AI for Surface Water Quality Prediction: A Comprehensive Review

Artificial intelligence for surface water quality prediction has evolved rapidly from classical regression and ensemble learning toward deep sequence, graph-based, physics-informed, and representation-learning approaches, while a largely disconnected systems literature has matured around digital twins, edge intelligence, and federated learning. This critical review synthesizes thirteen primary manuscripts against expanded 2020–2026 literature to examine the relationship between predictive performance and operational deployability. The evidence reveals a persistent bifurcation between an accuracy-maximizing research track and a deployability-maximizing systems track, corroborated independently at classical-modeling, digital-twin, and edge-hardware scales: the corpus's largest single study reports 31,289 samples from 23 stations, whereas a systematic review of 147 water-sector digital-twin studies identifies only 8 cases achieving genuine bidirectional control, and the largest review of environmental federated learning (361 studies) finds only 12 addressing water quality, with predominantly simulated deployments. Ensemble tree methods show the most consistently replicated predictive performance, while near-perfect results reported by simpler models remain constrained by the absence of external or cross-site validation. Explainable AI remains concentrated in feature-attribution methods, convergent SHAP findings require a causal-versus-correlational caveat, and the integration of spatial, physical, and temporal-representational deep-learning strategies remains unexplored. Building on these findings, this review develops a twelve-category Research Gap Matrix and proposes an Integrated Water Intelligence Framework linking prediction, model compression, governance-compliant decision-making, and closed-loop actuation. The review concludes that future progress requires predictive accuracy to be evaluated jointly with validation, explainability, governance, and deployability, and closes with a staged, stakeholder-differentiated roadmap.