Geospatial Analysis of Carbon Offset Projects: EldHollow's Approach and a Broader Scientific Outlook

The world has committed to protecting and restoring nature at an unprecedented scale. Whether that commitment delivers what it promises comes down project execution on the ground. Local socioeconomics and forest ecology intertwine to create complex challenges for projects to overcome during implementation, and at the end of the day, projects boil all of these complexities down to one single unit: the carbon credit. So the question becomes: can we actually measure what is happening to a forest, accurately and honestly, and everywhere at once?

For a long time, the honest answer has been no. Historically, many forest carbon projects overstated their impact. Usually it was because the baseline was too generous, or because the measurements underneath were flawed. For anyone with a stake in nature markets, that uncertainty is one of the core risks.

Here is how it fits together, and where it goes next.

For decades, mapping forests from space meant assembling images from many satellites by hand, correcting each sensor's quirks, patching over clouds, and stitching the years together so they lined up. It worked, but it was slow, brittle, and it had to be rebuilt for every new region. EldHollow's initial geospatial pipeline was built on this type of structure, and while it was functional and produced reasonable model results it often required close scrutiny and constant refinements.

Then Google DeepMind developed a single, unified view of the planet called AlphaEarth. For every ten-meter patch of ground on Earth, each year beginning in 2017, it takes the plethora of raw satellite data (optical, radar, spaceborne LiDAR, climate, elevation) and distills it into a compact 64-number matrix for each pixel. The messy work of reconciling sensors, filling gaps, and keeping things consistent from year to year has already been done once, for the whole planet. Google has committed to updating its AlphaEarth dataset every year for the time being, which will allow for continuous observations over time using a consistent data source.

That is what makes scalability possible. Because AlphaEarth's data is formed from the same foundation everywhere, a model we train over the Amazon runs, unchanged, over temperate rainforest or the boreal north. Nothing has to be re-engineered for each new geography, which is exactly what measuring nature at a global scale requires. Now, the AlphaEarth data is not a perfect product. It still has its flaws and results in models that are only slightly better than most custom analysis pipelines. The advantage is that most of the data processing is done for you, and the data is available worldwide.


Anyone can build a model that looks good on a map, and many organizations will market faulty geospatial products as being more robust than they really are. The harder question, and the one that matters, is whether the product is honestly accurate. EldHollow has devoted considerable time and resources into calibration, validation, post-processing, and estimating uncertainty to ensure our models are reliable and taken in their proper context.

We validate the way the science demands. It would be easy to test a model on data right next to where it learned, but nearby places look alike, so that flatters any model. Instead we hold out whole regions and see how the model does on forests it has never seen. When a model forecasts forward in time, we judge it the way you would judge a weather forecaster, on the years it never got to study. And we calibrate against reality, correcting the numbers so they mean what they say. Left uncorrected, models like ours tend to underestimate the tallest, most carbon-dense canopy, so we correct for exactly that.

Still, no model is perfect, and we are clear about our models' limitations. Every estimate comes with its uncertainty, and every model ships with its assumptions and its limits made as clearly as possible to the end-user. We would rather show an honest assessment of model accuracy than make unfounded claims about the model's quality. For example, EldHollow's models suffer from saturation like most other satellite-derived forest structure models, though the intensity of saturation varies from region to region. This is more an artifact of data quality than anything else and will continue to plague the remote sensing field until better data sources become available. 

Deriving a baseline

A credit does not reward a forest for standing. It rewards a forest that would otherwise have been harvested or would not have grown in the first place. Working out that alternate history, the baseline, is the most scrutinized step in all of carbon accounting, and it is where many projects have failed in the past.

We build a baseline to compare against the project's directly. First, a machine-learning model learns what actually drives deforestation or forest growth in a given landscape: the terrain, the roads and settlements, how close a place is to access, the condition of the forest itself. Then we search the surrounding area for places that closely match the project on those drivers, using the same matching techniques that economists and medical researchers use to build fair comparisons. We follow those matching areas over time as a stand-in for what the project's land would probably have experienced, on average.


Carbon only counts if it lasts, so the pipeline does not stop at a single snapshot. Every year it looks for change in both directions. It catches loss, whether that is fire, outright clearing, or the slow thinning of degradation. It also catches growth, as forests put on height and cover through recovery or active restoration.


Everything mentioned so far is operational in EldHollow's geospatial analysis pipeline today, but several new analyses are currently in development. The most interesting thing about measuring the living structure of forests everywhere, every year, is what it makes possible to observe on the landscape.

Start with forest connectivity. EldHollow's stocking index is, in effect, a map of habitat quality across a whole landscape. Read it the way ecologists do, with dense forest easy for wildlife to move through and cleared land acting as a barrier, and you can trace the corridors that hold a landscape together, measure where it is breaking apart, and show whether a project is reconnecting nature or leaving isolated islands of it. As markets for biodiversity and nature-positive outcomes take shape, that kind of evidence, drawn from data we already process, becomes valuable in its own right.

A few other directions follow from the same foundation:

- Catching degradation, not just clearance. A yearly record of forest structure can pick up the quiet losses, like selective logging and thinning, that simple "forest or not" maps miss entirely. Those quiet losses add up to a large, under-counted share of emissions.
- Proving co-benefits. Structure, connectivity, and disturbance history combine into real evidence of habitat and community value, which is the substance behind nature-positive claims.
- Beyond forests. The approach is not limited to forests. The same foundation extends naturally to mangroves and other coastal "blue carbon" systems, to grasslands, and to wetlands.

Why it matters

A market that works for nature needs measurements the everyone can trust. These include the carbon that is there, the carbon that would otherwise be lost, and the durability risks that could undo a project's hard work. We think this is, at its core, a scientific problem, and that solving it with rigor, honesty, and global reach is the right thing to do. It is a start, but there is so much more work to be done.

Further reading

A few places to go deeper on the science behind this work:

McRae, B. H., et al. (2008). Using circuit theory to model connectivity in ecology, evolution, and conservation. Ecology. The connectivity modelling our stocking index can feed. https://doi.org/10.1890/07-1861.1