Improved prediction of heavy metal concentration in typical agricultural soil in Hainan by Machine Learning method

As machine learning (ML) techniques continue to evolve, researchers are becoming more dedicated to applying these methods to model and predict heavy metals (HMs) in soil. Unfortunately, many existing studies neglect the impact of spatial stratified heterogeneity and parent material in soil. In this study, Spatial Stratified Heterogeneity Index (SHI) and Parent Material Index (PMI) were used in 10 ML models. The results showed that SHI enhanced the accuracy of all models, achieving accuracy rates between 0.706 and 0.954. The concentration of HMs in topsoil is greatly affected by the PMI of soil characteristics (SC). while 8 out of those 10 ML models whose environmental factors that did not include SHI, the model accuracy was decreased to 0.061 ∼ 0.388, and the PMI still contributed to 2.3 % and 40.1 % of the relative importance among all driving factors. In different land use types, the paddy fields had significantly lower concentration of chromium (Cr), copper (Cu), and nickel (Ni) compared to the dryland, other gardens and orchards. In contrast, lead (Pb) levels were significantly higher than in the those three types. This study provides actionable insights for ML model combined with SHI and PMI in environmental monitoring and risk management.