Statistical and Spatial Evaluation of Machine Learning Models for Imbalanced Soil Heavy Metal Contamination Classification

Abstract

Soil heavy metal contamination poses significant environmental and public health risks, requiring accurate and interpretable classification methods for effective risk assessment. This study presents a comparative evaluation of five machine learning models—Random Forest, XGBoost, LightGBM, TabNet, and Multi-Layer Perceptron (MLP)—for multiclass soil contamination classification using the publicly available Soil Heavy Metal Contamination Dataset (Kaggle) containing 1000 soil samples with heavy metal concentrations, soil characteristics, and geographic coordinates. To investigate the contribution of spatial information, models were evaluated under both spatial and non-spatial feature configurations using stratified 5-fold cross-validation. Performance was assessed using Macro F1-score, Matthews Correlation Coefficient (MCC), ROC-AUC, and G-Mean. Statistical significance was evaluated using the Wilcoxon signed-rank test, while Moran's I and SHAP analyses were employed to examine spatial dependence and model interpretability, respectively. Among the evaluated models, the non-spatial MLP achieved the best performance, with a Macro F1-score of 0.869 ± 0.051 and an MCC of 0.862 ± 0.027. The Wilcoxon test (p = 0.625) and Moran's I (0.0217, p = 0.119) indicated that incorporating raw spatial coordinates did not provide a statistically significant improvement and that contamination exhibited no significant spatial autocorrelation. SHAP analysis further identified Zn and Pb as the most influential predictors, whereas latitude and longitude contributed minimally. These findings demonstrate that contamination patterns in the studied dataset are driven primarily by heavy metal concentrations rather than geographic location, providing an interpretable and statistically validated framework for soil contamination assessment.