The elevated concentration of fluoride (F− ≥ 1.0 mg/L) in worldwide groundwater is a substantial environmental hazard to human health that necessitates measurement and monitoring. Water quality assessment and surveillance are labor-intensive and expensive. Consequently, novel modeling techniques, such as machine learning algorithms, serve as a powerful tool for predicting F− concentration, thereby facilitating the development of tailored mitigation plans. Predicted F− concentration with machine learning methods of varying complexity- classification and regression tree (CART), multivariate adaptive regression spline (MARS), random forest (RF), gradient boosting machines (GBM), and eXtreme gradient boosting (XGB) were coupled with GIS to map F− hazards. The concentrations of F− in 1178 wells were analyzed, together with the influencing groundwater quality parameters e.g., pH, EC, TDS, TH, TA, Ca2+, Mg2+, Na+, K+, HCO3−, SO42−, and Cl−. The feature importance analysis revealed that F− concentrations are significantly influenced by pH, TA, Na+, K+, Mg2+, and Cl−. The performance of the applied ML algorithms indicated that XGB had superior performance in forecasting F− concentration, with the lowest RMSE (0.238) and MAE (0.0756). Delineation of hazard quotient (HQ > 1.0) from the predicted data points suggests imminent risk to the infants than children and adults due to excessive exposure to F− in potable water. The integration of diverse approaches (GIS and machine learning algorithms) produces a synergistic effect, and helps identify areas with critical monitoring needs. Graphical Abstract The graphical abstract outlines hydro-chemistry, feature selection, predicting concentrations of fluoride in groundwater, integration of predicted fluoride with GIS, and consequent health risk assessment. It begins with physico-chemical parameters of groundwater, data pre-processing followed by features selection for model training. Various ML models viz. CART, MARS, RF, GBM, and XGB are evaluated as per ascending model complexity to ascertain the simplest and high predictive model. The performance comparison reveals XGB to be the most accurate model. These findings will be impactful in groundwater quality management and remediation strategies. For a deeper analysis and practical implications, we invite readers to explore the full article.

Enhancing Spatial Interpretability of Fluoride in Groundwater Using an Integrated GIS and Machine Learning Approach

期刊论文
发表日期：

2025年5月30日

查看全文

Rahul Singh,
Subhasish Tripathy,
Ashok Kumar Gupta,
Venkatesh Uddameri,
Sanket Ratan Bagal

摘要

The elevated concentration of fluoride (F⁻ ≥ 1.0 mg/L) in worldwide groundwater is a substantial environmental hazard to human health that necessitates measurement and monitoring. Water quality assessment and surveillance are labor-intensive and expensive. Consequently, novel modeling techniques, such as machine learning algorithms, serve as a powerful tool for predicting F⁻ concentration, thereby facilitating the development of tailored mitigation plans. Predicted F⁻ concentration with machine learning methods of varying complexity- classification and regression tree (CART), multivariate adaptive regression spline (MARS), random forest (RF), gradient boosting machines (GBM), and eXtreme gradient boosting (XGB) were coupled with GIS to map F⁻ hazards. The concentrations of F⁻ in 1178 wells were analyzed, together with the influencing groundwater quality parameters e.g., pH, EC, TDS, TH, TA, Ca²⁺, Mg²⁺, Na⁺, K⁺, HCO₃⁻, SO₄²⁻, and Cl⁻. The feature importance analysis revealed that F⁻ concentrations are significantly influenced by pH, TA, Na⁺, K⁺, Mg²⁺, and Cl⁻. The performance of the applied ML algorithms indicated that XGB had superior performance in forecasting F⁻ concentration, with the lowest RMSE (0.238) and MAE (0.0756). Delineation of hazard quotient (HQ > 1.0) from the predicted data points suggests imminent risk to the infants than children and adults due to excessive exposure to F⁻ in potable water. The integration of diverse approaches (GIS and machine learning algorithms) produces a synergistic effect, and helps identify areas with critical monitoring needs.

Graphical Abstract

The graphical abstract outlines hydro-chemistry, feature selection, predicting concentrations of fluoride in groundwater, integration of predicted fluoride with GIS, and consequent health risk assessment. It begins with physico-chemical parameters of groundwater, data pre-processing followed by features selection for model training. Various ML models viz. CART, MARS, RF, GBM, and XGB are evaluated as per ascending model complexity to ascertain the simplest and high predictive model. The performance comparison reveals XGB to be the most accurate model. These findings will be impactful in groundwater quality management and remediation strategies. For a deeper analysis and practical implications, we invite readers to explore the full article.