Signal versus Noise: When a Soil Carbon Change Is Real

A buyer asks an arable farmer's advisor whether the cover-crop programme is sequestering carbon. The advisor reports +0.4 tonnes C per hectare per year over five years, with 60 paired cores per time point. The verifier asks for the confidence interval. It crosses zero. The claim is silently dropped, the season's MRV budget is written off, and the buyer goes back to the supply-chain team with nothing to show.

The pattern is now well documented. Bradford and colleagues (2023) tested the feasibility of empirical SOC change detection across 45 commercial cropland fields and found that individual-field detection of meaningful accrual (around 3 Mg C/ha over ten years) is unreliable even at sampling densities of 1.2 hectares per sample1. The signal exists. It is just not larger than the noise at the scale most field designs assume. The 2024 collapse of the headline-grabbing Australian grazing-credit issuance, in which roughly 250,000 ACCUs were issued largely because of a Decile-10 wet year rather than management change, made the same point at policy scale2.

This article reframes that problem. SOC change detection is a signal-to-noise problem, governed by an inequality with three sources of noise on one side and a small management signal on the other. The sample size, the time horizon, and the design choices fall out of that inequality; they are not the starting point.

Anatomy of the noise floor

SOC measurement noise is the variance of the estimated stock change around its true value. It stacks from three sources, and the largest is rarely the one practitioners worry about.

Spatial variance within the inference unit

Soil carbon is heterogeneous at every scale. Within a single homogeneous-looking arable field, Poeplau and colleagues (2022) document plot-scale standard deviations of 5 to 8 Mg C per hectare on 0–30 cm stocks even when resampling the same profile with a minimal positional offset, a coefficient of variation around 10 percent that is mostly irreducible3. Across heterogeneous landscapes, where slope position, parent material, and management history vary, CVs of 30 to 60 percent are routine14. In tropical perennials, published standard deviations for cocoa agroforestry 0–30 cm stocks cluster around 10 to 14 Mg C per hectare, driven by hillslope position, shade architecture, and chronosequence age5.

Spatial variance is the noise source most projects under-budget for, because the pilot pass typically samples a few apparently uniform hectares. The variance unfolds when the design extrapolates to the full management unit.

Temporal and seasonal variance

Even at the same point on the same farm, repeat sampling at the same depth on different dates returns different values. Two effects dominate. First, soil moisture and the freshly mineralised organic-matter pool fluctuate with rainfall, which is why the Australian credit issuance disappeared in the year after triple-La-Niña conditions ended2. Second, sampling depth in the field is a stochastic operation, not a fixed plane: a few centimetres of difference between baseline and resample can swing 0–30 cm stocks by several Mg C per hectare when bulk density shifts under reduced tillage or organic amendment. Wendt and Hauser (2013) introduced the equivalent soil mass (ESM) correction precisely to absorb this artefact, and Raffeld et al. (2024) show that fixed-depth versus ESM choice can swing inferred change by more than the management signal at the same raw data67.

Analytical and processing variance

The lab adds variance too. Even at a single accredited lab, replicate measurements on the same sample carry a CV of 1 to 3 percent for total carbon by dry combustion. Cross-lab and cross-method variance (dry combustion vs. Walkley-Black, mid-infrared (MIR) calibration drift, sample preparation differences) is larger and rarely tracked through to the headline. Heuvelink and colleagues (2021) make the case that prediction variance from machine-learning models can dominate a final stock-change estimate at fine spatial scale, and that without explicit propagation of all three variance terms, a reported SOC change number is structurally uninterpretable8.

In a typical cropland baseline, the three contribute roughly 80, 15, and 5 percent of the total. In a paired baseline-resample design where the same plots are revisited, σ²_spatial collapses (the between-plot component cancels), and σ²_temporal plus σ²_analytical become the dominant noise. The ratio of the two is the central design lever, and the reason paired designs are now standard in serious crediting regimes.

