardized evaluation protocols, and applications across upstream–downstream linkages. Overall, this review reframes XAI as an integrative framework linking prediction, interpretation, and decision support across the water cycle.