Integrated PMF-XGBoost analysis of heavy metal source contributions and associated predictor patterns in a polymetallic mining region.

Heavy metal (HM) contamination in mining-impacted soils poses persistent ecological and human health risks, yet the relationships between source contributions and associated environmental variables remain poorly understood. To address the limitations of previous studies, we used an integrated Positive Matrix Factorization (PMF) and Extreme Gradient Boosting (XGBoost) framework based on 55 surface soil samples to estimate HM sources and explore the geochemical and environmental variables associated with source-specific contamination variability within the studied polymetallic mining region in Central China. PMF resolved four potential sources: metallurgical emissions (26.7%), geogenic sources (55.3%), combustion-derived deposition (8.1%), and mining-impacted geochemical reactivation (9.9%). The interpretation of the XGBoost models using SHapley Additive exPlanations (SHAP) further ranked Zn as the leading predictor associated with the metallurgical emission factor, with Mn and Cr as consistent secondary predictors. The geogenic factor was primarily associated with Mn and Zn, with Zn showing an inverse model pattern and Cr providing further contribution. Cr was also the dominant variable in the combustion-derived deposition factor. Ni and Cr together were the leading predictors of post-depositional reactivation. Repeated modeling under independent random partitions indicated that these predictor importance rankings were stable, increasing confidence in their reproducibility within the dataset. Spatial validation indicated acceptable within-region predictive utility for all four factors and no significant residual spatial autocorrelation, although the predictive performance varied among the factors. These exploratory findings suggest that, within the study region, variations in PMF-resolved factor contributions may be associated not only with individual metal concentrations but also with model-based interaction effects among metals and soil properties.