Liana cutting accelerates the recovery of logged tropical forest three times faster than tree planting, at one tenth of the cost. That is the headline of our new paper in Current Biology, out of the Sabah Biodiversity Experiment in Borneo. The result matters not only for what it shows but for the experiment behind it: the Sabah Biodiversity Experiment is one of the few places where forest restoration interventions have been tested against proper controls. A new Verra methodology, now in public comment, is being built on exactly this kind of evidence.

VM0047 made a brave call. It locks the developer's matched reference areas at validation and keeps them fixed for the project's full crediting period. For most ARR projects under the VCS Standard, that crediting period runs for 20 to 40 years. Long-term commitments are good for forests, but they mean you have to make the right choice up front.

The choice was deliberate. A baseline that lets developers quietly upgrade their comparators when the numbers start looking too good is not a baseline at all. By fixing the matched controls at validation, VM0047 closes that loophole. It is one of the more credibility-protecting moves in modern carbon methodology design.

But that brave methodological call makes the developer's matching a defining moment. One choice at validation, which is locked for decades.

Why your VM0047 control plots define your project for the full crediting period

Reference areas drive the performance benchmark for every issuance over a project's life. Each verification computes the benchmark on the same locked controls, using whatever new stocking-index data has accrued. The benchmark updates over time. The matched controls behind it do not.

VM0047 makes this explicit. Once the match passes Verra's quality test, "the final selection of control plots and their respective weights are then fixed... for the duration of the crediting period." The same control plots, with the same weights and coordinates, will define your project's benchmark for twenty, thirty, sometimes more than forty years.

This is not one of many small calls in project development. It is the single largest methodological choice a developer makes.

What makes a VM0047 baseline defensible?

Reference area selection is also the part of the methodology most exposed to challenge. Even when the developer does the work with care and integrity, with expert analysts and the best available data, the selection is open to questions and attack. Which covariates were used? What distance threshold was applied? Which polygons were excluded? Why one set of plots and not another?

This is not a hypothetical concern. Across the broader carbon credit literature, baseline construction is consistently identified as the lever where overstatement enters the system. Recent work in Nature Communications attributes systematic over-crediting in forest carbon programmes to selection-bias mechanisms in baseline determination (Swinfield et al. 2026). Reference area selection is the step under scrutiny in most of these accounts.

A global evaluation of voluntary REDD+ projects published in Conservation Biology (Guizar-Coutiño et al. 2022) found that a more defensible matching procedure can both confirm real reductions and substantially adjust their estimated scale. The methodology behind the match matters.

Buyers know this. A project that can show its reference areas were chosen by a third-party process, with documented covariates and reproducible outputs, has a stronger position with both buyers and ratings than one that picked its own.

The challenge for the developer, of course, is that even doing the selection well is not the same as being seen to have done it well. Conflict-of-interest logic is hard to refute when the developer holds the pen.

What does independent VM0047 reference area selection look like?

Independence alone is not enough. A consultant or service provider can be accused of the same kinds of bias as a developer the moment they are paid to produce a result. What matters is whether the selection is reproducible, auditable, and removes the role of judgement as much as possible.

There are two key elements to selecting reference areas: finding matching areas, and the criteria for how they match. VM0047's prescribed procedure for matching is based on the closest controls on the stocking index, which are then locked in and used for the benchmark. But how the candidate pool is built is even more important. Two areas can look identical on the stocking index while sitting in completely different landscapes. A back garden surrounded by houses can match a forest patch on greenness alone. Without that context, the match returns plausible but wrong comparators. This is the pixel-level problem we discussed at ML4EO 2026: per-pixel accuracy means little if it produces matches that do not make landscape sense.

Independent VM0047 reference area selection, done well, means a third party identifies the candidate areas before the prescribed matching runs. Best practice is to use a process that another analyst running the same procedure would replicate exactly. The developer does not choose the inputs. The methodology runs as written within the regions returned.

