ramet5: reference areas for nature finance and carbon baselines

The list of characteristics includes things like slope, distance to a road, forest type, rainfall; reasonable things to care about. But two things go wrong. Areas that look similar on four variables can be quite different in ways nobody listed. And the person choosing the variables has an interest in the answer. Even with the best intentions, two analysts will choose different variables and get quite different answers, and neither can prove there was no unconscious bias.

Pixel matching pairs isolated points across an enormous region and calls the result a comparison. It ignores what everyone knows about landscapes: neighbouring land behaves alike, and isolated pixels far apart do not.

The consequence of ignoring geography shows up in published work, often with misleading results. In one well-known critical assessment, the comparison areas carrying the most statistical weight for a Colombian project were 900km away in a different ecoregion. Nothing in the method flagged it, because distance and ecoregion were not among the variables being matched on. Our response to that assessment, with its map, is on the avoided deforestation page.

1. The satellite record becomes a fingerprint

A foundation model is a very large machine-learning model, trained on an enormous quantity of satellite imagery. It learns, without being told what to look for, how to summarise a patch of land as a list of numbers. These lists are called embeddings, because so much information about the patch is embedded in them.

The useful property is that patches which look alike to the model end up with similar numbers. Not similar in one dimension that a person chose, but across everything the imagery contains: forest structure, terrain, water, the marks of roads and clearings and cultivation, how all of it changes through the year.

We describe an embedding as a fingerprint of that patch of land. While not a precise metaphor, we use it to distinguish how this is different from a hand-picked list of variables; it does not depend on anyone guessing correctly in advance about what matters. Crucially, it also takes away the possibility of choosing the covariates for the matching, and with it the wiggle room to steer the outcome of the analysis.

We work on the A5 geo grid, an equal-area grid built from pentagonal cells. Equal area matters because credits are calculated per hectare, and the commonly used hexagonal grids change area with latitude. The A5 grid is open and we publish our tools for working with it.

The embeddings we currently use come from two published foundation models, AlphaEarth Foundations and Embedded Seamless Data (ESD).

2. The project area sets the specification

This is the step that most distinguishes what we do. We do not start from a list of what a good comparison looks like in general. We start from the project. The project area’s own fingerprint defines what a reference area has to match, so the specification is derived from the similarity of the landscape itself, not from a methodology document or from an analyst’s judgement.

If a project area contains three kinds of terrain, its reference areas have to contain them too, in something like the same proportions. The match is made on a single year’s fingerprint, so that the reference areas are the most similar kinds of land. What happened in the years before the project, in the project area and in each reference area, comes in at step 5, where those histories build the weights, which ensures the histories of disturbance are included.

We use the embeddings to constrain the donor pool, which is how we select the comparison areas. The embeddings mean we match on all kinds of things without choosing what to match on: a reference area is selected because its whole satellite record resembles the project’s, not because it scores alike on a handful of variables somebody picked in advance. That selection step is the part our patent application covers.

3. The engine searches the landscape

With the specification set, ramet5 searches the surrounding landscape for the areas whose fingerprints come closest to it, and selects them automatically. A match is made on a whole year of imagery, not on a single image. The embeddings summarise the year, in the same way that many biomass composites do, so a place that looks right in one season and wrong in another does not come out as a close match.

The result is reproducible. Run it again on the same project with the same inputs and you get the same reference areas, which is not true of a process involving human selection. It is also fast: the work takes days, not weeks, and a jurisdiction that has never been analysed before takes no longer.

4. The project and its references define the training region

The project area and its selected reference areas, taken together, span a region. That region is what we use to train the model that estimates carbon stocking.

A global biomass model is fitted to everywhere at once, so it fits no specific place well. A model trained on the specific region a project sits in, and on the areas it is being compared against, is fitted to the conditions the comparison actually depends on.

That model is trained on canopy structure measured by NASA’s GEDI lidar, validated against other extensive lidar datasets, and checked against two independent biomass products, CTrees and ESA CCI Biomass, over the project and its reference areas.

5. Causal inference does the rest

Once the donor pool exists, the matching is done, and that is when we switch to much more traditional statistics: what is used in the econometric literature, particularly synthetic controls and difference in differences. These are models used in economics to understand the effect of a policy decision or a change in a financial market, where you cannot run a typical experiment. You cannot invest in two regions, protect one and leave the other, in order to measure the difference. The controls are not perfect, but by taking a combination of several you can build a good synthetic control to compare the project against after the intervention.

You do not need a one-to-one perfect match. What you need is enough similar areas that you can weight them: the years before the project build the weights, so that the synthetic control follows the project’s own trajectory, and a donor that does not help that match gets zero weight. Synthetic controls are statistically very sound, but they need good donors. It is not down to the causal inference method, it is down to the donor selection, which is what the steps above are for.

The final product is an estimate of what would have happened in the project area, with uncertainty attached, not a single number. It is recalculated at every monitoring event, from the areas matched at the start of the project.