Artificial Intelligence for Water Quality Prediction: From Data-Driven Models to Explainable, Physics-Informed and Generalizable Hydrological Intelligence

Abstract

Artificial intelligence (AI) and machine learning (ML) are increasingly used to predict water-quality dynamics from high-frequency sensors, routine monitoring, hydrometeorological observations, remote sensing and process-model outputs. However, the field has progressed faster in algorithmic complexity than in hydrological validation. This focused review synthesizes recent developments in conventional ML, deep learning, hybrid and ensemble methods, graph-based spatiotemporal learning, transfer learning, explainable AI (XAI), probabilistic prediction and physics-informed learning for surface-water quality prediction. Rather than ranking algorithms by reported accuracy, we evaluate their suitability against seven criteria: predictive skill, temporal generalization, spatial transferability, physical consistency, interpretability, uncertainty representation and reproducibility. Tree ensembles remain strong baselines for structured monitoring data; recurrent and attention-based networks are advantageous for sequential prediction; graph neural networks explicitly represent river-network connectivity; and transfer learning provides a route from data-rich to data-poor sites. Yet common practices—random data splitting, preprocessing before partitioning, single-site evaluation and deterministic point forecasts—can overstate operational skill. XAI improves model interrogation but does not establish causality, while physics-informed approaches are promising only when constraints are hydrologically defensible and independently tested. We propose a validation framework that couples chronological and spatial holdouts, event-stratified evaluation, uncertainty calibration and physical diagnostics. The next generation of water-quality AI should therefore be judged not by marginal reductions in RMSE alone, but by its ability to remain accurate, physically plausible, interpretable and uncertainty-aware under hydrological conditions that differ from the training data.