From fuzzy-TOPSIS to machine learning: A holistic approach to understanding groundwater fluoride contamination

Abstract

Fluoride (F−) contamination of groundwater is a prevalent environmental issue threatening public health worldwide and in India. This study targets an investigation into spatial distribution and contamination sources of fluoride in Dhanbad, India, to help develop tailored mitigation strategies. A triad of Multi Criteria Decision Making (MCDM) models (Fuzzy-TOPSIS), machine learning algorithms {logistic regression (LR), classification and regression tree (CART), Random Forest (RF)}, and classical methods has been undertaken here. Groundwater samples (n = 283) were collected for the purpose. Based on permissible limit (1.5 ppm) of fluoride in drinking water as set by the World Health Organization, samples were categorized as Unsafe (n = 67) and Safe (n = 216) groups. Mean fluoride concentration in Safe (0.63 ± 0.02 ppm) and Unsafe (3.69 ± 0.3 ppm) groups differed significantly (t-value = −10.04, p < 0.05). Physicochemical parameters (pH, electrical conductivity, total dissolved solids, total hardness, NO3−, HCO3−, SO42−, Cl−, Ca2+, Mg2+, K+, Na+ and F−) were recorded from samples of each group. The samples from ‘Unsafe group’ showed alkaline pH, the abundance of Na+ and HCO3− ions, prolonged rock water interaction in the aquifer, silicate weathering, carbonate dissolution, lack of Ca2+ and calcite precipitation which together facilitated the F− abundance. Aspatial distribution map of F− contamination was created, pinpointing the “contaminated pockets.” Fuzzy- TOPSIS identified that samples from group Safe were closer to the ideal solution. Among these models, the LR proved superior, achieving the highest AUC score of 95.6 % compared to RF (91.3 %) followed by CART (69.4 %). This study successfully identified the primary contributors to F− contamination in groundwater and the developed models can help predicting fluoride contamination in other areas. The combination of different methodologies (Fuzzy-TOPSIS, machine learning algorithms, and classical methods) results in a synergistic effect where the strengths of each approach compensate for the limitations of the other.