Predicting heavy metal distribution coefficient in global soil via machine learning: The effect of mineral...

Heavy metal adsorption by soil particles significantly influences its pollution risk in global soil. However, the global distribution of soil adsorption ability is still lacking due to the difficulty of obtaining large-scale adsorption data. Here, we proposed a novel method to evaluate the adsorption ability in a global scale through predicting the distribution coefficient (K_d) of HMs in soil via machine learning and big data. Based on the data selected, which included soil properties, adsorption properties, and mineral contents, the Random Forest (RF) achieved prediction accuracies of 0.91, 0.85, 0.90, and 0.87 for K_d values of As, Cd, Cr, and Pb, respectively. The main factor influencing K_d values was found to be the mineral contents. The corresponding contents of Fe_oxide, chlorite, and kaolinite exhibited significant positive correlations with As, Cd, and Cr, respectively. The pollution risk areas were further delineated based on the U.S. EPA metrics of K_d. It is noteworthy that the environmental risk areas of farmland on each continent exceed 20%, with Asia had 72%, 64% and 58% of the risk areas for As, Cd and Cr, which are obviously larger than those on other continents. Moreover, the distributions of these high - risk areas were consistent with those of the corresponding minerals below or close to the global mean contents. The approach proposed in this study enables the prediction of K_d values and pollution risk assessment of soils globally, revealing the impact of mineral heterogeneity on K_d values. These results are valuable for evaluating environmental risks and formulating soil remediation strategies.