Statistical Sampling for Carbon Credits: A Project Developer's Guide

There is a number in every carbon project that decides how many credits it can sell, and it is not the biomass estimate. It is the width of the interval around that estimate.

Statistical sampling is how you control that width. It means measuring a designed subset of the project area and scaling the result with quantified uncertainty and standards require it because a designed sample has properties an attempted census does not: a calculable confidence interval, a defensible sample size, and a stated relationship between what was measured and what is claimed.

Carbon stock varies sharply over short distances. Species composition, stand age, soil depth, hydrology and disturbance history all drive biomass, and they change at scales of tens of metres.

That variance is the thing a sampling design exists to handle. Measuring more trees without addressing variance structure does not improve precision proportionally, it just costs more. Measuring fewer trees in a design that reflects the variance structure can produce a tighter estimate for a fraction of the effort.

The output that matters to a verifier is not the mean. It is the interval around the mean, and whether you can justify it.

Stratification divides a heterogeneous area into strata that are internally more uniform. Sampling then happens within each stratum, and results are combined with weights proportional to stratum area.

The leverage comes from arithmetic. Total variance decomposes into within-stratum and between-stratum components. Good strata push variance into the between component, where it is handled by the weighting rather than by the sample size. Within-stratum variance is what drives how many plots you need.

Strata that work in nature-based projects are usually built from:

Ecosystem type — mangrove, peat swamp, lowland dipterocarp, dryland, agroforestry. These differ in carbon density by multiples, not percentages.

Stand age or successional stage — biomass accumulation is non-linear, so mixing ages inflates variance badly.

Soil and hydrology — especially where soil carbon is a counted pool, and decisively so on peat.

Management history — logged, unlogged, restored, actively managed.

Strata that do not work are ones drawn for administrative convenience: ownership blocks, compartment boundaries, or anything that does not correlate with biomass.

Sample size follows from variance, target precision and the population's structure — not from a rule of thumb.

The practical sequence is: run a pilot or use prior inventory data to estimate within-stratum variance; set a target precision, usually expressed as a relative half-width of the confidence interval at a stated confidence level; then compute the plots needed per stratum, allocating more to strata that are larger or more variable.

Two errors recur. The first is picking a plot count first and deriving precision afterwards, which inverts the logic and usually under-samples the variable strata. The second is allocating plots proportional to area alone, ignoring variance — which over-samples uniform strata and under-samples the messy ones that actually drive the interval.

Every conversion step adds uncertainty, and it compounds: measurement error on diameter and height, allometric model error converting dimensions to biomass, wood density variation, expansion factors for unmeasured pools, and the sampling error itself.

Standards require that uncertainty be quantified and reported, and they apply conservative deductions scaled to it. That is the mechanism that connects sampling quality directly to revenue: a wider interval means a larger deduction means fewer issuable tonnes.

Projects that treat uncertainty as a reporting formality discover this at verification. Projects that treat it as a design parameter optimise against it from the start — which usually means better stratification rather than more plots.

Crediting runs over years. Re-measuring the same permanent plots rather than drawing fresh ones each cycle removes a large source of between-period variance, because change is observed directly rather than inferred from two independent estimates.

This matters more than it sounds. The uncertainty on a difference between two independent samples is larger than the uncertainty on a measured change in the same plots. For projects crediting incremental growth, that difference can dominate the result.

Remote sensing extends a ground sample across the full area, but the relationship between sensor data and carbon must be established by regression against those ground plots — and is only valid within the conditions it was calibrated on.

The plots are therefore doing two jobs: producing the estimate, and calibrating the extrapolation. A design that serves only one of those jobs will disappoint at verification.