Comparing uncertainty quantification methods for Random Forest-based digital soil mapping

Abstract. Uncertainty quantification is essential for using machine learning-based soil maps in decision-making, but the uncertainty communicated to users can depend strongly on the method used. This is especially relevant for soil organic carbon (SOC), where spatial predictions are used in carbon accounting, land management, and climate-related reporting. This study compares five uncertainty quantification methods within a shared Random Forest-based SOC modelling workflow for Estonia: 1) Quantile Regression Forest, 2) Conformalized Quantile Regression, 3) land use-conditional Conformalized Quantile Regression, 4) Random Forest-based calibration of conformity scores, and 5) land use-conditional split conformal prediction. Using 1004 SOC observations and 16 environmental covariates, the study evaluates how these methods differ in coverage, sharpness, land use- and SOC-specific calibration, environmental patterns of high uncertainty, and relationship with Area of Applicability.
All tested methods produced reasonably good marginal coverage for 90% prediction intervals, but differed in their local behaviour and sharpness. The results show that decomposing prediction interval diagnostics by land use and predicted SOC range provides a more informative view of uncertainty quality. High uncertainty was linked to specific environmental covariate combinations, suggesting that interval-based uncertainty patterns could help identify conditions where additional sampling may be useful. Prediction interval width and Area of Applicability were only weakly related. The main contribution of this study is a structured diagnostic framework for comparing prediction interval-based uncertainty methods at both marginal and more local levels, while also improving our understanding of how interval-based uncertainty relates to environmental conditions and model applicability in Estonia’s SOC model.