Title : Explainable ensemble learning for spatiotemporal water quality assessment in the Narmada River Basin, India
Abstract:
Sustainable river basin management depends on accurate water quality assessment. However, conventional Water Quality Index (WQI) methods assign fixed weights to parameters and thus fail to re?ect spatiotemporal variability across large, hydrologically heterogeneous basins. This study proposes a framework that integrates multivariate statistical analysis, ensemble machine learning, and explainable artificial intelligence (XAI) to evaluate water quality in the Narmada River Basin, India. We used 3,341 observations collected between 2017 and 2025 at 50 monitoring stations operated by the Madhya Pradesh Pollution Control Board. The data cover eight physicochemical and microbiological parameters. WQI was computed with the standardized weighted arithmetic method and grouped into five quality classes. The results show widespread degradation across the basin: 63.6% of samples were classified as poor, only 0.7% as good, and none as Excellent. We benchmarked nine machine learning classifiers. CatBoost performed best, with 95.67% accuracy and a weighted F1-score of 0.9562, exceeding linear and single-tree baselines by 3.9 to 8.5 percentage points. SHAP analysis identified turbidity, biochemical oxygen demand, and dissolved oxygen as the main contributors to degradation, and Linear Discriminant Analysis confirmed this result. The study shows that interpretable ensemble models can achieve high predictive accuracy while also yielding physically meaningful insights, making them a scalable decision-support tool for water quality monitoring and river basin governance. Future work will add adaptive multi-criteria weighting and formal uncertainty quantification to support operational deployment.