The signal we are trying to detect

The other side of the inequality is the management-driven signal. The literature is finally rich enough to ground realistic ranges, and they are smaller than most marketing decks acknowledge.

• Cover crops. Poeplau and Don's meta-analysis of 37 sites finds 0.32 ± 0.08 Mg C per hectare per year in topsoil9. Honest, real, small.
• No-till and reduced tillage. Apparent gains of 0.3 to 0.7 Mg C per hectare per year in 0–15 cm largely disappear when integrated to 0–40 cm. Powlson and colleagues (2014) showed this is principally a depth-redistribution artefact, not net sequestration10.
• Temperate agroforestry. Mayer et al. (2022) report a topsoil mean of 0.21 Mg C per hectare per year across 51 studies, with a between-study standard deviation of 0.79; hedgerows reach 0.32, alley-cropping 0.26, silvopasture slightly negative11. The cross-study SD is larger than the mean, which is the central problem.
• Cocoa agroforestry. Adiyah et al. (2022, 2023) document large variation by age class and slope position, with topsoil accrual rates of 0.2 to 0.5 Mg C per hectare per year and substantial subsoil gains in stands older than ~18 years5.
• Biochar and amendment. Discrete one-off events of 5 to 15 Mg C per hectare followed by slow attenuation; effectively a step rather than a slope.
• Adaptive multi-paddock grazing. Site-dependent and equilibrium-dependent; mean differences of order 0.5 to 1.5 Mg C per hectare per year where they exist, but with large between-site variance.

Two implications follow. The realistic management signal is in the 0.1 to 0.5 Mg C per hectare per year range for slope-style practices, smaller than the lab error on a single sample in absolute terms. And the cross-site SD of the effect, in the meta-analyses, is comparable to or larger than the mean, which constrains how much variance compression a single project can extract from prior literature.

The detectability inequality

For a paired baseline-resample design across two time points, the smallest mean change a campaign can statistically distinguish from zero is the minimum detectable difference (MDD):

with z_α/2 = 1.96 at α = 0.05 (two-sided) and z_β = 0.84 at 80 percent power, so (z_α/2 + z_β)² ≈ 7.85. σ_paired is the standard deviation of the per-plot change between time points (much smaller than the absolute σ of the stock if the same plots are revisited and within-plot autocorrelation is high). n is the number of paired plots.

Three properties of this inequality matter for design.

The MDD is symmetric. Halving it quadruples the required n. Doubling the σ also quadruples n. There is no smooth path from a campaign that detects 1 Mg C per hectare per year to one that detects 0.25; it costs sixteen times the sample.

σ_paired collapses where σ does not. This is why paired designs are now the verification-credible structure. A 60-plot paired design at σ_paired = 4 Mg C per hectare can detect a 5-year cumulative change of about 2 Mg C per hectare with 80 percent power. The same 60 plots, sampled independently at baseline and resample, would need σ closer to 2 Mg C per hectare to reach the same MDD, which is not achievable on real cropland.

The signal accumulates with time, not with sample count alone. A 0.3 Mg C per hectare per year accrual cumulates to 1.5 Mg C per hectare in five years and 3 Mg C per hectare in ten. The same n that fails to detect change at year five may comfortably detect it at year ten without adding a single core. This is the time-horizon paradox: the cheapest way to buy detection power is to wait.

Signal envelope simulator

The first diagram makes the inequality concrete. Choose an annual accrual rate, a paired standard deviation, and a number of plots; the chart draws the true mean trajectory and the noise corridor at ±ε, and marks the year at which the trajectory first clears the corridor. Pull σ down with a paired design and the detection year drops by half; push the accrual rate up and the trajectory clears earlier; cut n and the corridor widens until the trajectory cannot escape it inside 15 years.

The simulator visualises why most published “we sequestered X” claims are unverifiable in their own data: the sliders that resolve a credible signal sit outside what the field design actually paid for.

