Disentangling anthropogenic inputs from natural heterogeneity: An adaptive machine learning-Geostatistics workflow for soil heavy metal risk assessment

The background value is crucial in determining anthropogenic inputs and risk level of soil heavy metals (HMs). However, the spatial heterogeneity of the HMs frequently cause errors in the background value calculation, which further lead to underestimation or misjudgment of potential environmental risks. Here, we developed a sequential adaptive workflow utilizing expectation-maximization (EM) and Gaussian Process Regression (GPR) to eliminate spatial heterogeneity and improve the quality of HM risk assessments. This workflow operates through sequential diagnostic and application stages. It first conducts spatial variation driving factor analysis, which then guides the spatial heterogeneity identification of background values. This workflow integrated spatial variation driving factor analysis and spatial heterogeneity identification of background values for risk assessment. Spatial variation driving factor analysis is based on source apportionment and multiscale geographically weighted regression. Spatial heterogeneity identification is based on the correlation between the HMs background and the soil chemical composition. Methodological research was conducted using 3916 intensive sampling data points involving a dataset of element concentration data for 28 components. This workflow improves discrimination the HMs background from the anthropogenic inputs in the dataset, and applies the background values to risk identification and environmental bearing capacity assessment. To validate the proposed method, it was compared with the traditional methods. Comparison results demonstrated that the proposed method more effectively enhance the accuracy and reliability of regional environmental risk assessments by eliminating the spatial heterogeneity of the HMs background. The applicability analysis indicates that this workflow can be applied to geologically complex regions if sufficient sample quantities and analysis items can be obtained. The findings provide a robust tool for HMs background determination and offer accurate information and reliable evidence for soil pollution remediation projects. • A sequential adaptive ML-geostatistics workflow was developed to decouple soil HM backgrounds. • Spatial heterogeneity of backgrounds was eliminated via multi-source data integration. • Static background thresholds were proven to cause significant risk misjudgments. • Application significantly improved risk quantification and environmental assessment.