Spatial prediction of groundwater nitrate by explainable machine learning models and remote sensing indicators under conditions of imbalanced datasets

This study employed explainable ensemble machine learning models to predict groundwater nitrate (NO3−) contamination and identify key influencing factors. Using 446 groundwater samples, four binary classification machine learning (ML) models, including decision tree (DT), random forest (RF), multilayer perceptron (MLP), and LightGBM were trained with class imbalance handled via SMOTE-NC algorithm and hyperparameters optimized using Tree-structured Parzen Estimator (TPE) method. LightGBM and RF had the best performance (AUROC: 0.9600 and 0.9557), and compared to interpolation methods, ML-based maps offered greater spatial detail and heterogeneity. SHapley Additive exPlanations (SHAP) analysis indicated that agricultural intensity, human footprint, and soil organic carbon were major drivers, along with topography and proximity to NO3− sources. Livestock farming, waste storage, and fertilization were the primary contributors, while industrial sources had limited impact. By integrating multi-dimensional indicators with ML and SHAP analysis, this integrated approach effectively supports accurate NO3− concentration prediction and targeted sustainable groundwater management.