What a soil core knows

A soil-carbon project pays to sample a field, then makes a claim about all of it. Soil sampling runs a quarter to two fifths of a carbon project's budget, so the sampling design is the single largest cost decision in the work. Here is what that money buys, and the one number that says what it cannot buy at any density.

The sampling design

Field 100 ha

Cores per hectare 20

Sampling depth 30 cm

Cost per core $12

Autocorrelation range · as published

Nugget / sill · as published

Both bottom rows are set from values reported in the soil-carbon literature rather than invented. Published ranges for soil organic carbon span roughly 234 m to 2,102 m depending on site, and nugget-to-sill ratios near 0.20 and 0.31 are both described as strong spatial dependence.

What the money bought

2,000cores taken

$24,000at the cost per core above

0.63 ppmof the soil volume physically sampled — parts per million

22 mspacing between cores on a regular grid

Two questions, one sampling design

The parts-per-million figure is arresting and, on its own, unfair. Soil carbon is spatially autocorrelated, so a core speaks for far more ground than it occupies, and the field mean comes out well. What does not come out is everything varying below the spacing you chose, and geostatistics has a name and a number for that.

The field mean covered

Variation below the spacing 20% of the total variance sits in the nugget — variance at separations shorter than any pair of cores. It does not fall when you add cores at the same spacing. It is the unknown fraction, already measured, already published, and treated as noise.

How much rests on the range

One core speaks for a circle of the autocorrelation range. Published ranges for the same quantity differ by nearly an order of magnitude between sites, and this review did not locate a requirement that a project measure the range for the field it is sampling.

Range assumed	One core speaks for	Cores needed for areal cover	vs. the design above

What this is. A first-order model of a sampling design, not a carbon protocol. Sampled volume is exact geometry. Areal coverage treats a core as representative over a circle of the autocorrelation range, which is the standard geostatistical reading of a variogram range and is generous to the design. The nugget is reported as published, not derived. What it is for. A carbon credit is a claim about a volume. This shows how much of that volume the claim rests on, and names the part of it no sampling density at this spacing can reach. Companion to the grain bin instrument, which asks the same question where the physics is thermal rather than statistical. Part of the Glass Lab × Charlot Lab track on Physical AI and integrated sensors for agriculture surveillance.