Towards decision-ready explainable machine learning for water quality management using a consistency index and...

Reliable prediction and interpretation of water quality dynamics are essential for environmental monitoring and risk-informed water resources management. Explainable machine learning (XML) offers means to interpret complex predictive models; however, commonly used explanation methods often yield inconsistent feature attributions, and feature selection frequently relies on subjective correlation thresholds. This study develops a unified consistency index (CI) and a data-driven cross-validated recursive feature elimination (RFECV) workflow to quantitatively assess and improve the reliability of XML-based interpretation. Using a 30-year river-water dataset and nine machine-learning algorithms, two XML frameworks were evaluated (correlation-based XML and RFECV-based XML). Correlation-based models achieved strong predictive performance (RMSE = 0.77, 0.64, 1.07), whereas RFECV reduced input dimensionality by 69-85% (from 6 to 12 features) while maintaining comparable accuracy (RMSE = 0.75, 0.57, 1.20). Across the correlation-based workflow, CI values ranged from 0.42 to 0.72, 0.27 to 0.48, and ∼0.60, indicating strong rank-level agreement among interpretation tools. RFECV-based XML preserved predictive accuracy but produced lower CI values (0.00-0.65), reflecting tighter top-k agreement but weaker global ranking coherence. This pattern represents a practically relevant form of reliability in which agreement on core drivers is maintained despite reduced ranking stability. High and α-stable CI values indicate that interpreter disagreements are benign, whereas low and α-sensitive CI values reveal instability in explanations. By providing a quantitative diagnostic check on the robustness of explanations, this study helps ensure that XML-based water-quality assessments are clearer, more trustworthy, and more practical for real-world decision-making.