Unlocking hidden Carbon: A likelihood-based spatial framework to reduce SOC measurement uncertainty

Problem: The estimation of soil organic carbon (SOC) using biogeochemical models such as RothC and Century is a key component in the verification of soil carbon credits. However, current methodologies and approaches [e.g., "measure and model" (QA1)], apply uncertainty propagation models that do not explicitly account for spatial and temporal autocorrelation in SOC stocks. This omission leads to an overestimation of uncertainty and excessive deductions in carbon credits. Given that credit issuance is directly related to reported uncertainty, the current approach may lead to unnecessary financial losses for project developers and farmers.

Objective: To tighten SOC stock measurement uncertainty we develop and test a likelihood-based spatial framework that explicitly models spatial dependence in RothC-predicted SOC change.

Methods: Spatial dependence is represented with a Matérn covariance function whose parameters are estimated by maximum likelihood directly from RothC outputs. To avoid artefacts we introduce coordinate perturbation (±0.2 m for training, ±5 m for testing). We apply spatially coherent K = 5 fold cross-validation (clustered in geographical space) and repeat the split 100 × K times via Monte Carlo (total = 500 replicates). Model performance is summarised with RMSE relative to a fixed-effects (no-spatial-structure) model.

Results: Across 500 spatially coherent Monte-Carlo folds, layering a Matérn random field on top of the RothC grid cut the cross-validated RMSE from 21.04 ± 2.35 to 19.42 ± 1.27 t C ha⁻¹—an average improvement of 7.7 % (95 % CI 6.99–8.44 %). The fitted correlation range (~30 ± 20 km) and modest nugget variance (<200 (t C ha⁻¹)²) echoes the meso-scale gradients already present in the input layers—the field-observed SOC stock—showing that the process-based RothC model faithfully carries this spatial pattern through to its predictions. 

Conclusions: By explicitly modelling spatial autocorrelation with a likelihood-based mixed model, we trim prediction error by 7–8 %, offering a practical pathway to reduce SOC measurement uncertainty. Because credit issuance is linearly penalised by uncertainty, this precision gain can recover thousands of credits per project and improve market confidence. Importantly, the framework achieves these benefits without imposing an artificial structure; it simply recovers the inherited spatial structure preserved in RohC predictions.

Implications or Significance: The framework advances the MMRV (Measuring, Monitoring, Reporting & Verification) agenda by pairing deterministic biogeochemical models with modern geostatistics. Future work will extend this approach to a hierarchical Bayesian setting and explore higher-resolution, point-based SOC measurements to further diminish microscale noise.