Ensemble Learning for Predicting Groundwater Iron and Manganese Distribution and Identifying Potential Environmental Controls in an Arid Irrigated Region

The enrichment of iron (Fe) and manganese (Mn) in arid groundwater is jointly influenced by multi-source environmental processes, which complicates the delineation of their spatial risks using traditional methods. Taking the Yinchuan Plain as the study area, 228 groundwater samples collected in 2021 were integrated with multi-source environmental variables were integrated to construct a multi-paradigm machine learning predictive framework, aimed at identifying potential environmental controls and high-risk distributions of Fe/Mn. Six machine learning algorithms, including Logistic Regression, Random Forest, LightGBM, XGBoost, TabNet, and Multi-Layer Perceptron, were compared, and a probability-based ensemble model was developed. The results demonstrate that tree-based ensemble algorithms (LightGBM, Random Forest, and XGBoost) outperformed other models, with multi-model probability ensembles providing moderate predictive performance for Fe (AUC = 0.748) and better predictive performance for Mn (AUC = 0.852). Interpretability analysis revealed distinct enrichment pathways for the two elements: Fe enrichment is strongly correlated with soil nitrogen and silt content, whereas Mn exhibits a higher dependency on climatic variables, specifically vapor pressure deficit. Because groundwater samples from multiple aquifer systems were integrated into a single modeling framework, the predictions represent integrated groundwater conditions across the study area, which may limit the model’s ability to capture aquifer-specific heterogeneity. The model generated risk probability maps help identify potential contamination hotspots and priority areas for monitoring and control, providing a scientific foundation for formulating differentiated groundwater safety strategies in arid regions.