Ensemble learning-based applied research on heavy metals prediction in a soil-rice system

Area of Science:

• Environmental Science
• Ecology
• Agricultural Science

Background:

• Heavy metal accumulation in soil poses risks to ecosystems and food safety.
• Accurate prediction of heavy metal concentrations (HMC) is essential but challenging.
• Traditional models often lack the accuracy and robustness needed for effective management.

Purpose of the Study:

• To develop and evaluate advanced prediction models for HMC in soil-rice systems.
• To assess the performance of ensemble learning (EL) techniques, specifically Random Forest (RF) and Gradient Boosting Machine (GBM), against benchmark models.
• To investigate the robustness and practical applicability of EL-based HMC prediction models.

Main Methods:

• Constructed a dataset of 490 multidimensional environmental covariates.
• Developed EL-HMC models (RF-HMC, GBM-HMC) and compared them with Multiple linear and Bayesian regressions (BMs).
• Utilized R², Mean Absolute Error (MAE), Root Mean Square Error (RMSE), sensitivity analysis, and spatial autocorrelation (SAC) for evaluation.

Main Results:

• EL-HMC models demonstrated significantly higher accuracy (R² increase of 48.0% for soil Cd, 58.2% for rice Cd, Pb, Cr, Hg) compared to BMs.
• RF-HMC and GBM-HMC achieved R² values of 0.654-0.690 for soil Cd and 0.618-0.850 for rice heavy metals.
• Sensitivity and SAC analyses confirmed the excellent robustness and stability of the EL-HMC models.

Conclusions:

• Ensemble learning technology provides practical and feasible prediction models for HMC with superior accuracy and stability.
• The developed EL-HMC models offer a new perspective for sustainable management and precise prevention of heavy metal pollution in farmland.
• This study highlights the significant application potential of EL technology in pollution ecology and environmental management.