Machine learning framework for irrigation water quality assessment in Kerala Rivers, India.

Accurate assessment of irrigation water quality is essential for sustainable agricultural management, particularly in regions facing increasing pressure on freshwater resources. Conventional water quality indices often require extensive parameters and may fail to capture nonlinear hydro-chemical relationships. An integrated machine learning framework to predict and classify the Irrigation Water Quality Index is developed in this study, using routinely monitored physicochemical parameters from 44 rivers in Kerala, India. Fuzzy Irrigation Water Quality Index (FIWQI) were used for Irrigation Water Quality assessment and were computed using fuzzy analytic hierarchy process (Fuzzy-AHP), incorporating uncertainty in parameter weighting, and used as the target variable for model development. Three categories of machine learning models: tree-based ensemble, neural network, and classical models were evaluated. Data preprocessing included normalization, multicollinearity analysis, and class balancing using SMOTE, while hyperparameters were optimized via five-fold cross-validation. Results show that tree-based ensemble models outperformed others, with CatBoost achieving the highest accuracy (R2 = 0.996; RMSE <1), and a substantial (>70%) improvement over classical models, while classification accuracy exceeded 97% in tree-based ensemble models compared to classical models. Uncertainty analysis using quantile regression indicated stable prediction behaviour, while the Wilcoxon signed-rank test confirmed statistically significant differences between models (p < 0.01). Sensitivity analysis indicated high robustness, with <5% variation under ±20% parameter weight changes. Feature importance analysis identified boron, fluoride, and electrical conductivity as key predictors in classification. The proposed framework offers a scalable, data efficient solution for irrigation water quality monitoring and decision support systems in river-dependent agricultural regions.