The Carbon Signal We Can't Find: Why Atmospheric CO₂ Measurements Struggle to Verify What Countries Report
The Smoke Detector Problem
Imagine you've been cooking something on the stove. Smoke fills the kitchen. You crack open a window, turn on the fan, and wait for the air to clear. But the smoke detector never stops beeping — or worse, it never even goes off. That's roughly where we are with climate policy.
Countries spend enormous effort reporting their carbon emissions: coal burned in German power plants, gasoline consumed on Los Angeles highways, gas flared from oil rigs in the Niger Delta. These bottom-up inventories form the backbone of how the world tracks progress under the Paris Agreement. But according to a sweeping new analysis that maps the relationship between what we emit and what the atmosphere actually shows, the signal of human-caused carbon dioxide is frequently drowned out by the planet's own breathing.
The finding, published by a multi-institutional team spanning researchers in Germany, Serbia, Spain, the Netherlands, France, Italy, and Slovenia, is both humbling and clarifying. By combining satellite-reconstructed atmospheric CO₂ data with emission inventories, vegetation activity, and a measure of the El Niño cycle, the researchers found that natural carbon-cycle processes and global background trends dominate regional atmospheric CO₂ growth rates — so much so that even a dramatic, worldwide drop in human emissions might not show up clearly in local atmospheric measurements.
The year 2020 offered a near-perfect natural experiment to test this. COVID-19 lockdowns caused the largest single-year decline in fossil fuel emissions ever recorded — roughly 5.4 percent globally, according to the EDGAR inventory the researchers used. And 2020 fell during a neutral phase of the El Niño–Southern Oscillation, meaning there was no strong climate-driven interference. You might expect the atmospheric CO₂ fingerprint of that emissions drop to be visible across the globe.
It wasn't.
"The resulting reductions were not consistently reflected in regional atmospheric CO₂ growth rates," the researchers write — a spatial decoupling that underscores just how thoroughly biospheric dynamics, atmospheric transport, and what they call "regional carbon-cycle memory" can overwhelm the human signal.
This matters enormously. If policymakers can't reliably verify that emission cuts are working by looking at atmospheric measurements — the most direct evidence that emissions are actually changing atmospheric composition — then the entire framework for tracking climate progress rests on something harder to trust than it appears.
The Science: Reading the Atmosphere From Space
The study took a top-down approach. Rather than starting with reported emissions and asking what they might do to the atmosphere, the researchers began with what the atmosphere is actually showing — and worked backward to understand what drives it. To do this, they stitched together four datasets that capture different pieces of the carbon puzzle.
The first is the Copernicus Atmosphere Monitoring Service (CAMS) greenhouse gas reanalysis. This product reconstructs atmospheric carbon dioxide concentrations using a combination of satellite observations, atmospheric modeling, and data assimilation — essentially, it creates a globally consistent picture of CO₂ concentrations in the atmosphere every three hours, on a one-degree-by-one-degree grid (roughly 111 kilometers at the equator). The researchers didn't just look at concentrations; they calculated growth rates — how fast CO₂ was accumulating in a given location year to year — using a technique called Dynamic Linear Model fitting that strips away seasonal cycles to reveal the underlying trend.
The second dataset is the Emissions Database for Global Atmospheric Research (EDGAR), which provides bottom-up estimates of anthropogenic CO₂ emissions — fossil fuel combustion, cement production, and biofuel burning — on a fine 0.1-degree grid going back to 1970. This is the standard source for country-level emission reporting and the kind of data that fills the tables in international climate assessments.
The third captures what the biosphere is doing. The researchers used the Global OCO-2 Solar-Induced Chlorophyll Fluorescence (GOSIF) dataset, which measures a subtle light signal — fluorescence — emitted by chlorophyll in living plants. This signal, called SIF, is tightly linked to photosynthetic activity and serves as a real-time window into how much carbon vegetation is pulling from the atmosphere. When the Amazon rainforest photosynthesizes vigorously, it draws down CO₂; when a drought grips it, the drawdown weakens. GOSIF combines satellite observations from NASA's Orbiting Carbon Observatory-2 with land-surface data to produce a global map of this vegetation signal every eight days.
The fourth is the Southern Oscillation Index (SOI), a long-standing measure of the El Niño–Southern Oscillation that tracks pressure differences between Tahiti and Darwin, Australia. ENSO is one of the dominant sources of year-to-year climate variability on Earth; it shifts rainfall patterns, alters ocean upwelling, and affects how much carbon forests absorb or release. Previous research had shown that SOI leads changes in atmospheric CO₂ growth by about seven months, so the researchers applied that lag before using it as a predictor.
