Comparing AI-driven approaches for predicting river water quality: a systematic review of water quality indices and...

River water quality (RWQ) monitoring faces growing challenges from geographical data gaps, methodological inconsistencies, and limited uncertainty quantification in predictive modelling frameworks. This systematic review aims to critically synthesise, compare, and identify knowledge gaps in AI-driven RWQ prediction studies, by synthesizing 71 peer-reviewed case studies published between 2018 and 2024, grouped into (i) traditional water quality index (WQI) approaches (n = 44) and (ii) remote sensing (RS) approaches (n = 27), both employing machine learning or deep learning (ML/DL) methods. A strong geographical bias is identified, with 83% of studies concentrated in Asia, and limited representation from other regions. The review identified that WQI studies exhibit substantial variability in parameter selection, sampling type and frequency, weighting, and aggregation. RS-based studies predominantly retrieve optically active (OA) parameters using Sentinel-2 and Landsat-8/9 imagery, with optically inactive (OIA) parameters inferred indirectly. Across both groups, performance evaluation relies largely on conventional accuracy metrics, with sensitivity and uncertainty analyses rarely reported. Overall, this review identifies inconsistent preprocessing protocols, and the absence of standardised benchmarking as critical barriers to the transferability and operational reliability of current RWQ prediction frameworks. This review provides a consolidated evidence base and practical recommendations for developing more transferable and operationally reliable RWQ prediction framework with priority directions for future research.