How Large Should Your Soil Carbon Sampling Campaign Be?

A sustainability team calls about a soil carbon baseline, usually because a Scope 3 removal report or a supplier programme needs a defensible starting number. The second question is almost always the same: how many samples will we need? It is a reasonable question. It is also the wrong one.

Sample size is a consequence, not a starting point. It is the output of four upstream decisions: the scale of inference (field vs. project vs. programme), the variance of the stock across that scale, the minimum detectable difference you care about, and the confidence you need. Lock those down before the budget conversation, not after. Campaigns that skip this step arrive in year five with data that cannot resolve the effect everyone hoped it would show, a pattern now documented in high-density sampling data from 45 commercial cropland fields1.

Scale of inference: the decision that reshapes the budget

Before any formula, fix the scale. A baseline report for a single supplier farm is a field-scale precision problem: you want a tight confidence interval on the mean stock now. A 500-farm sourcing programme claiming Scope 3 removals is a project-scale change-detection problem: you want to distinguish a sequestration signal from natural noise across many fields over time. The two problems have different formulas, different data requirements, and very different price tags.

Bradford and colleagues tested this empirically on 45 commercial cropland fields and found that individual-field detection of meaningful accrual rates (around 3 tonnes C per hectare over ten years) is unreliable even at sampling densities of 1.2 hectares per sample, marked apparent gains and losses appear where no real change exists1. At project scale, by contrast, 30 field pairs at the same density put the mean estimate of stock change within one tonne C per hectare of the true value in 83 runs out of 100. Potash et al. (2025) made the economic case that sampling only around 10 percent of fields in a project of several thousand, with paired resampling, delivers a competitive return on the whole project investment within five years, provided the accrual rates reported in the literature hold on commercial farms2. The takeaway is stark: if your design assumes every field must individually prove its own sequestration, it is the wrong design.

Three numbers that drive sample size

Once the scale is fixed, three statistical inputs drive the calculation. None is optional, and none can be guessed without cost.

Spatial variance (σ²)

Soil carbon is heterogeneous at every scale. In temperate cropland and grassland, Poeplau and colleagues documented plot-scale mean absolute errors of 5.1 and 7.6 tonnes C per hectare at 0–30 cm (7.5 and 8.5 percent of stock) even when resampling the same profile with a minimal positional shift, and found that spatial dependence is weak3. Weak spatial dependence limits how much stratification can help in relatively uniform fields. In tropical perennials, variance is typically higher. For cocoa in Ghana and Côte d'Ivoire, published standard deviations for 0–30 cm SOC stocks cluster around 10 to 14 tonnes C per hectare4, driven by hillslope position, shade architecture, and chronosequence age.

In design, σ² is either taken from comparable published work on similar soil and management, the tropical SOC literature is finally rich enough for cocoa, coffee, and oil palm to do this honestly, or estimated from a 20 to 30 core pilot. Do not skip the pilot: the sensitivity of the final n to σ² is quadratic.

Minimum detectable difference (MDD)

The MDD is the smallest change the campaign must distinguish from noise. For a baseline it is the precision of the stock estimate itself (for instance ±3 tonnes C per hectare at 95 percent confidence). For change detection it is the smallest annual or cumulative change that matters for the claim (for instance 0.5 tonnes C per hectare per year cumulated over five years). Halving the MDD quadruples the sample count; it is the most punishing relationship in the design.

A warning follows. Credible MDDs for cocoa and coffee agroforestry on a 5 to 10 year horizon are in the range 2 to 5 tonnes C per hectare total. The temperate evidence sets the scale: across five silvoarable agroforestry sites in France, mean SOC accrual in the top 30 cm was 0.24 tonnes per hectare per year, with a site range of 0.09 to 0.465. Designs targeting less than 1 tonne per hectare of cumulative change are not detecting management effects; they are detecting rainfall. The 2023 Australian grazing-credit issuance dissected this exact failure mode: roughly 250,000 tonnes of credits were issued largely because of a Decile-10 wet year, not because of management change6.

Confidence (α) and power (1 − β)

α is the tolerance for a false positive; β for a false negative. Conventional values are α = 0.05 and power = 0.8. Two formulas are in use, and they are not interchangeable. For a baseline precision design, the relevant form is n ≈ (z_α/2 · σ / ε)², where ε is the half-width of the desired confidence interval. For a paired change-detection design across two time points, the relevant form is the two-sample power formula below. The original framing of this article mixed the two; both are shown here with their appropriate use cases.