The core of the analysis was pixel-wise linear regression: at every land grid cell on the planet, the researchers asked whether the local atmospheric CO₂ growth rate in a given year was statistically associated with local anthropogenic emissions, local vegetation activity, global ENSO conditions, or some combination. They also ran models that used year-over-year changes in emissions and vegetation instead of absolute levels — a way to isolate the effect of shifts in activity from background levels.
To make sense of the global patterns, they applied unsupervised machine learning — specifically, K-means clustering — to classify the world's land surface into regions that behave similarly in terms of their carbon dynamics. They didn't tell the algorithm what to look for; they just asked it to find natural groupings in the data. Then they ran a "persistence analysis": instead of looking at a single snapshot, they tracked which grid cells ended up in the same cluster across all ten years of the study (2015–2024). Pixels that reliably clustered together over time — connected components in the clustering output — were identified as stable, persistent carbon-cycle regimes.
This yielded five large-scale regimes that the researchers labeled by their dominant characteristics: carbon-inactive polar regions, carbon-limited arid and boreal zones, carbon-transition drylands, carbon-source anthropogenic cores, and carbon-sink active biospheres. They also identified four smaller transitional clusters that shuttle between regimes depending on the year — suggesting that some parts of the world sit on fault lines in the global carbon map.
What They Found
The atmosphere speaks, but not in the language of emissions
The global mean atmospheric CO₂ growth rate in 2019 — the year before the pandemic — was 2.58 parts per million per year. That means each year, the atmosphere accumulated roughly 2.58 ppm of CO₂ above what natural processes absorbed. That number conceals enormous regional variation: some places saw growth rates above 4 ppm yr⁻¹, while others showed rates near zero or even negative — places where the land was pulling carbon out of the air faster than it was being added.
The regression models tell a striking story. When the researchers predicted atmospheric CO₂ growth rates using only local emissions and the global SOI index, they achieved close agreement with observed growth rates across most of the globe — even though the model was simple and didn't include vegetation directly. This sounds like good news for the emissions signal, but it's not: the SOI term, which captures the global, ENSO-driven component of CO₂ variability, was consistently the stronger predictor. In most regions, year-to-year swings in atmospheric CO₂ growth were driven more by El Niño and La Niña cycles — and the cascading effects they have on vegetation and ocean carbon uptake — than by local changes in human emissions.
Look at the year-by-year breakdown at nine representative locations that the researchers analyze in detail, spanning from New York City to the Amazon to the Sahara to boreal Alaska. At most of these locations, the model's prediction falls within the error bars of the observed growth rate — the model works. But the decomposition of why it works reveals something important: in almost every case, the dominant driver of predicted growth rate variation from year to year isn't the emission term. It's the global SOI term. The baseline — the intercept in the regression — accounts for the bulk of the absolute growth rate in most places. The emission signal, when it appears, is secondary to the climate signal.
Global Atmospheric CO₂ Growth Rates, 2015–2024
Estimated global mean atmospheric CO₂ growth rates from 2015 to 2024, showing interannual variability dominated by ENSO cycles and natural carbon-cycle processes rather than anthropogenic emission changes.
| Label | Value |
|---|---|
| 2015 | 2 ppm yr⁻¹ |
| 2016 | 3.2 ppm yr⁻¹ |
| 2017 | 2.1 ppm yr⁻¹ |
| 2018 | 2.5 ppm yr⁻¹ |
| 2019 | 2.58 ppm yr⁻¹ |
| 2020 | 2.3 ppm yr⁻¹ |
| 2021 | 4.8 ppm yr⁻¹ |
| 2022 | 2.6 ppm yr⁻¹ |
Global CO₂ growth rate in 2019 estimated from deseasonalized atmospheric measurements. The map shows substantial regional variation, with the global mean at 2.58 ppm yr⁻¹.
The five clusters tell a geographic story. Carbon-inactive regions (Antarctica, Greenland, and parts of the high Arctic) showed almost no association with either human emissions or biospheric activity — their atmospheric CO₂ growth simply followed the global background trend. Carbon-limited regions (Alaska, central Siberia, the Sahara) had low vegetation activity and minimal anthropogenic influence; they behaved almost like the poles, but with slightly more sensitivity to ENSO-driven climate swings. Carbon-transition regions (the central United States, parts of central Asia) showed modest influence from both drivers but nothing dominant. Carbon-source regions (industrialized corridors including the U.S. Northeast, parts of Europe, East Asia) had the clearest — though still modest — association with local emissions. And carbon-sink regions (the Amazon, central Africa, Southeast Asia) were dominated by biospheric dynamics: in the Amazon especially, the year-to-year variability in atmospheric CO₂ growth reflected how hard the rainforest was photosynthesizing, not how much CO₂ was being emitted locally.
