Robust water quality prediction across multiple indicator formulations using an explainable ensemble learning model

Reliable assessment of water quality is critical for sustainable resource management. The Water Quality Index (WQI) is a valuable tool for evaluating water quality. However, the existence of multiple WQI techniques often results in inconsistent assessments and considerable uncertainty in the correct water quality classification. This study aimed to examine the robustness of machine learning (ML) methods for water quality assessment across several widely used WQIs. A stacking ensemble learning framework was developed, integrating multiple ML base learners with a meta-learner based on the Ensemble Deep Random Vector Functional Link structure. Experimental results show that the proposed approach consistently achieves superior performance, outperforming existing state-of-the-art models with accuracy, precision, recall, and F1-score ranging from 99.17% to 99.59% across all evaluated WQIs. SHAP-based analysis was employed to interpret the model’s output and clarify the contribution of each base learner to the final decision. It revealed that the most influential interactions affecting classification outcomes were those between BOD and DO, as well as between conductivity and pH. Meanwhile, the main contributors among the base learners were the deep learning models, particularly LSTM and MLP. These findings demonstrate that the proposed framework not only resolves inconsistencies in multiple index-based water assessments but also provides a transparent and reliable decision-support tool for water quality monitoring and management.