<p>Elevated heavy metal concentrations in farmland soils pose a serious threat to crop yield, quality, and human health, necessitating rapid and accurate monitoring for effective pollution control. This study investigated heavy metal pollution (HMP) in farmland soils across Ningxia, China, using five tree-based models to predict concentrations of eight heavy metals. The SHAP method was applied to elucidate the role of environmental variables in determining the concentrations of eight heavy metals, and a combined modeling approach was developed to predict the Nemero Pollution Index (NPI). Results showed that the Random Forest (RF) model attained a superior level of accuracy in predicting arsenic (As), copper (Cu) and zinc (Zn). However, the LightGBM and XGBoost models performed best for cadmium (Cd), chromium (Cr), mercury (Hg), as well as lead (Pb) and nickel (Ni), respectively. Among the input features, soil parameters constituted the most significant predictors, with vegetation indices, land use percentages, and landscape pattern index, with topographic data contributing the least. For NPI prediction, a combined model integrating RF and XGBoost further improved performance (<i>R</i>² = 0.4761). However, incorporating LightGBM with other models introduced noise and reduced predictive stability. In summary, specific tree-based models showed high accuracy for particular heavy metals under consistent feature inputs, while the RF + XGBoost combination proved most effective for NPI prediction. These findings provide a novel and interpretable framework for predicting and evaluating HMP in farmland systems.</p> Graphical Abstract <p></p>