Distinguishing Natural and Anthropogenic Cadmium Sources for Regional Pollution Zoning: A Two-Stage Machine Learning Framework

Widespread cadmium (Cd) anomalies in soil and associated pronounced spatial heterogeneity in high-geochemical background regions are primarily attributed to the superimposed effects of natural enrichment and anthropogenic disturbances. These superimposed effects hinder accurate identification and management of critical pollution zones. To address this issue, a pollution differential (PD) indicator was developed based on the DBSCAN algorithm to delineate the distinct pollution patterns arising from the superimposed effects. By integrating the DBSCAN-derived results with bivariate local indicators of spatial association analysis, critical pollution zones predominantly influenced by mining, manufacturing, and agricultural activities were effectively identified. The PD indicator was subsequently transformed into model-interpretable structured data, serving as an input that captures heterogeneity information for the subsequent prediction stage. Validation using an independent data set demonstrated that incorporating the PD indicator significantly improved model predictive performance and generalization, achieving a testing R 2 RMSE of 0.86 and 0.13, respectively. Considering the inconsistency in Cd content and its bioavailability, an anthropogenic activity-oriented spatial prediction of Cd distribution was further generated using the optimized model after excluding the influence of natural enrichment. The two-stage model developed in this study provides an enhanced framework for identifying critical pollution zones, thereby supporting more effective pollution prevention and control strategies in high-geochemical background regions.