An Interpretable Machine Learning Framework for River Water Hardness Classification

Aditya Rajiv Ratnam 

Determining water hardness in the field currently relies on laboratory Ethylenediaminetetraacetic acid (EDTA) titration or single-proxy electrical conductivity (EC) measurements, both of which have significant limitations. This study investigates whether routine in situ physicochemical parameters can accurately classify river water hardness without laboratory analysis. Using samples from a river basin in Argentina, four ML classifiers were evaluated against a trivial majority-class classifier (71.4% accuracy) and an EC threshold baseline (75% accuracy). A Random Forest model achieved 95.45% accuracy, a 20.45 percentage point improvement over the EC baseline. Five complementary feature importance and model interpretability approaches (Gini importance, permutation importance, SHapley Additive exPlanations (SHAP), feature ablation, and correlation analysis) show that the model captures physicochemical relationships pertaining to ionic interference, hydrological dilution, and temperature-dependent variations in conductivity. A minimal three-sensor suite comprising EC, Total Suspended Solids (TSS), and sample temperature matched full-model performance, demonstrating that accurate hardness classification is feasible with a low-cost, reagent-free field system. These findings support a practical pathway toward simpler and more sustainable water quality monitoring, in alignment with green chemistry principles of waste prevention and reduced chemical use.