The five persistent carbon-cycle regimes identified through unsupervised clustering, showing mean human-driven emissions (DSe) and bio-driven activity (GOSIF) within each cluster.
The 2020 anomaly: what the pandemic didn't show
The researchers describe 2020 as a "unique natural experiment," and it's worth dwelling on why. The COVID-19 pandemic produced what any climate scientist would call a massive perturbation to the human emissions system: a 5.4 percent drop in fossil fuel CO₂ emissions globally, the largest single-year decline since the industrial revolution. If any year should have shown up as a clear dip in atmospheric CO₂ growth, it would be 2020.
But the atmospheric record tells a murkier story. The researchers show that even in this year — during a neutral ENSO phase, meaning no strong climate interference — the regional atmospheric CO₂ growth rates didn't uniformly reflect the emissions drop. Some regions showed slower growth that year, but many didn't, because local vegetation dynamics, atmospheric transport patterns, and the accumulated carbon already present in regional atmospheric pools were dominating the signal.
This gets at something subtle and important: the atmosphere doesn't respond instantly to changes in surface emissions. There's a kind of memory in the system. CO₂ that was emitted in previous years continues to influence atmospheric concentrations. Air masses transport carbon from one region to another. And vegetation — the biosphere — responds to weather, not to human emission reports. A wet year in the Amazon, or a cool summer in North America, can swamp the atmospheric signature of millions of cars staying in driveways.
The researchers put it this way: even in the absence of strong climate forcing (i.e., a neutral ENSO year), local atmospheric responses are dominated by "biospheric dynamics, atmospheric transport, and regional carbon-cycle memory."
This is a profound challenge for climate policy. The Paris Agreement relies on countries reporting their emissions reductions. Those reports are then checked against atmospheric measurements as a form of independent verification. But if atmospheric measurements don't reliably reflect regional emission changes, that verification pathway is compromised.
What the clusters reveal about averaging
When the researchers aggregated their pixel-level data within the five persistent clusters and re-ran the regression models, they found something that cuts both ways. On one hand, spatial averaging within clusters largely smoothed out the unique regional growth patterns, leaving large-scale climate as the dominant control in most regimes. The carbon-source clusters, the carbon-transition clusters — in all of them, the SOI term explained most of the interannual variance after averaging.
On the other hand, the active biosphere cluster — the tropical rainforest regions — was the notable exception. In the Amazon and its analogs, the biogenic signal remained strong enough to survive the averaging process. The GOSIF vegetation signal continued to be a meaningful predictor of atmospheric CO₂ growth even at the cluster level, underscoring what the researchers call "the critical role of tropical forests in shaping atmospheric CO₂ variability."
Explained Variance by Driver and Carbon-Cycle Regime
Comparison of how much of the interannual variance in atmospheric CO₂ growth is explained by large-scale climate (SOI) versus vegetation activity (GOSIF) across five carbon-cycle regimes. The active biosphere is the only cluster where biogenic signals dominate.
| Label | Value |
|---|---|
| C1: Carbon Inactive (Polar) | 85 % of explained variance |
| C2: Carbon Limited (Arid/Boreal) | 72 % of explained variance |
| C3: Carbon Transition (Drylands) | 58 % of explained variance |
| C4: Carbon Source (Anthropogenic) | 45 % of explained variance |
| C5: Active Biosphere (Tropical) | 25 % of explained variance |
Annual predictor contributions to modeled CO₂ growth rate per persistent cluster for two model configurations: emissions plus SOI (left) and emissions plus GOSIF vegetation signal (right). The active biosphere cluster (C5) retains a strong biogenic signal even after spatial averaging, while other clusters are dominated by large-scale climate variability.
This matters because it identifies which regions are actually informative for atmospheric verification of emissions. Tropical forests are loud — their carbon signal is so strong that it can be seen even in coarse averages. But industrialized regions, where emission reductions are most politically and economically significant, are quiet: their atmospheric CO₂ signal is easily lost in the climate noise.
Why This Changes Things
The gap between accounting and atmosphere
There is a fundamental mismatch at the heart of climate policy that this paper makes viscerally clear. The world tracks emissions using political boundaries — a factory in Pennsylvania, a coal plant in Shandong, a wildfire in the Congo Basin. Each country tallies its contributions and reports them upward. The Paris Agreement, the national climate pledges, the race to net zero: all of it rests on this bottom-up accounting system.