The selection should be reproducible: same inputs, same outputs, regardless of who runs the procedure. It should be auditable, with every covariate, threshold, and exclusion documented and inspectable. It should be vendor-agnostic, sitting upstream of any biomass or stocking-index provider so the developer is not locked into a downstream stack. And it should be compliant with the methodology: VM0047's prescribed matching procedure runs as written within the candidate areas the search returns.

This is the service belian provides. We sit upfront in the developer's pipeline and run VM0047's prescribed matching as written within the regions we identify. We do not modify or replace any of Verra's prescribed procedures. We add an upstream step the methodology leaves to the developer. We make sure the procedures are run in a way that is reproducible and auditable, and free from any conflict of interest.

That distinction matters. The developer ends up with reference areas they can defend to a VVB, to a buyer, and to a ratings agency, because they did not pick them.

Faster baselines, faster credits

Reference area selection is also one of the most time-intensive parts of VM0047 baseline preparation. In-house, building the donor pool, running the matching procedure, and validating the quality thresholds can stretch across weeks or months of analyst time per project.

Independent automated selection compresses that timeline. The output arrives in days, not weeks. For developers, that means shorter project preparation, smaller upfront cost, and an earlier first issuance. For programmes operating across many sites, it shifts a one-off bottleneck into something predictable and budgetable. ARR economics are notoriously thin. Anything that compresses the path to first issuance moves projects that would otherwise be marginal into financial viability.

Faster, defensible baselines also help on the validation side. The cleaner and more auditable the selection record, the less time the validating body spends on questions during validation review. The benefits compound.

VM0047 is the right methodology. Make sure it runs on the right ground.

VM0047's lock-in is one of its strengths. It is also the reason the developer's first decision is heavier than any decision they will make later. Choose well, and the methodology rewards the project for the full crediting period. Choose poorly, and the project lives with that decision for the same length of time, and may lose credibility.

If you are working on VM0047 projects and want to stay up to date with developments in independent reference area selection, sign up below.

Last year we supported ML4EO, the Machine Learning for Earth Observation conference in Exeter. Many different researchers and practitioners showing what's now possible with satellite data, foundation models, and geospatial AI across conservation monitoring, carbon markets, and environmental integrity. We're sponsoring again this year because we think this community matters.

What machine learning for earth observation gets wrong

The advances in geospatial AI are exciting, and this is why we are supporting the conference. Foundation models trained on petabytes of satellite imagery. Embeddings that capture landscape similarity without hand-crafted features. New ways to monitor forests, map habitats, and track land use change at global scale. It's what we work on every day.

But one of the talks that stuck with us most from ML4EO 2025 wasn't about a new model or a new dataset. It was about caution. Karen Anderson and Bri Pickstone from Exeter presented "Why ML != EO: From Black-Box to the Accuracy Paradox", arguing that the remote sensing community is adopting machine learning faster than it's thinking about what these tools actually represent. We agree.

Remote sensing scientists work with composites every day. Mosaics stitched from images captured at different times, from different viewpoints. Few people stop to question them. In an age of machine learning for earth observation, this is getting more pronounced. Models built on composites that have not been questioned. Pixels treated as more accurate than ground truth when the same tree can look completely different one day apart because of sun angle. Pixels are useful abstractions, but they are abstractions.

Why this matters for carbon and biodiversity conservation

This is where it connects to what we do. If you can't be certain what a pixel represents, what does that mean for carbon credit baselines? For conservation monitoring? For any claim that says "this forest is intact" or "deforestation happened here"?

Deforestation doesn't occur on a pixel. It occurs across landscapes, at different scales, driven by processes that don't respect grid boundaries. The question isn't "what class is this pixel?" It's "what would have happened here without intervention?" That's a fundamentally different question, and it needs different tools. At belian.earth we work with embeddings rather than pixel classifications. We compare trajectories across larger scales rather than labelling individual pixels. We do this not because it's fashionable, but because the pixel-level assumptions don't hold up when the stakes are real.

