Machine learning-based prediction of elevated uranium concentrations in shallow groundwater of Punjab, India

Excess uranium (U) in the shallow aquifers of Punjab, India, has become a significant public health concern for the population dependent on groundwater for drinking and irrigation purposes. Although prior investigations have statistically established the control of geogenic and anthropogenic factors on U enrichment, a comprehensive and high-resolution spatial distribution of the extent of contamination remains lacking. To address this gap, we employed the Random Forest (RF) machine-learning classifier to model 1,852 data points of groundwater U concentrations compiled from different districts of Punjab. Spatial prediction and mapping were performed using spatially continuous predictor variables pertaining to meteorological, topographical, geological, soil, and other relevant parameters. A highly accurate prediction map of the occurrence probability of U surpassing the WHO drinking water limit of 30 µg L^-1 at a 250 m spatial resolution, with an accuracy of 85% for test data and 87% for validation data, was generated. The predicted U hazard was strongly influenced by potential evapotranspiration, elevation, and aquifer thickness, with a moderate to low influence from soil physical and chemical properties. Based on the predicted hazard map, the probability of U contamination was higher in the south and southwestern districts (Malwa region) than in other regions of Punjab, comprising approximately 1.7 million hectares (~ 35%) of the state's total area. This study represents the first attempt to spatially predict the occurrence of high groundwater U levels, providing valuable insights for government agencies and policymakers to make informed decisions and manage groundwater sustainably.