Employing machine learning to predict the occurrence and spatial variability of high fluoride groundwater in intensively irrigated areas

Abstract

Groundwater serves as a vital source of drinking water in the Hetao Basin of northwestern China. However, the occurrence and intense spatial variability of high fluoride (F−>1.5 mg/L) groundwater poses a significant health threat to local residents. In this study, the Random Forest (RF) machine learning algorithm was employed to develop two prediction models for high fluoride groundwater in Qiantao and Houtao plain, northwestern China, based on 650 groundwater samples. The models incorporated hydrochemical and geospatial datasets (climatic factors, topographic factors, and soil properties) to predict the probability of fluoride in groundwater exceeding 1.5 mg/L at a 250 m spatial resolution. The results show that the high-risk areas of the Qiantao Plain and Houtao Plain cover 33.0 % and 6.7 % of the total area of the plains, respectively. The ratio of high risk (F > 1.5 mg/L) samples correctly detected in Qiantao plain and Houtao plain were 86.7% and 94.5%, indicated an increased sensitivity to high fluoride category. Hydrochemical parameters were proven to be the most effective predictors. The spatial variability of high fluoride groundwater in two plains was dominated by the lithology, tectonic evolution, and hydrogeologic settings. Hyper-parameter optimized RF models demonstrate an outstanding performance in predicting the risk of high fluoride groundwater. This research presents a cooperative and successful effort to integrate hydrochemical and geospatial variables for prediction model, which can enhance the understanding of the geological origin and anthropogenic causes, and facilitate decision-making to identify potential risks in the areas affected by high levels of fluoride in groundwater.