For the change-detection form with α = 0.05 and power = 0.8, (z_α/2 + z_β)² ≈ 7.85. In either case, variance and MDD dominate everything, and they trade off against each other in real, calculable ways.

Stratification: the variance multiplier that usually (but not always) pays for itself

Stratified sampling cuts the samples needed for a target precision, but by less than the design brochures suggest. Neyman allocation sizes each stratum by its own variance rather than by its area, and should outperform the area-proportional allocation used in most published field trials7. Across eight commercial Midwest fields, stratifying on remote-sensing, topographic, and soil-survey covariates cut the required sample size by 17 percent on average, with a 95 percent interval of 11 to 23 percent8. The gain depends on whether the strata actually capture variance: where the covariates explained little of the within-field variation, Potash and colleagues found it modest, under 5 percent at the weakest site, and where they explained most of it the saving reached about 45 percent8. In cocoa landscapes where topography, shade architecture, and chronosequence drive large differences in SOC, stratification is transformative: Adiyah et al. (2022, 2023) show significant hillslope-position effects4 and, over matched 15-year intervals, SOC gains that differ roughly threefold to fourfold between age classes, from 13.7 to 46.9 tonnes C per hectare at 0–60 cm9.

A defensible stratification combines three layers: remote sensing (canopy and bare-soil signals from Sentinel-2, with Sentinel-1 backscatter where moisture or structural variance matter), topography and drainage from a DEM, and a reconstructed management history of the preceding decade, now routinely recoverable from grower records or time-series imagery. Two cautions. First, stratum boundaries are themselves a source of variance when the covariates are noisy; inspect stratum means from the pilot before locking allocation. Second, Verra's VM0042 v2.2 , ICVCM-endorsed under the Core Carbon Principles in October 2025, sets a minimum of 15 samples per stratum regardless of what Neyman would allocate, and does not yet accept digital soil mapping as an approved CCP method. If the project intends to credit under VM0042 or the forthcoming EU CRCF methodologies, align the stratification with those constraints from day one.

A worked example, grounded in published cocoa data

A 500-hectare cocoa plantation in southern Côte d'Ivoire. Baseline target ±3 tonnes C per hectare at 95 percent confidence on a mean 0–30 cm SOC stock of roughly 45 tonnes C per hectare, in line with Adiyah et al. (2022) for Ashanti hillslope cocoa4. Pilot sampling (25 cores) yields σ ≈ 11 tonnes C per hectare, consistent with the published literature rather than an assumed value.

Stratified into three management zones (young replant, mature stand, recent renovation) with within-stratum σ of 7, 8, and 9 tonnes per hectare, Neyman allocation reduces the requirement to roughly 28 to 30 cores, a ~45 percent saving. If VM0042 v2.2's minimum of 15 samples per stratum applies, the count climbs back to 45, essentially erasing the stratification gain at this farm size. This is a real-world tension practitioners should confront explicitly rather than paper over.

For change detection, whether measurement pays at all depends on project size, and the threshold sits far higher than most programmes assume. At an average treatment effect of 0.3 tonnes C per hectare per year, Potash and colleagues find measure-and-remeasure economically feasible only above roughly 5,000 fields at 40 dollars per tonne; 250 fields need a 100 dollar price, and 100 fields need at least 2002. The structural point is separate, and better established: project-level paired resampling across many farms beats a heroic effort on one. Bradford's results make it explicit: insisting on per-farm detection with 60 to 100 paired cores per time point will still produce unreliable per-farm estimates, while 30 field pairs sampled at 1.2 hectares per sample recover the project mean within one tonne C per hectare in 83 runs out of 1001. That design is defensible, but it costs several times the per-farm coring effort rather than the same. This is the shift in framing the 2023–2025 literature has forced on MRV design.

