VM0047 data provider: reference areas and baselines for tree planting projects

So the whole claim depends on knowing what the land would have done without the project.

VM0047 sets the baseline by comparison, not by projection. Instead of assuming what the land would have done, the methodology looks at comparable land that did not get a project, and uses what happened there. That is what makes the baseline dynamic: it moves with observed conditions through the life of the project.

A projection cannot be checked. A comparison can, because the places are real and anyone can look at them. The methodology documents are published on Verra’s VM0047 page.

Under the Paris Agreement’s Article 6.4 mechanism the equivalent planting methodologies are still being written; we follow them on our A/R PACM methodologies page.

VM0047 has two ways of quantifying removals. The area-based approach measures change across the whole project area from satellite data and field plots, and sets its baseline from matched control plots. The census-based approach counts the planted trees themselves and does not use control plots; it suits dispersed planting such as agroforestry, shelterbelts and urban trees. A project may use one, the other or a combination, depending on which applicability conditions it meets. This page is about the area-based approach, the one that needs reference areas.

It works in three steps. First, at validation, the project is matched to a set of comparable areas, known in the methodology as control plots. Second, those areas are watched for as long as the project runs. Third, at each verification the benchmark is recalculated from what has grown in them since the last check.

The project is credited only for the difference. If the comparable land regrew a certain amount on its own, the project gets no credit for matching it, and credit only for what it achieved beyond it. This is the core of the design, and it is why the methodology needs comparison areas at all. Land that is left alone does not stay bare. Some of it regrows without anyone planting anything, and a project that took credit for that regrowth would be selling something it did not cause.

A useful consequence is that a good year for the whole region is not automatically a good year for the project’s credit count. If the rains were kind and everything grew, the comparison areas grew too, the benchmark rises with them, and the project is measured against the better conditions everyone had. A baseline projected forward from history would have missed that entirely.

One part of the arrangement does not move. The benchmark is recalculated, but the comparison areas it is calculated from are not reopened. They are chosen once and kept.

The stocking index is built from the satellite record, not from a survey, which is what makes it possible to calculate the same measure consistently across a project area and across every area it is being compared to.

Under VM0047 it does two jobs. They fail in different ways.

It decides what is eligible. Land that already carried substantial tree cover before the project is not land where a project can claim to have added trees. The stocking index is how that line gets drawn, which means it determines which parts of a site can be included at all.

It is the measuring stick. The same index is calculated for the comparison areas, and VM0047’s prescribed matching procedure selects the closest controls on it. The benchmark is then recomputed from stocking index data in those areas at each verification.

Because it does both jobs, an error in the stocking index does not stay in one place. It moves the eligibility line and it moves the bar the project is measured against, and it does so for every issuance the project ever makes.

There is a limit. Matching on the index alone compares places on how much is growing, not on what kind of place they are. As we have written elsewhere, two areas can look identical on a stocking index while sitting in completely different landscapes: a back garden surrounded by houses can match a forest patch on greenness alone. That is why the pool of candidate areas the matching runs within matters as much as the matching itself.

Our stocking index measures canopy structure, with NASA’s GEDI lidar as the reference: canopy height and plant area index, a measure of canopy density, used together. It is built from two published satellite embedding products, Embedded Seamless Data (Chen et al., 2026; Landsat and MODIS, consistent from 2000) and AlphaEarth Foundations (Brown et al., 2025; from 2017), with a model trained for the project’s region against GEDI canopy measurements. It is produced for every year of the baseline and monitoring period, with no gap years.

Every value carries a prediction interval, checked against GEDI data held back from training; on held-out data, 95% intervals cover about 95% of observations. Accuracy is reported against GEDI held out geographically, so the check area is never seen in training: for canopy height, about 72% of the variation explained and a typical error of about 2.5 m. We supply the index only where we can show it tracks biomass in the project’s forest type, cross-checked against independent biomass products and local field plots where they exist.

The control plots, their weights and their coordinates are set once and held for the duration of the crediting period, which for most ARR projects under the VCS Standard runs 20 to 40 years.

Verra chose that deliberately, and it was right to. A baseline that let a developer quietly upgrade its comparators whenever the numbers looked unflattering would not be a baseline at all. But it does mean there is no second chance. Everything the project is credited on for the rest of its life is measured against a comparison chosen before anyone knew how it would turn out, and the first anyone usually hears about a poor choice is when an auditor or a buyer asks a question that cannot be answered.

The pool of candidate control land starts as VM0047’s own eligibility rules, set out in the methodology’s Table A1: same jurisdiction and ecoregion, matched policy and tenure, within 100 km, with the project, protected areas and registered carbon projects removed. Within that pool, we narrow the candidates to land that resembles the project, using the fingerprint of the landscape, a step the methodology permits under Appendix 1, Section A1.4 Step 2.1. It only ever narrows the pool; it never admits land the rules exclude, and it is reported as an explicit, auditable step. The second extension we flag is the matching variable itself: a stocking index trained for the project’s region, in place of a coarse global product.

VM0047 asks for uncertainty to be quantified. We treat that as the starting point, not an extra, because a baseline is an estimate of something that did not happen, and the way to report an estimate is with the range around it.

A planting project can measure what grows to the millimetre, but not how much would have grown back on its own. The uncertainty in an ARR baseline does not come from one place. It comes from the matching, from the inputs the stocking index is built on, and from the selection of the comparison areas themselves. We report a range around every estimate that covers all three, not a single figure that hides which part of the work is least certain.

Reporting a range is not the same as rating a project. We publish the estimate and the bounds around it and leave the judgement to buyers, auditors and standards bodies. More on why we report a range.

For the stocking index, the range around each value covers the model’s error combined with the error of the GEDI lidar it is checked against, so it is conservative for the true state of the canopy. It is validated to its stated coverage on data held back from training. Field plots, where they exist, calibrate the index to ground-measured biomass and tighten that range.

What a developer gets is a baseline they did not choose and can therefore defend: the comparison areas selected by a process that runs the same way every time, the maps behind them, and a record an auditor can follow. Because the selection is automated, the work arrives in days, not the weeks of analyst time it takes to assemble by hand.

Our product for VM0047 is called ramet47. It runs on ramet5, the matching engine underneath everything we build.

Reference area selection. Comparable areas chosen from the satellite record, not from a shortlist of variables, with the selection derived from what the project area contains. Reproducible, so running it again gives the same answer.

Stocking index mapping. Historical stocking across the project and its reference areas, from a model trained on that region, not a global product.

The dynamic baseline and performance benchmark. Updated through the crediting period from the reference areas, as the methodology requires.

Uncertainty. Reported as a range around every estimate, covering the matching, the inputs and the selection. VM0047 asks for uncertainty to be quantified.

Audit materials. A pack written for someone examining it critically, not a summary. If our outputs generate findings during validation or verification, we answer them in writing until they are closed.

We price per project. Reference areas are always included, however many the analysis needs, because the reference areas are the product. Larger projects cost more than smaller ones.

There is no setup fee, and we quote the current year’s price without asking anyone to commit to a multi-year arrangement.

Not every run is for an audit. If you are still choosing a provider, checking reference areas selected in house, deciding whether a site can go ahead, or putting a project in front of funders, we run the same analysis at a lower price than an audit year. The answer is the same; what you are not paying for is the audit materials and the work at each verification. Ask us for the preview price.

Knowing early whether a site can go ahead is cheaper than finding out at validation.