Abstract
Heavy metal enrichment in agricultural soils can affect crop safety, ecosystem functioning, and long-term land productivity, yet farm-scale screening is often constrained by limited routine monitoring data. This study develops a GIS-based framework that combines field-scale spatial analysis with explainable machine learning to characterize and predict heavy metal enrichment on an intensively managed cereal farm in eastern Kazakhstan. Topsoil samples (0 to 20 cm) were collected from 34 fields across eight campaigns between 2020 and 2023, yielding 241 composite field–campaign observations for eight metals (Pb, Cu, Zn, Ni, Cr, Mo, Fe, and Mn) and routine soil properties (humus, pH in H2O, and pH in KCl). Concentrations were generally low but spatially heterogeneous, with wide observed ranges for several elements (for example, Pb 0.06 to 2.20 mg kg−1, Zn 0.38 to 7.00 mg kg−1, and Mn 0.20 to 38.0 mg kg−1). We synthesized multi-metal structure using an HMI defined as the unweighted mean of z-standardized metal concentrations, which supported field-level screening of persistent enrichment and emerging hot spots. We then trained Extreme Gradient Boosting models using only humus and pH predictors and evaluated performance with field-based spatial block cross-validation. Predictive skill was modest but nonzero for several targets, including HMI (mean R2 = 0.20), indicating partial spatial transferability under conservative validation. SHAP analysis identified humus content and soil acidity as dominant contributors to HMI prediction. Overall, the workflow provides a transparent approach for field-scale screening of heavy metal enrichment and establishes a foundation for future integration with satellite-derived covariates for broader monitoring applications.