Global ecological risk assessment of soil contamination by heavy metal(loid)s based on machine learning

Soil contamination by heavy metal(loid)s, due to their persistence and bioaccumulation potential, has emerged as a global environmental issue. Rapid and accurate identification of soil contamination risk levels is critical for ensuring ecosystem safety and human health. In this study, a machine learning-based framework for multi-class classification was developed to identify and interpret ecological risk levels associated with soil contamination by heavy metal(loid)s. Based on 9489 soil samples collected from multiple countries and regions, a comprehensive dataset encompassing soil properties, metal(loid) descriptors, and total metal(loid) content was constructed. The performance of six commonly used machine learning classification models was systematically evaluated, with hyperparameter tuning applied to enhance overall model performance. Model interpretation techniques were employed to elucidate the relationships between ecological risk levels associated with soil contamination by heavy metal(loid)s and input features. The results demonstrated that the categorical boosting model outperformed others on both training and testing sets, achieving an accuracy of 0.681 and an area under the curve (AUC) value of 0.885 on the testing set. Permutation importance analysis identified total metal(loid) content, pH, organic carbon, clay, cation exchange capacity, and electron affinity as key features influencing risk levels classification, with partial dependence plots and Shapley additive explanations further characterizing their nonlinear impact on predictions. This study presents a novel approach for developing effective and interpretable systems to rapidly identify ecological risk levels associated with soil contamination by heavy metal(loid)s, providing a scientific basis and priority restoration guidance for global contamination management and sustainable land management.