But atmospheric CO₂ doesn't respect borders. It circulates globally. It responds to natural fluxes that dwarf the human signal in most regions at most times. And the relationship between what a country emits and what shows up in the atmosphere above that country is loose at best.
The researchers aren't the first to point this out. Previous work — by teams using ground-based measurement networks like TCCON, by satellite missions like GOSAT and OCO-2, by atmospheric inversion studies — has documented discrepancies between bottom-up inventories and top-down atmospheric constraints. What's new here is the global, gridded, year-by-year analysis that simultaneously quantifies the contributions of emissions, vegetation, and climate across the entire land surface. The scale and the integration are what make the result striking.
Think about what this means for verification. If a country claims a 10 percent reduction in CO₂ emissions between 2019 and 2020, and that claim is based on activity data (how much coal was burned, how much cement was produced), the only way to verify it from above is to look at the atmosphere. But if the atmospheric signal of that reduction is smaller than the natural year-to-year variability from ENSO, the verification fails. The signal-to-noise ratio is too low.
This isn't a reason to abandon emissions reporting — it's a reason to be honest about what atmospheric measurements can and can't do, and to invest in better ways of using them.
What the five regimes mean for climate monitoring
The five-cluster framework the researchers developed offers a practical geography of the carbon problem. It categorizes the world's land surface by the dominant control on atmospheric CO₂ growth, which has direct implications for what kind of monitoring is useful where.
In the carbon-inactive polar regions, there is essentially no local signal to detect. Atmospheric CO₂ growth tracks the global mean, and the best way to monitor change is the global mean itself — which is already being done by monitoring stations worldwide. No amount of regional monitoring will reveal more about these areas than the global trend.
In the carbon-limited arid and boreal zones — the Sahels and Alaskas of the world — the system is too sparse to generate a strong local signal. These regions are dominated by climate variability and have limited biospheric or anthropogenic influence. Monitoring them has limited value for verifying emission reductions; they are more useful as indicators of climate-driven carbon-cycle changes.
In the carbon-transition regions, the signal is mixed, which makes attribution difficult. A change in atmospheric CO₂ growth rate in these zones could come from emissions, from vegetation shifts, or from climate variability — and disentangling those contributions requires information beyond what atmospheric CO₂ alone provides.
In the carbon-source anthropogenic cores — the New York-to-Boston corridor, the Po Valley, the Pearl River Delta — the atmospheric signal should reflect human emissions most clearly. But even here, the analysis suggests, the signal is modest and competes with global climate variability. Verification of emission reductions in these regions requires careful accounting of the global ENSO contribution and, ideally, comparison with pre-industrial-era baselines that are difficult to establish.
In the carbon-sink active biosphere — the Amazon above all — the vegetation signal dominates. Monitoring atmospheric CO₂ in these regions is really monitoring vegetation health. That makes these sites enormously valuable for understanding the global carbon budget, but almost useless for verifying human emission reductions. An emission cut in São Paulo might be invisible against the atmospheric signature of an Amazonian dry season.
The five persistent clusters (C1–C5) identified through ten-year persistence analysis, with transitional clusters (T1–T4) shown as colored crosses that oscillate between primary clusters over time.
The biosphere as the great amplifier
Perhaps the most striking insight from this paper is the unique status of tropical forests in the global carbon system. The active biosphere cluster — dominated by the Amazon, but also including tropical Africa and Southeast Asia — is the only region where the biogenic signal survives spatial averaging. When you average over hundreds of grid cells across the entire Amazon basin, the vegetation signal is still the dominant predictor of atmospheric CO₂ growth. That's remarkable. It means that tropical forests are not just a large carbon sink (which scientists have known for decades); they're a coherent, large-scale carbon sink whose behavior is visible even in crude averages.
This is both a vulnerability and an opportunity. It's a vulnerability because it means that anything that disrupts tropical photosynthesis — deforestation, drought, fire, warming-induced tree mortality — will have an outsized effect on the global atmospheric CO₂ growth rate. The 2015–2016 El Niño caused a massive die-off in Amazonian vegetation, and atmospheric CO₂ growth rates spiked globally that year. The 2019–2020 fires in Australia had similar, if smaller-scale, effects. These are not minor perturbations; they're visible in the global mean.
It's an opportunity because it means that satellite measurements of vegetation activity — the SIF signal that the researchers used — may be a more reliable indicator of what's happening to the global carbon budget than emissions inventories in many regions. If the atmosphere's CO₂ content is being overwhelmingly shaped by whether tropical forests are photosynthesizing hard or going dormant, then the thing to watch is the forests, not the smokestacks.
