Integrating soil properties, remote sensing indices, and machine learning for predicting soil heavy metal...

Soil heavy metal contamination has become a serious environmental problem particularly in industrial areas. This study investigates the potential of integrating soil physiochemical parameters and remote sensing (RS) datasets with machine learning (ML) models including random forest (RF), extreme gradient boosting (XGB), k-nearest neighbors (KNN), and support vector regression (SVR) to predict soil heavy metal concentrations. A total of 130 systematic soil samples were collected and analyzed for nine heavy metals (i.e., Cd, Co, Cr, Fe, Mn, Ni, Pb, V, and Zn) using ICP-MS, and fourteen different RS indices derived from Landsat 9 imagery. Among the ML models, RF and XGB revealed moderate performance, yielding the highest R² values for Fe and Cd reaching 0.52 and 0.41, respectively, and the lowest RMSE (0.15 and 0.14), respectively, across several metals. While KNN and SVR demonstrated weaker performance, the KNN generating the lowest R² value for Pb and Cr achieves 0.04 and 0.07, respectively. The underestimated and overestimated bias values among heavy metals were low ranged between -4.60 and 2.41%. Spatial prediction maps showed heterogenous contamination patterns throughout the study area and outlined hotspot zones across industrial region. The investigation suggests that ML models captured broad spatial contamination trends across the study area, although predictive performance varied among heavy metals. However, RS data alone cannot be used to predict heavy metal concentration. These findings offer valuable insights to inform environmental monitoring strategies and risk evaluation within industrial landscapes.