An Integrated GIS and Explainable Machine Learning Framework for Groundwater Contamination Risk Assessment and Drinking Water Security

Abstract

Geogenic contamination of groundwater by arsenic and fluoride threatens drinking-water security across the Ganga basin, and Bihar is among the most affected Indian states. GIS overlay indices rely on expert weights, while machine learning (ML) models are accurate but rarely interpretable enough for public-health decisions. This paper presents an integrated framework that couples GIS-based predictor engineering, spatially cross-validated ML classification, SHapley Additive exPlanations (SHAP) and a drinking-water security index (DWSI) that converts contamination probabilities into population-weighted intervention priorities. Because explanations must be shown to recover true processes before they guide policy, the framework was evaluated on a controlled synthetic Ganga-plain benchmark (120 km × 80 km, 1,800 wells) with known arsenic and fluoride generating rules. Under spatial-block cross-validation, logistic regression (AUC 0.929), random forest (0.925) and XGBoost (0.917) performed similarly and all exceeded SVM (0.878); fluoride was predicted with an AUC of 0.870. SHAP recovered the imposed controls, namely clay-rich Holocene sediments, proximity to the river and the 20–60 m shallow-well risk zone, and it exposed a spurious rainfall signal arising from spatial correlation. The DWSI placed 53.7% of the population in high or very high priority classes on 44.7% of the area, and these classes accounted for 90.8% of expected exposure. The framework is transparent and transferable to Bihar's monitoring data for well-testing and safe-source planning.