Consider the accuracy paradox. A land cover map with 92% overall accuracy sounds good. But at ML4EO, Bri showed that heathland was only detected 34.8% of the time, misclassified as grassland. High overall accuracy hiding a catastrophic failure for the habitat class that actually mattered. As Karen highlighted, "finer resolution data does not always equal better quality information." When decisions about carbon credits or conservation outcomes depend on these classifications, that gap matters.

Come find us at ML4EO 2026

We will be at ML4EO 2026, where we're presenting "Better Baselines for Carbon Markets: Using Earth Observation Foundation Models to Build Credible Counterfactuals."

This year's keynotes include Dr Emily Lines (Cambridge), who is developing new methods for monitoring forest structure and biodiversity using remote sensing from the ground to satellites, and Tomislav Hengl (OpenGeoHub) on open geospatial science, which we are big supporters of. It's a strong programme.

The workshops extend the conversation. Jakub Nowosad is running a session on where your models can be trusted, using spatial cross-validation to test whether predictions hold outside training areas. It's exactly the kind of rigour the field needs more of. IBM are back with hands-on sessions on TerraTorch and foundation model fine-tuning. And there are rumours of an introduction to TESSERA, the Cambridge geospatial foundation model, from the lead developer. For anyone working with earth observation embeddings, these are worth the trip alone.

If you're working in conservation monitoring, carbon markets, or environmental integrity and want to talk about what credible counterfactual baselines look like, come find us.

ML4EO 2026

The carbon market has an integrity problem. Everyone knows that. But it is not the one anyone is talking about anymore. The problem is in the carbon baseline.

The major market upset was driven by West et al. (2023), who argued that many forest carbon projects overestimated deforestation rates in their baselines (see our response to their article). We contest the methods used, but the authors raise a pertinent point: existing methods do not do a good enough job of causally measuring the true impact of carbon projects.

Since then, the market has shifted. There is a greater desire for "high-integrity" credits. Good. But many appear to wilfully ignore the criticisms the market has faced, claiming we can restore trust through better measurement and improved predictive modelling. Better measurement is welcome, of course. But while the industry obsesses over which biomass map to trust, the real uncertainty, the real risk, is somewhere else entirely. It is in the carbon baseline.

Efforts continue to eke out greater performance from forest allometry and machine learning upscaling. New global and regional canopy height models are released on a near-weekly basis, with growing potential for model ensembles and fine-tuning. Perhaps all of this is soon to be made redundant by ESA's P-Band SAR BIOMASS mission.

None of this is wasted effort. We have been building biomass maps for years and welcome every advance. But having built them ourselves across multiple carbon projects, we have come to realise something uncomfortable. The real uncertainty is not in precisely how many tonnes of carbon are in the forest. It is in what would have happened to the forest without the project. The spread between biomass maps is small. The spread between a good baseline and a bad one is enormous.

The carbon credit integrity crisis everyone is ignoring

Carbon credit integrity depends on many things. But the one question that determines whether a credit represents a real change in atmospheric CO₂ is simple: would it have happened anyway? Get that wrong and five different ratings agencies will each charge you to tell you the credit is worthless. It does not matter how precisely you measured the trees if your counterfactual scenario is more fiction than fact.

The industry knows this. It is exactly where past integrity failures came from. Carbon baselines that assumed the future would look like the past. Projects claiming credit for "saving" forests that were never under threat. That is why baseline has become a loaded term. People would rather focus on biomass maps, where you can at least get precision, than touch the baseline question, where you cannot hide behind a pretty map.

It is not just about the biomass maps

A road gets built. A policy changes. Commodity prices shift, or a new president takes office. The real trajectory of deforestation changes, and a historical baseline built from historical trends cannot keep up. It is looking backwards when it should be taking in the full picture. What is happening right now, to similar forests, that are not protected? That wider view is what matters.

