Smart Ensemble Learning Framework for Predicting Groundwater Heavy Metal Pollution

Abstract: Groundwater in the Densu Basin is increasingly threatened by heavy metal contamination, yet conventional assessment methods struggle to capture the statistical complexity and spatial heterogeneity of pollution indicators. A critical challenge lies in modelling the Heavy Metal Pollution Index (HPI), which is typically skewed and influenced by correlated contaminants, leading to biased predictions when modelled without transformation. This study develops a predictive framework that integrates response transformations with nested cross-validated ensemble machine learning to address these limitations. Three transformations; raw, log, and Gaussian copula, were applied to HPI and evaluated across six learners: support vector regression (SVM), k-nearest neighbours (k-NN), CART, Elastic Net, kernel ridge regression, and a stacked Lasso ensemble. Diagnostic evaluation showed that raw-scale models produced deceptively high fits, with Elastic Net and the stacked ensemble reporting values close to 1.0, raising concerns of over-optimism and potential information leakage. The log transformation stabilised variance, improving prediction for SVM (, RMSE) and k-NN (, RMSE), though the performance of Elastic Net deteriorated. The Gaussian copula transformation yielded the most reliable outcomes: the stacked ensemble achieved with RMSE, while other learners such as SVM (, RMSE) and k-NN (, RMSE) maintained high accuracy. Importantly, copula-based models improved residual behaviour and produced spatially plausible prediction maps, reinforcing their potential for groundwater quality management. Clustering analysis using DBSCAN further revealed the dominance of Fe, followed by Mn, as the primary contributors to HPI, consistent with regional hydrogeochemical processes. Limitations include the reliance on random rather than spatial cross-validation and the basin-specific nature of the analysis, which may constrain transferability. Future research should explore spatially explicit validation schemes and extend the framework to diverse hydrogeological settings. Overall, the study advances predictive hydrogeochemistry by demonstrating that distribution-aware ensembles, complemented by clustering diagnostics, can provide robust and interpretable assessments of groundwater contamination.