The time-horizon paradox

The implication of the MDD inequality is uncomfortable. For most regenerative-style practices, the honest detection window is 8 to 12 years, not the 3 to 5 years on which corporate roadmaps and crediting cycles are typically built112. Smith (2004) modelled this two decades ago: under a 20 to 25 percent uplift in carbon inputs, detection at 90 percent confidence requires roughly 6 to 10 years if the regime tolerates a 3 percent shift in background SOC, and is essentially never detectable on shorter cycles if the regime tolerates only 15 percent of background change12.

Three consequences follow for project design. First, accept measurement-and-modelling hybrid accounting for the early years rather than over-claim. Verra's VM0042 v3.0 consultation (the Soil Sampling and Analysis Handbook released March 2026) and the Potash et al. (2025) measure-and-remeasure economic case both move in this direction1314. Second, structure the contract around the verification window, not the marketing one: a 3-year report that says “below detection limit, signal trajectory consistent with model” is more durable than a 3-year report that says “+1.4 Mg C per hectare” with a CI that crosses zero. Third, take the depth question seriously. SOC gains in deep-rooted perennials older than ~18 years often exceed those in topsoil; a 0–30 cm design simply does not see them57.

Sampling design as the variance-reduction lever

If σ is fixed by the soil and waiting longer is constrained by the contract, the remaining lever is design. Three design moves compound, and a serious project uses all three.

Stratification. Allocating sampling effort to strata that capture variance (topography, management age, soil type) reduces effective σ and therefore n. de Gruijter et al. (2016) introduced the Ospats algorithm for cost-optimal stratified allocation using a prior SOC map, and showed the gain is substantial when prior information is good15. Empirically, follow-on work on the Bradford et al. data shows stratification on readily available covariates buys a mean 17 percent sample-size reduction (95% CI 11 to 23 percent) versus simple random sampling. Strong priors plus doubly balanced sampling can push that to 30 to 50 percent1. Note the floor: VM0042 v2.2 sets a minimum of 15 samples per stratum, which can erase the Neyman gain on small management units.

Paired baseline-resample. Returning to the same plots at resample collapses the between-plot component of variance. Typical paired-design noise reductions are 30 to 70 percent depending on the temporal autocorrelation of micro-site SOC. The USDA NRCS Soil Carbon Monitoring Network operationalises this at six samples per 4-hectare field on a paired-design schedule, roughly one core per 0.7 hectares16.

Composite-within-stratum, independent-across-stratum. Pool cores within a stratum to absorb micro-scale variance into a stratum-mean estimate at lower cost; keep strata as independent replicates so the test still has degrees of freedom. The opposite (pooling across strata to a single composite) is fast and cheap and useless for inference.

Sampling design dial

The second diagram lets the reader compose the three levers. A baseline simple-random design at user-set σ produces a years-to-detect estimate. Toggling stratification reduces effective σ by a literature-realistic factor. Toggling paired baseline-resample collapses the between-plot variance further. The chart updates the years-to-detect, the per-stratum count after the VM0042-style 15-floor is applied, and the resulting total core count.

The point of the dial is not to optimise to a single number. It is to surface where each design choice pays back, and where the verification floor (the 15-per-stratum constraint, the depth requirement, the ESM correction) costs more than Neyman saves.

What the model-data fusion turn changes

The most consequential shift in 2024–2026 has been the move from measurement-only or model-only crediting toward hybrid designs. The architecture is now well described: Tier-3 process-based or empirical models calibrated locally with paired measurement networks, with prediction variance and parameter variance propagated jointly into the reported uncertainty interval1714. Verra's VM0042 v3.0 consultation document and the GHG Protocol Land Sector and Removals Standard (effective 1 January 2027) both require empirical, pool-specific data for any reported soil-carbon removal, ending the “model-only” loophole that several first-generation programmes used1318. CarbonPlan's independent buyer-guide review concluded in 2023 that no commercially available protocol simultaneously met adequate rigor, durability, additionality, and depth requirements; the 2026 standards cycle is, in part, an explicit response19.

