Soil Carbon Isn’t Too Messy to Measure

(Yale School of the Environment) Re-measuring the same spots a year later cuts through the natural variation that has made soil carbon so hard to verify, researchers at the Yale School of the Environment found. -- Dig into one Hudson Valley farm field, and you might pull up soil that’s 5% carbon. A few steps away, it’s closer to 3%.

That patchiness is a major reason many scientists have doubted that anyone can directly track whether farming practices add carbon to the ground. It matters because governments and companies are increasingly paying farmers to build soil carbon, in the hope of removing carbon dioxide from the atmosphere. If the changes from a farming practice are small compared with that spot-to-spot variation, which scientists call “noise,” it’s hard to know whether anyone is getting what they paid for.

As a result the field has largely turned to computer models. Those models may well be right. The problem is that, without field measurements to test them against, nobody can say.

“Right now there is a reliance on model-based predictions of what different practices are expected to achieve,” said Alexander Polussa, lead author of a new study in Geoderma and a research project Manager at the Yale Applied Science Synthesis Program (YASSP). “Part of the reason is an assumption that directly measuring these effects across working landscapes is prohibitively difficult.”

If you wanted to know whether a diet works, you’d treat each person as the unit of the experiment, since the diet is applied to people, not to individual body parts. You wouldn’t compare the weights of 80 different people. Some are tall, and some are short—and that variation would swamp any change from the diet. You’d weigh each person twice and look at how each one changed.

Polussa and colleagues applied the same logic to farming. A farmer’s management applies to a whole field, so the study treated each field, not each soil sample, as the unit being tested. “For quantifying a management intervention, we need to use the unit at which the intervention is applied,” Polussa said. That differs from the classic approach of sampling across a region to take inventory. The team selected 80 fields on working farms, 40 managed regeneratively and 40 conventionally, matched by soil type and starting carbon. In the winters of 2023–24 and 2024–25, they collected 1,592 samples from 796 precisely marked locations, comparing each field with itself a year later. That step alone removed about two-thirds of the noise.

Returning to the same spots within each field helped further. Carbon levels varied from place to place, but consistently. Spots high in carbon one year tended to be high the next. By sampling the same marked locations both years, the team could cancel out those built-in differences. A field’s average reflected how much its carbon changed, not where the cores happened to land. Averaged across many fields, that reveals the typical effect of a practice. Overall, changes in carbon varied far less from field to field than carbon levels themselves, making real change roughly six times easier to detect.

“The question was whether that natural variability would drown out any signal, and our results suggest the concern is overstated,” Polussa said.

Regenerative fields, which relied on practices such as manure, compost, and perennial grasses, gained about 0.04 percentage points more soil carbon than conventional fields over the year. The statistical range around that estimate includes zero, and one year of data can’t show a lasting trend.

But that’s not the focus of the research, Polussa emphasized. The goal was to test whether the approach works under real-world conditions, not to deliver a verdict on any one practice.

The study also showed that the fields you sample shape the answer. When the team looked only at sandier fields, the estimated benefit nearly vanished; in less sandy fields, it grew. Those subgroup estimates were too imprecise to draw firm conclusions, but the lesson was clear: no amount of sampling can recover from a poorly chosen set of comparison fields. Adding more fields also helped more than adding more samples within each field, though more fields means recruiting more farmers and driving more miles.

For carbon registries and project developers, it could change how they measure and re-measure the same points. These choices are increasingly related to regulation. They’re being considered in the frameworks that will govern carbon accounting for aviation fuels, including CORSIA, the international program setting those rules, where the Yale team is serving as technical advisors to the working group. The momentum there is toward empirical measurement, which puts these sampling decisions at the center of whether a credit holds up.

“Ideally, programs paying for these practices would build in measurement from the start by sampling fields before they switch, and sample comparable fields that don’t switch, then keep measuring both over time,” Polussa said. “That’s also how we build justified confidence in the models we’ll still need to work at scale.”

Models still have a role.

The authors aren’t arguing against models, Polussa said, and not every project needs an intensive sampling campaign. The study has limits. It covered one year, one region, and the top six inches of soil, and it measured carbon concentration rather than total carbon stored. But direct measurement should become a routine part of the evidence behind agricultural and climate policy. He pointed to public health as a comparison: “We wouldn’t want to make major public health decisions relying only on modeled expectations without making substantial efforts to validate those predictions with empirical evidence.”