Minimal-Variable Water Quality Prediction Enabled by Explainable Machine Learning in an Urbanized Plain Watershed

Increasing anthropogenic pressures have intensified challenges to surface water environments in urbanized plain watersheds. This study developed an integrated framework that combines hydrochemical analysis, multivariate statistics, the Weighted Arithmetic Water Quality Index (WAWQI), and interpretable machine learning to investigate hydrochemical processes, evaluate water quality, and predict variations in water quality in an urbanized plain watershed in southwestern China. Based on 112 surface water samples collected across four hydrological periods, results showed that silicate weathering and carbonate dissolution dominated hydrochemical evolution, resulting in a prevailing Ca–HCO 3 type. Anthropogenic inputs, particularly nutrient enrichment, exerted increasing influences during the dry season. The WAWQI assessment indicated that over 70% of samples were suitable for drinking, but significant spatiotemporal variability was observed. The poorest water quality occurred in Period 4, with a mean WAWQI of 82.83, and was mainly distributed in urbanized and downstream areas. Sensitivity analysis and SHAP interpretation identified total phosphorus (TP) and ammonium (NH 4 + ) as the primary drivers of water quality variability. Support Vector Regression (SVR) achieved the best predictive performance (R 2 = 0.98, RMSE = 4.26, MAE = 1.65, and PBIAS = -0.52) with the lowest uncertainty. SHAP-based feature optimization revealed predictor redundancy: a reduced model using only 13.3% of the variables maintained comparable accuracy. This framework provides an efficient and transferable approach for water quality assessment and prediction in rapidly urbanizing plains.