Integrating explainable artificial intelligence and policy attribution to evaluate water-chemistry and hydro-climatic drivers over river hydrology in the Yamuna River

River ecosystems are critical socio-ecological systems to support human well-being but face increasing pressures from pollution and climate variability. The Yamuna River has experienced persistent water quality (WQ) degradation despite long-term restoration efforts. This study evaluated whether predictor importance changed across 25 years of restoration activities (1997–2022). A Hydro-Climatic (HC) and Water-Chemistry (WC) Domain Decomposition Framework was combined with explainable artificial intelligence (SHAP), and twelve machine-learning algorithms were evaluated using leakage-controlled spatial and temporal cross-validation. WC models showed high spatial transferability, with R 2 values of 0.962–0.975 for the five strongest algorithms, whereas all HC models produced negative spatial R 2 values (−0.166 to −0.693). Among the stable HC models with positive temporal performance, R² ranged from 0.354 to 0.582. SHAP analysis identified Fe, PO₄-P, NO₂-N, NO₃-N, and NH₃-N as the five highest-attributed WC predictors. Their ranking remained stable across the three policy eras, while the mean pairwise rank correlation across held-out station folds was 0.942. The five predictors accounted for 58.0% of total absolute SHAP attribution. A Water-Chemistry-to-Hydro-Climatic Attribution Ratio R W C / H C was calculated from a single combined-domain model, with a full-record reference value of 21.88 (95% CI: 21.63–22.13). The framework provides a transparent basis for interpreting water-quality prediction and supporting evidence-informed river-basin management.