Some have taken this further, arguing that because social-ecological systems are complex, causal attribution is fundamentally impossible and carbon markets should be abandoned altogether (Rana et al., 2026). We disagree. Complexity is not a reason to give up. It is exactly the kind of problem science exists to solve, and as we argued at ML4EO 2026, pixel-perfect classification means little if the landscape-level counterfactual is wrong.

The carbon market has spent years refining how we measure trees. It is time to be just as rigorous about what we compare these measurements against. Designed experiments like the Sabah Biodiversity Experiment show one path: build the controls in from the start.

belian.earth exists to solve the carbon baseline problem. Read about how counterfactual baselines work.

It was an early start as I was lucky enough to get a place on "The Davos Express", and it was a flashback to my yearly school trips on the Hogwarts Express - treating the journey as a chance to meet and catch up with contacts on the way from Zurich.

For many years there have been demonstrations against WEF, but now, since the start of the Climate Hub Davos run by Green Up Switzerland, nature, biodiversity, climate and community are part of the conversation. This is the place to be, with people from diverse backgrounds sharing their experiences and solutions for a greener future.

I joined Clarmondial's session 'How Localised Nature Data can Empower Business and Unlock Finance', which included some really interesting perspectives. We heard from scientists supporting sustainable cacao farmers in Indonesia, nature and climate risk specialists from reinsurance, and sustainable timber investment groups, amongst others. It was encouraging to see so many different perspectives all converging on protecting nature.

I was struck by a couple of comments. On assessing climate risks: "How you interpret data is crucial to interpreting risk." This resonated with me. There is so much talk about data, and so little consideration of experience interpreting it - something we constantly consider while building belian's products. On nature credits: "These already exist, and a market already exists - high quality carbon credits that can demonstrate their community and biodiversity benefits." There certainly seems to be consensus in many areas of the market.

Just as when skiing in Davos, restaurants tend to be a bit busy and pricey, so I got a packed lunch from Coop and found a bench in the sun as I often do at conferences. While eating a boiled egg and some fruit, I was lucky enough to be joined by various friendly people, including Cameron from Bottletop.

I learnt that #togetherband is made with Humanium - metal repurposed from decommissioned illegal firearms seized in Central America. The bands are handmade at an atelier in Nepal by women rescued from human trafficking. Their mission is to share the UN's 17 Sustainable Development Goals, and they do it in a beautiful way. As Cameron was tying this togetherband around my wrist he explained how the dismantling of guns is an important part of the process for the young men processing their trauma. Certainly a project I would like to support.

I had many conversations at the Climate Hub around the sustainable development goals, and I particularly appreciate learning about concepts that are not my core expertise. While belian may appear focused on life on land and climate action, all of the sustainable development goals align with our values. I wonder if the next togetherband might feature upcycled Bornean ironwood beads?

I then took a little break from work to look at skis. I had some really good advice from Silvan at Angerer Sport. I would certainly recommend checking out their ski test centre on Jakobshorn. I certainly plan to - we all need to take a break from Trump's new world order.

After years of working on forest carbon baselines, it's good to see the carbon market conversation maturing, and after years of WEF, it's good to see that nature is properly on the table, despite what some make you believe.

Forest carbon markets face a fundamental credibility challenge: how do you estimate what would have happened without a conservation intervention? Recent high-profile analyses have questioned whether existing carbon baseline methods deliver accurate results, a debate our team has contributed to directly through our critique of synthetic control approaches in Science.

This six-month feasibility study will let us systematically test emerging techniques in geospatial machine learning and causal inference against this problem, with a focus on quantifying uncertainty in counterfactual estimates.

Our team brings more than two decades of field research in tropical forests, including work published in Science on carbon recovery rates, combined with operational experience in forest carbon project development and assessment. We also bring extensive expertise in cloud computing and open-source tools for reproducible ecological workflows. ESA's support lets us bridge that applied research background with new computational approaches.

