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

Efficient spatial prediction models for soil heavy metals are crucial for maintaining soil ecosystem health, promoting high-quality regional agriculture, and national food security. Traditional machine learning (ML) models often overlook spatial autocorrelation, a limitation that reduces their predictive accuracy. This study employed geographically weighted random forest (GWRF), a spatial extension of random forest (RF), incorporating terrain, air quality, vegetation, soil properties, human activity, and hyperspectral covariates to predict the distribution of Cr, Pb, As, and Hg in a city in eastern China. We compared GWRF with geographically weighted regression (GWR) and an optimized global RF. The GWRF model performed best, achieving test set R² values of 0.303 (Cr), 0.419 (Pb), 0.428 (As), and 0.561 (Hg). This represents a significant improvement over GWR (up to 209.18 %) and global RF (up to 53.28 %). Furthermore, this study investigated the influence of different bandwidth and weight parameters on GWRF performance. Results revealed that selecting an appropriate bandwidth and the suitable integration of local GWRF and global RF model results can enhance prediction accuracy. Interpretable ML analyses (MDI and SHAP) identified air quality (especially O₃), topography, distance to mines, and rainfall as key influencing factors. The GWRF-generated spatial distribution maps showed high-concentration areas mainly in the northwestern plains and south-central regions, with distinct patterns for different metals. This research confirms the effectiveness of GWRF for high-accuracy, interpretable spatial mapping of soil heavy metals, offering new insights for environmental risk assessment and pollution source tracing.