A machine learning and geostatistical hybrid method to improve spatial prediction accuracy of soil potentially toxic elements

Abstract

Effective environmental management and contamination remediation require accurate spatial distributions and predictions of potentially toxic elements (PTEs) in the soil. However, no single method has been developed to predict soil PTE accurately. This study evaluated the advanced geostatistical method of empirical Bayesian kriging regression prediction (EBKRP), machine learning algorithm of random forest (RF), and hybridized model (RF-EBKRP) to predict and map PTEs content in greenspace soils. As identified by RF, soil organic carbon, organic matter, total (nitrogen, phosphorus, and potassium), topographic features, and urban functional types were used as significant covariates to improve the prediction accuracy of PTEs in soil. The model prediction performance was evaluated using the root mean square error (RMSE), mean absolute percentage error (MAPE), and coefficient of determination (R2). Results showed that RF performed much better than EBKRP in predicting soil PTEs, with lower prediction errors and a higher R2. The RMSE, MAPE, and R2 values for the RF model were 0.25–85.32 mg/kg, 3.86–25.40%, and 0.77–0.90, respectively, while the values for the EBKRP method were 0.51–99.03 mg/kg, 5.42–32.13%, and 0.40–0.66. Moreover, the RF-EBKRP method produced more accurate spatial predictions and distributions of PTEs than the individual models, with R2 improvements of 122.5% for EBKRP and 15.58% for RF. The better performances of RF-EBKRP are due to its incorporation of various covariates and its ability to handle complex nonlinear relationships between soil PTEs and covariates. In the end, a hybrid RF-EBKRP method is a promising approach to improving the spatial distribution map of soil PTEs.