Two measurement-stack upgrades that reshape the economics

First, sample to 0–60 cm in at least three increments (0–10, 10–30, 30–60). Adiyah et al. (2023) measured SOC gains at 20–60 cm that exceeded topsoil gains in every matched 15-year interval in Ghanaian cocoa agroforestry, 34.9 against 8.5 tonnes C per hectare in the youngest pair and 40.3 against 6.6 in the oldest9. A 0–30 cm design misses most of that: on the paper's own depth increments, roughly 60 percent of the 0–60 cm gain sits below 30 cm. Raffeld et al. (2024) reach the same conclusion for annual row crops, where sampling to 60 cm in multiple increments is their recommended practice, and note that deep-rooted perennial and agroforestry systems may need to be sampled deeper still10. For deep-rooted perennials, subsoil sampling is no longer optional.

Second, use Equivalent Soil Mass (ESM), not fixed depth, for stock-change calculations. Fowler and colleagues show that fixed-depth accounting underestimates stock change by about 17 percent for a realistic drop in bulk density, from 1.5 to 1.1 grams per cubic centimetre, and that in their worked case it reverses the sign entirely, reading a true 14.8 percent gain as a 16.2 percent loss11. The equivalent-mass correction itself is not new: Ellert and Bettany set it out three decades ago12, and von Haden and colleagues have since shown how far depth-based comparisons drift once bulk density changes13. Raffeld and colleagues put field numbers on it in two long-term trials, where bulk density moved under no-till and under organic amendment and fixed-depth accounting reversed the sign of the stock change10. Verra's forthcoming VM0042 v3.0 Soil Sampling and Analysis Handbook (in public consultation through March 2026) makes ESM the default. A 2026 article (or project) that ignores ESM is a credibility flag.

Counterfactuals and climate confounding: why controls are non-negotiable

A change-detection design without a counterfactual does not measure management, it measures weather. On a southern Queensland grazing trial, SOC rose 7.7 tonnes C per hectare at 0–100 cm over 2.5 wet years, or 3.1 tonnes per hectare per year, and half of the 0–30 cm gain was gone once rainfall returned to average. That rate is five times the 0.6 tonnes per hectare per year the authors set as a realistic 25-year upper bound for Australian grazing land, and ten times their medium scenario of 0.36. Any credible change-detection campaign therefore needs either a co-located control (paired fields under baseline management, sampled on the same dates) or an explicit climate normalisation using a biogeochemical model calibrated locally. A third-party verifier in 2026 will ask which one the project used. “Neither” is not an answer that survives.

What the 2026 MRV landscape actually requires

Three frameworks now translate the general principles of soil-carbon measurement, reporting, and verification 15 into rules any serious SOC campaign must follow. Verra VM0042 v2.2, ICVCM-endorsed under the Core Carbon Principles in October 2025, specifies stratification rules, minimum samples per stratum, uncertainty deductions, and ESM-based accounting; v3.0 with the Soil Sampling and Analysis Handbook is expected during 2026. EU Regulation 2024/3012 (CRCF) entered force in December 2024; carbon-farming certification methodologies arrive through 2026, with the QU.A.L.ITY criteria (quantification, additionality, long-term storage, sustainability) shaping how EU-facing claims will be audited. The GHG Protocol Land Sector and Removals Standard takes effect 1 January 2027, requires a quantitative uncertainty estimate and a confidence interval for every reported removal, and requires removals on productive agricultural land to rest on empirical data specific to the carbon pools, with resampling by consistent methods at least every five years16. The estimate may come from measurement, from a calibrated model, or from a hybrid of the two, but the field data behind it are not optional. SBTi FLAG aligns with these.

Design the campaign to the standard the buyer will be audited against, not the one most familiar to the soil lab.

The punchline

The right sample size is whatever makes the claim survive the scrutiny it actually faces. For a baseline on a single cocoa farm with realistic σ, that is typically 30 to 60 cores with proper stratification, 0–60 cm depth, and ESM accounting. For a credible multi-year sequestration claim, it is a project-level paired design across many farms, not a heroic effort on one. Proximal sensing has changed the arithmetic. Ramírez and colleagues costed one FT-IR setup against dry combustion for a grassland project of 400 to 1,000 samples, and report throughput rising from 4 to 12 samples per hour and cost per sample falling by about 2.5 times, with annual instrument maintenance running the other way17. Hutengs and colleagues show that portable VisNIR and mid-infrared instruments can be calibrated for field-condition soils once moisture and particle-size effects are removed, which is what makes scanning in the field rather than the laboratory realistic18. That saving should be spent on more samples, deeper profiles, and more farms, not on hitting the same old number for less money.

Key takeaways