The main methodologies governing the voluntary carbon market (VCM) and Scope 3 accounting now allow the use of Digital Soil Mapping (DSM) for soil organic carbon (SOC) stock estimation. Verra’s VT0014 has quickly become the industry standard for DSM-based SOC stock estimation and provides step-by-step guidance on uncertainty estimation. This guidance is largely based on Wadoux & Heuvelink (2023) and is applicable for a large class of models. However, the generality of the approach has the cost of suboptimality: uncertainty of the resulting SOC stock estimates is larger than necessary, which directly influences project economics.

In this work, we empirically study the uncertainty estimator from Wadoux & Heuvelink (2023) and compare it to a Random Forest Residual Kriging (RFRK) approach (Hengl et al., 2015; Szatmári et al., 2024). We also include the classical Horvitz-Thompson estimator (HTE) calculated from soil samples only. For a fair comparison, we apply the Wadoux & Heuvelink (2023) estimator on top of a Quantile Regression Forest (QRF) model that uses the same calibration data, covariates and hyperparameters as the Random Forest model from the RFRK approach. We benchmark all estimators across three agricultural projects. Special attention is given to the uncertainty of the estimators across differing project sizes and sampling intensities.

Both DSM approaches are compliant with VT0014 and passed the model validation requirements with comparable validation metrics. RFRK consistently showed slightly higher R² scores, with empirical coverage probabilities closer to the nominal coverage. Throughout the benchmark, RFRK was consistently more precise than HTE, reducing uncertainty of the SOC stock estimate substantially across all scenarios. In contrast, QRF improved over HTE only in the largest-sample project, while inflating the uncertainty in the other two. DSM uncertainty estimates showed diminishing returns as sample sizes increased, and dropped sharply as project area increased.

These findings confirm that a DSM-based SOC stock estimate can be substantially more precise than a classical soil sampling approach: RFRK roughly halved the uncertainty of the estimate, translating to reduced uncertainty deductions. However, a DSM-based estimate can also inflate uncertainties, despite being validated with state-of-the-art performance metrics, as demonstrated by the QRF-Wadoux approach. The exact choice of the DSM model appears to govern precision more than the model accuracy in terms of R² score. Therefore, project proponents should carefully select the most suitable modelling approach for their project.