We'll be sharing methods and findings openly as the work progresses. Our aim is to contribute tools that make rigorous counterfactual baseline assessment more accessible across the sector.

With restoration and tree-planting projects back in the spotlight, partly driven by growing scrutiny of baselines in avoided-deforestation projects, we are revisiting a paper we published five years ago in Science: Active Restoration Accelerates the Carbon Recovery of Human-Modified Tropical Forests. The study combined two decades of permanent-plot measurements with high-resolution biomass maps, produced with airborne LiDAR, in Sabah, Malaysian Borneo. This allowed us to directly compare carbon recovery in areas that had been allowed to regenerate naturally with areas that had received active restoration, including enrichment planting of dipterocarps and routine cutting back of climbers and competing vegetation. Ultimately we found that in addition to the already impressive natural regeneration in this area, active restoration accelerated biomass accumulation substantially, around 50% faster than in the areas naturally regenerating. Follow-up research published in 2026 has since shown that climber (liana) cutting alone drives much of that gain — separating the contribution of seedling planting from climber cutting required a designed experiment with proper controls. As funding shifts towards restoration projects, often viewed as more 'additional' than avoided-deforestation, this work is relevant because it established a robust empirical counterfactual baseline on which to measure the efficacy of active restoration.

Why forest protection matters for forest carbon recovery

Natural regeneration performed well, but only because the forest had been protected from further disturbance. The surrounding forest, outside the study area, was logged repeatedly (Reynolds et al., 2011). Without sustained protection, the recovery rates we measured would not have been possible, and neither the impressive gains from natural regeneration nor the additional gains from active restoration would persist; both depend on the balance between growth and degradation. Growth and degradation are not separate silos; they result from interacting anthropogenic and natural processes that coexist and must be measured simultaneously.

The counterfactual question: what would have happened?

The growing scrutiny of baselines in avoided-deforestation projects essentially questions the counterfactual, i.e. what would have occurred in the absence of the conservation intervention and, more specifically, how much deforestation and degradation would have occurred. However, the factor that is often missed is that a counterfactual is only an estimate as the true baseline is unknowable. This is inevitable because observing "what would have happened" would require leaving the area unprotected and observing the deforestation rate, which is obviously incompatible with implementing a conservation project. There are, however, a variety of methods that enable us to estimate this counterfactual and quantify the uncertainty empirically, although these methods are still in development and remain underutilised in forest carbon projects.

Restoration projects need counterfactual baselines too

Just as with avoided deforestation projects, restoration and tree-planting projects also require a counterfactual to estimate their impact. In the restoration case, the primary questions are: if you hadn't planted a tree, what would have been there in its place? Bare ground, or perhaps some natural regeneration? And, if forest recovery is the most likely alternative scenario for the site, what would the rate of recovery have been? In both restoration and avoided-deforestation projects, the underlying question is the same: how does the carbon balance change over time compared with what would have happened without the intervention?

The real question is whether counterfactual baselines are robust

Although active restoration accelerated carbon recovery, the additional gains need to be considered alongside the restoration costs. Comparing the difference between natural regeneration and active restoration with global cost estimates makes it clear that, given the expense, restoration is more justified in the most degraded or disconnected areas where natural regeneration would be most inhibited. Recent global assessments also show that the majority of remaining forest carbon potential lies in existing forests, meaning that protection delivers much of the climate benefit, with restoration valuable where forests have been heavily degraded or removed (Mo et al., 2023). The baselines of avoided-deforestation projects have been criticised, but restoration projects also depend on a counterfactual baseline to demonstrate impact. The issue is not whether baselines exist, but whether they are robust. Impact evaluation of both restoration and avoided-deforestation ultimately relies on the same principle: clearly estimating the difference in carbon balance between the intervention and a robust counterfactual baseline.

Updated 14 March 2026: added subheadings for readability, updated SEO metadata and internal links.