Estimation of spatial distribution of heavy metals in soils of mining areas based on ensemble learning and...

Accurate prediction of soil heavy metal concentrations is crucial for environmental management in mining watersheds. Traditional machine learning models cannot fully capture the spatial heterogeneity of soil contamination. This study proposes a novel Lightweight Gradient Boosting Machine-Geographically Weighted Regression (LGGWR) model that combines the lightweight gradient boosting machine (LightGBM) with geographically weighted regression (GWR). The model uses seven environmental covariates and incorporates the Risk Factor for Soil Index (RFSI) technique to predict heavy metal concentrations. LGGWR improves prediction accuracy by accounting for spatial heterogeneity and nonlinear relationships. Results show that after cross-validation, the coefficients of determination (R²) for five heavy metals (Lead (Pb), Zinc (Zn), Copper (Cu), Nickel (Ni), and Cadmium (Cd)) are 0.84, 0.71, 0.76, 0.86, and 0.84, respectively. Compared to random forest (RF), extreme gradient boosting (XGBoost), ordinary kriging (OK), GWR, and LightGBM models, LGGWR improves R² by 30.95-42.09 %. The study is conducted in the Baiyin mining area, Gansu, China, using 227 samples from mining and agricultural lands, along with 14 control samples. This study provides a fundamental data source for environmental management and targeted soil remediation strategies.