Managing Uncertainty in Agricultural Soil Carbon Project

How to manage the different sources of errors in an agricultural soil carbon project

For a project developer who wants to start an agricultural carbon project, one of the most difficult questions is also one of the most important: 

how much has soil carbon (SOC) actually changed thanks to the adoption of a better management practice?

At first sight, the answer seems straightforward. Take soil samples, measure the carbon stock, repeat the measurement later, and calculate the difference.

In practice, it is much more complicated.

Soil carbon is highly variable in space, while the changes we are trying to detect—particularly over short periods such as one year—can be relatively small. This creates a fundamental challenge for carbon projects: the uncertainty in the measurement can be as large as, or even larger than, the change we are trying to quantify.

Different approaches can be used to address this problem. A project can rely on measure-and-remeasure sampling, use sampling together with Digital Soil Mapping (DSM), or combine one of them with a dynamic biogeochemical model (measure-and-model).

All three approaches use measurements, but the difference is what they do with those measurements. The introduction of a biogeochemical model in the process, in particular, changes the perspective of the sampling scheme.

Instead of asking:

“Can we directly measure the change at enough locations to characterize the entire project?”

The question is:

“Can we measure enough strategically selected locations to understand the system, and then use a validated model to predict carbon dynamics across the wider project?”

Before talking about the advantages and challenges that affect the uncertainty, let’s not forget also other advantages the use of a biogeochemical model has to offer:

• No need for baseline control sites for the SOC counterfactual
• Estimation for SOC quantification does not require sampling (e.g., every year)
• Potentially less intensive soil sampling, given the different goals
• Better suited to large, heterogeneous projects
• More flexibility for changing management and crop rotations

For our discussion, introducing a model inevitably introduces a new source of uncertainty, the model prediction error (MPE). Yet, if the model is appropriately designed, calibrated and validated, that additional source of error can be more than compensated for by reducing other sources of uncertainty, particularly sampling error and measurement/input-data error. Let’s examine this by first briefly describing the sources of uncertainty.

Measurement error: every number has a margin

This error captures the uncertainty of values measured for SOC content, bulk density, soil texture, and management data, where applicable. These errors cannot be considered intrinsically low, especially if not deriving from physical measurements but from estimations (for example soil spectroscopy techniques for measuring SOC content).

In particular, the management data can be an important source of errors if they are incomplete and need to be gap-filled using secondary sources (regional or national statistics). Examples of data that could require a gap filling are fertilizations amounts and number of applications.

Sampling error: measuring only part of the project

No one measures every square meter of the project area, so initial soil measurements are taken at a series of sampling points. The uncertainty therefore depends partly on how representative those sampled points are of the wider project area. Stratification and good sampling point schemes can reduce this component of uncertainty. VERRA VM0042, for example, explicitly references stratification and describes the criteria to follow in defining the different strata of the project.

Model prediction error: the “price” of using a model

A biogeochemical model is an imperfect representation of reality. Model prediction error can arise from uncertainty in model parameters or from structural errors, where the model does not fully represent the actual biogeochemical processes occurring in the soil.

The impact of the Model Prediction Error

All these sources of errors (so of uncertainty) must be combined (“propagated”) to obtain the final value of uncertainty for the project. So now we come back to the topic of this discussion: find the best trade-off balancing the error introduced by the model (MPE) trying to lower the other two sources of error.

Measurement error

As described by VERRA protocol VM0042: “In many cases, the impact of these measurement errors on the error of estimated reductions and removals is assumed to be captured in Model Prediction Error”. This concept has also been expressed in the recent scientific literature (for example “Bayesian uncertainty analysis of soil organic carbon stocks and stock changes from croplands in the U.S. Midwest“, Gurung et al. 2026).

So, on which occasions is that “in many cases” assumption not valid? Generally, when your sources of data are not primary. For example, as described above, in case of gap filling.

Sampling error

This component of the uncertainty can only be influenced by the use of a model indirectly in two ways:

• Helping define a sampling scheme emphasizing the processes triggered by the different agronomic practices
• Simulating and determining the uncertainty at field level, potentially ensuring a better description of the dynamics for the entire project

Both of these aspects can contribute to lowering the sampling error, leveraging the capability of the model to better predict the evolution of the SOC produced by the implementation of different practices.

Ultimately, the purpose of introducing a model is to obtain a more accurate evaluation of the SOC changes and, at the same time through the simulation of the agronomic processes, lower the influence of the other sources of error.

How Armosys Has You Covered

As outlined in the previous sections, determining uncertainty in such a complex scenario can be a daunting task, particularly because introducing a model adds another layer of complexity.

That is why at Armosys,uncertainty management starts before the model is run. 

We continue to calibrate and validate, and improve our model across different scenarios, climate regions and agronomic practices to lower the model prediction error.

At project level, uncertainty can then be quantified by propagating the relevant sources of uncertainty through the modelling framework.  This allows project developers not only to report a final uncertainty value, but also to understand where that uncertainty comes from and which components can be improved through better data, sampling or model validation. 

Join the conversation on LinkedIn