An Uncertainty-Aware Explainable Ensemble Framework for Water Quality Index Prediction Using Multi-Scale Spatial Encoding

Abstract

Water Quality Index (WQI) prediction is used to maintain a sustainable water resource management, environmental protection, and public health. The traditional machine learning approaches achieve high accuracy. However, they do not quantify uncertainty, and suffers from spatial adaptability, and interpretability. This paper presents an integrated, uncertainty-aware framework for WQI prediction that combines ensemble machine learning, Bayesian optimization, explainable artificial intelligence, and deep learning. The performance of existing algorithms is evaluated and compared with proposed methodology. A multi-scale spatial positional encoding is implemented to effectively capture geographical variability, as it enables the model to learn both local and global spatial dependencies. Bayesian optimization is utilized for efficient hyperparameter tuning. Also, SHAP-based explainability identifies the most influential hydrochemical parameters affecting WQI predictions. Furthermore, Bayesian deep learning quantifies both epistemic and aleatoric uncertainties, as it enhances model reliability and decision-making confidence. The experimental results show that the proposed framework outperforms conventional models, and achieves an $\mathbb{R}\mathbb{M}\mathbb{S}\mathbb{C}$, $\mathbb{M}\mathbb{A}\mathbb{E}$, $\mathbb{M}\mathbb{A}\mathbb{P}\mathbb{E}$, and $\mathbb{C}\mathbb{D}\mathbb{E}$ of 3.92, 2.87, 7.54%, and 0.98, respectively. Further, SHAP analysis highlights the significance of fluoride, magnesium, and calcium-magnesium ratio in water quality prediction. The proposed framework provides a robust and interpretable solution for predictive water quality assessment. It supports risk-informed environmental monitoring, adaptive resource management, and sustainable water governance.