What's Next
Open questions and honest limitations
The researchers are appropriately careful about what their analysis can and can't claim. A few caveats deserve attention.
First, the study covers only 2015–2024 — a decade, which is substantial but not long in climate terms. ENSO cycles every two to seven years, and the researchers acknowledge that a longer record would help distinguish between genuine associations and decadal coincidences. Ten years of data is enough to identify patterns, but longer records would strengthen confidence in the results and help characterize how stable the five clusters actually are over multi-decadal timescales.
Second, the atmospheric CO₂ growth rates are derived from a reanalysis product — CAMS — not direct satellite observations. Reanalysis combines models and data, which is powerful but introduces model dependency. The researchers acknowledge that CAMS has been validated against ground-based TCCON measurements and shows "acceptable agreement," but they also note that the product involves assumptions about atmospheric transport and chemistry that could introduce biases. Future work using direct satellite column observations — from missions like GOSAT-3 or the upcoming CO2M mission — could sharpen the top-down picture.
Third, the regression models are statistical associations, not causal claims. The researchers are careful to say that $\beta_E(r)$ represents the "local emission-related association with atmospheric CO₂ growth," not a direct measurement of how much of the CO₂ in the atmosphere above a pixel came from local emissions. Atmospheric transport moves CO₂ far from its sources, and the regression framework doesn't fully account for this. Causal attribution of CO₂ sources to atmospheric signals requires more sophisticated inverse modeling — a whole separate field of atmospheric science — which the researchers acknowledge as a direction for future work.
Fourth, the study looks only at CO₂. Methane, nitrous oxide, and other greenhouse gases have their own atmospheric dynamics and policy relevance, and the interaction between them — the "carbon-climate feedback" — is a major area of ongoing research.
There is also the question of what the findings mean for near-term climate action. The paper suggests that atmospheric monitoring of CO₂ alone, at regional scales, is insufficient for verifying emissions reductions. This points toward a need for multiple lines of evidence: not just atmospheric CO₂ measurements, but satellite-based estimates of fossil fuel combustion (using NO₂ as a proxy, for example), city-level energy data, and continuous ground-based monitoring networks. The authors suggest their framework could be extended to include additional variables that might help isolate the human signal — things like atmospheric carbon monoxide (a combustion byproduct), or satellite-derived measurements of power plant activity.
What this opens up
Despite the limitations, the paper offers a genuinely useful lens for thinking about the carbon cycle — one that integrates human and natural drivers, that maps them geographically, and that honestly acknowledges where the signals are strong and where they're weak.
The five-cluster framework is a starting point. The researchers suggest it could be extended by including additional Earth-system variables — ocean carbon fluxes, soil moisture, fire activity — and by applying the same analysis to other greenhouse gases. The transitional clusters they identified (T1–T4) — regions that oscillate between regimes depending on the year — are particularly interesting and underexplored. A drought year might shift part of the Amazon from carbon sink toward carbon source; a mild winter might shift parts of the carbon-transition zones toward carbon-limited behavior. Understanding what controls these regime transitions is important for predicting how the carbon cycle will behave under a warming climate.
The finding about 2020 is a cautionary tale that extends beyond COVID-19. It suggests that short-term emission perturbations — whether from economic slowdowns, fuel switches, or policy changes — may be invisible against the atmospheric noise in most regions. This doesn't mean emission reductions don't matter; it means that their atmospheric signature may take years to emerge, and verifying them requires looking at longer timescales and larger spatial scales than any single country or region.
At the same time, the dominance of the active biosphere in the signal suggests that protecting and restoring tropical forests is not just one item on the climate agenda — it may be the most visible thing humans can do for the atmospheric carbon signal. A world with a healthy Amazon draws down more CO₂ than a world with a degraded one, and that difference is, in principle, measurable from space.
Finally, the gap between bottom-up inventories and top-down atmospheric measurements that the paper quantifies is a call for better integration of the two approaches. Several international efforts are already working toward this — the global stocktake process under the Paris Agreement includes components that compare reported emissions with atmospheric observations, and the upcoming CO2M satellite mission (part of Europe's Copernicus program) is explicitly designed to provide authoritative measurements of atmospheric CO₂ for policy verification. This paper provides a scientific foundation for why that's hard and why it matters.
The atmosphere is not a simple ledger. Emissions go in, but the balance shown at any given location, in any given year, reflects a tangle of natural processes, transport dynamics, and accumulated memory that can easily overwhelm the human fingerprint. That's a fact. What we do with it — how we build better monitoring systems, more honest accounting, and more realistic expectations for what atmospheric verification can deliver — is a choice.