Comparing the prediction performance, uncertainty quantification and extrapolation potential of regression kriging and random forest while accounting for soil measurement errors

Summary

This comparative study evaluated regression kriging (RK) and random forest (RF) algorithms for mapping three soil properties in Cameroon, explicitly accounting for measurement error in soil data. Regression kriging demonstrated superior predictive performance with 2–12% lower RMSE and better spatial extrapolation capacity, though random forest provided better local uncertainty quantification. The work addresses a significant gap in geostatistical and machine learning practice by incorporating measurement error sensitivity and extrapolation risk assessment.

UK applicability

The methodological framework for comparing geostatistical and machine learning approaches whilst accounting for soil measurement error is transferable to UK soil mapping contexts, particularly for national soil surveys and regional property prediction. However, the case study was conducted in Cameroon using proximal soil sensing methods; UK applicability would require validation against British soil types and climate conditions.

Key measures

Model Efficiency Coefficient (MEC), Root Mean Squared Error (RMSE), prediction uncertainties, cross-validation metrics, accuracy plots for uncertainty quantification

Outcomes reported

The study compared regression kriging and random forest models for predicting spatial distribution of soil pH, clay content, and organic carbon whilst accounting for measurement errors. Performance was evaluated using Model Efficiency Coefficient, Root Mean Squared Error, and cross-validation metrics, with assessment of uncertainty quantification and extrapolation capability.

Topic tags