Geographically weighted random forest fusing multi-source environmental covariates for spatial prediction of soil...

Area of Science:

• Environmental Science
• Geospatial Analysis
• Soil Science

Background:

• Efficient spatial prediction of soil heavy metals is vital for ecosystem health, agriculture, and food security.
• Traditional machine learning models often neglect spatial autocorrelation, limiting predictive accuracy.
• Soil contamination by heavy metals poses significant environmental and health risks.

Purpose of the Study:

• To develop and evaluate a geographically weighted random forest (GWRF) model for predicting soil heavy metal distribution.
• To compare GWRF performance against geographically weighted regression (GWR) and global random forest (RF).
• To identify key environmental factors influencing soil heavy metal concentrations.

Main Methods:

• Employed GWRF, a spatial extension of RF, integrating terrain, air quality, vegetation, soil properties, human activity, and hyperspectral data.
• Compared GWRF with GWR and an optimized global RF model.
• Utilized interpretable machine learning (MDI and SHAP) to identify influential factors and analyzed bandwidth/weight parameter effects.

Main Results:

• GWRF achieved superior prediction accuracy for Cr, Pb, As, and Hg, with R² values ranging from 0.303 to 0.561.
• GWRF significantly outperformed GWR (up to 209.18% improvement) and global RF (up to 53.28% improvement).
• Air quality (O₃), topography, mine proximity, and rainfall were identified as key influencing factors.

Conclusions:

• GWRF is a highly effective method for accurate and interpretable spatial mapping of soil heavy metals.
• Optimizing GWRF parameters and integrating local/global model results enhances prediction accuracy.
• The findings provide valuable insights for environmental risk assessment, pollution source tracing, and land management.