This matters for the signal-versus-noise framing. A model alone has a deterministic prediction with a structurally optimistic uncertainty bound. A measurement campaign alone has the noise floor described above. A measurement-calibrated model can compress the uncertainty interval below either component, but only if the measurements are dense enough to constrain the parameters that matter and the model's structural error is included in the propagated variance. “Model-light, measurement-light” hybrids are not credible; “measurement-as-model-calibration” is the operational standard the 2026 regimes are converging on.

Honest reporting templates

The discipline that makes a signal-versus-noise framing operational is in the reporting. Four numbers should travel together with any SOC change claim, and a 2026-era verifier will ask for them.

1. The point estimate of the change, with units and depth (e.g., “0–30 cm SOC change of +1.4 Mg C per hectare over 5 years, ESM-corrected”).
2. The 95 percent confidence interval computed from the design's actual variance, not from a generic literature CV.
3. The MDD the design can resolve at 80 percent power, computed before the campaign, and the result expressed as “claim within / outside detection limit”.
4. The propagated uncertainty taxonomy following IPCC AFOLU Volume 4 conventions20: activity-data variance, reference-stock variance, and stock-change or model-error variance, additively combined.

A claim of +1.4 Mg C per hectare with a CI of −0.2 to +3.0 and an MDD of 1.6 is honest, useful, and unverifiable; it goes in the report as such and the project plans for the next remeasure. A claim of +1.4 Mg C per hectare with a CI of +0.6 to +2.2 and an MDD of 0.8 is honest, useful, and defensible; it goes in the inventory.

What follows for project design

The signal-versus-noise framing simplifies several procurement decisions that have, over the past five years, repeatedly been litigated as if they were a matter of opinion.

The unit of inference is the project, not the field. Per-field detection of slope-style practices is statistically unreliable at affordable density, while project-mean detection across 20 to 60 paired fields is operationally tractable. Design accordingly, and route the buyer's claim to the level the data can defend.

The honest detection window is the management horizon, not the corporate roadmap. For cover crops, no-till, and most agroforestry types, that is 8 to 12 years from baseline. Three-year reporting cycles are still useful, as long as they report against the MDD trajectory rather than against a presumed signal.

The deep profile is no longer optional. Subsoil gains in perennial systems older than ~18 years often exceed topsoil gains and are invisible to a 0–30 cm design5721. ESM correction is similarly non-negotiable under tilled-versus-no-till comparisons67.

The protocol the buyer is audited against sets the floor. VM0042 v2.2/v3.0, the EU CRCF (Regulation 2024/3012), the GHG Protocol Land Sector and Removals Standard, and SBTi FLAG converge on stricter empirical-data requirements through 2026–202713182223. Designing to the most permissive standard the team is familiar with, when the buyer will be audited against the strictest, is a slow-motion failure mode.

Key takeaways

• Soil organic carbon change detection is governed by an inequality between a small management signal (typically 0.1 to 0.5 Mg C per hectare per year for slope-style practices) and a noise floor with three components (spatial, temporal, analytical), of which spatial dominates in unpaired designs.
• A defensible claim travels with three numbers: the point estimate, the 95 percent confidence interval, and the minimum detectable difference the design could resolve at 80 percent power.
• Paired baseline-resample designs are the operational structure for credible change detection; per-field claims at affordable density are not statistically supportable for slope-style practices.
• The honest detection window for cover crops, no-till, and most agroforestry types is 8 to 12 years, not the 3 to 5 years built into most corporate carbon roadmaps.
• Stratification, paired designs, and within-stratum compositing compound; the 15-per-stratum verification floor caps the gain on small management units.
• Depth integration to 0–60 cm and ESM correction are no longer optional under the 2026 verification regimes; designs that skip them produce credible-looking numbers that do not survive a Tier-3 audit.