The argument rests on a subtle but important distinction that applied mycology has long recognized yet rarely built into risk assessments. Every mycotoxigenic fungus has two overlapping but distinct ecological ranges: its growth window, the full set of temperature and water activity conditions under which it can colonize a substrate, and its toxin-production window, the narrower range in which the biosynthetic machinery for a given toxin is actually switched on. In nearly every species studied, the toxin-permitting range sits inside the growth range like a kernel inside a shell. Most models have treated these boundaries as fixed species properties, measured once on synthetic laboratory media such as potato dextrose agar and applied as constants ever since. Medina argues that this treatment is doubly flawed: fungi experience the food matrix itself as part of their environment, and boundary values measured on agar can differ considerably from those on the actual grain, nut or fruit at risk.

The factor that forced a reconsideration of these fixed windows is carbon dioxide. When elevated CO2 is added to the interaction between temperature and water availability in experimental systems, it does not simply increase toxin output under already favorable conditions. Instead, it widens the range of conditions under which toxin is produced at all, drawing the toxin-production window closer to the growth window. In studies of Aspergillus flavus on maize, interacting elevated CO2 with drought stress and increased temperature raised the expression of aflatoxin biosynthetic genes aflD and aflR and boosted aflatoxin B1 production, in some cases several-fold, under conditions that previously supported growth only. A parallel pattern, in which fungal growth changes little while toxin production expands, has been documented in pistachio-associated A. flavus and grape-associated A. carbonarius, although the degree and direction of the shift vary by species, strain and substrate.

There is also evidence that these windows are not merely plastic but can shift with repeated exposure. Strains of Fusarium langsethiae, F. poae and F. sporotrichioides sub-cultured for ten generations at 1000 parts per million of CO2 altered their subsequent growth and toxin-production responses relative to non-acclimatized strains, suggesting that successive generations can acclimatize to a changing atmosphere. Medina stresses that the direct CO2 effect has so far been demonstrated most convincingly in Aspergillus systems, and that confirming it across the wider fungal community remains a research priority. Still, the emerging picture is that climate change, acting through the interaction of temperature, water activity and CO2, does not merely make existing conditions more frequent. It modifies the ecological space available to toxigenic fungi, and often widens it.

The second half of the argument concerns timescale. Research on climate change and mycotoxins since the early 2000s has focused on combined rather than single factors, because studies varying one variable at a time consistently mis-estimated fungal behavior compared with combined-factor data. What has changed is the frequency with which these factors now cluster within a single cropping season rather than being spread across successive years. A season that once brought a single anomalous event, an early heat spike, a mid-season drought, a late intense rainfall, can now deliver several in sequence in the same field. Meanwhile, crop pests and pathogens have been shifting polewards at a mean rate of 2.7 kilometers per year since 1960, and temperature-dependent infection risk for fungal crop pathogens is projected to rise fastest at higher, increasingly productive latitudes. A framework built around isolated, spaced-out anomalies simply cannot describe a regime in which extremes compound within one crop cycle.

At the community level, this has a counterintuitive consequence. If growth and toxin windows widen and shift for multiple species at once, the previously separate temporal and spatial niches occupied by Fusarium, Aspergillus, Penicillium and Alternaria species within a crop cycle begin to overlap. Species that once took turns now compete for the same favorable conditions, and the outcomes are strain- and species-specific rather than predictable from single-factor data. Wider ecological windows, in other words, do not simply mean more toxin from one fungus. They may mean a larger, more crowded set of species with a genuine opportunity to contribute toxin to the same commodity at the same time. Medina proposes that the degree of overlap between the growth window and the toxin window could serve as a practical ecological metric for anticipating mycotoxin risk under climate change, one that can be estimated for a given species and matrix rather than assumed constant.

If the mechanism is general, its fingerprints should be visible in field surveys that were never designed to test it, and the evidence assembled in the Perspective is strikingly consistent with that prediction. Household maize in Kenya has been found to carry 22 different mycotoxins spanning both Aspergillus and Fusarium products, alongside 31 metabolites of A. flavus, with sharp regional differences between aflatoxin- and fumonisin-dominant areas. Brazilian cocoa has been reported carrying up to 15 mycotoxins from three fungal genera simultaneously. An 11-year survey of northern Italian maize found more than 85 percent of samples contaminated with two or more mycotoxins from Fusarium and Aspergillus jointly, and a seven-year survey across 23 European countries documented co-occurrence rates of 52 to 74 percent between trichothecenes and other Fusarium-derived emerging toxins. A ten-year global screen of 74,821 animal feed samples from 100 countries found that 64 percent of samples tested for three or more toxins contained at least two.

Multi-species coexistence is also visible within single crops, and species substitutions track the changing climate. In maize, hotter, drier conditions favor a shift from Fusarium verticillioides towards Aspergillus flavus dominance, and one modeling study projects a 92 percent increase in aflatoxin B1 risk in European maize under a two-degree warming scenario. PCR surveys of Spanish vineyards have found A. flavus on 67 percent of organically grown berries, A. niger on 44 percent and A. carbonarius on 14 percent, three toxigenic species routinely sharing the same fruit. On cereals, F. graminearum has displaced F. culmorum as the head-blight pathogen of Danish wheat, while Fusarium head blight has intensified under warming in southern England and central China, the latter accompanied by a shift toward a more toxigenic chemotype. The pattern even persists through processing: a UK survey of plant-based meat alternatives found emerging Fusarium toxins in up to 98.9 percent of samples and Alternaria toxins in up to 90.4 percent, while 78 percent of wheat flour samples from Shandong, China, carried deoxynivalenol together with two Alternaria toxins.

Medina is careful to flag the central confounder. Improved analytical capability, particularly broader liquid chromatography tandem mass spectrometry panels, undoubtedly accounts for a substantial share of the co-occurrence now being recorded, and multi-mycotoxin contamination is unlikely to be an entirely new phenomenon. But the same trend appears across independent surveys differing in continent, commodity and study design, which is compatible with a genuine biological contribution layered on top of better detection. Climate change should be viewed as a major interacting driver rather than the sole determinant, since agronomic practices, cultivar choice, trade patterns, storage conditions and sampling strategies all shape what is recorded. The proposed framework is offered explicitly as a plausible hypothesis rather than an established transition.

The practical implications are considerable, because the infrastructure of mycotoxin management remains anchored in a world of single toxins and single species. Predictive models that treat temperature, water activity and CO2 as separate, additive factors will systematically underestimate risk, since it is their interaction that widens the toxin window. Regulatory limits calibrated around single-compound thresholds do not address the toxicological consequences of co-occurring mycotoxins. Medina calls for surveillance to move to routine multi-mycotoxin analysis and near-real-time monitoring, noting that low-cost CO2 sensing has recently been proposed as an early warning signal because fungal CO2 output rises before toxin becomes detectable. Mitigation strategies validated under current conditions need re-testing under interacting future scenarios, and regulators from EFSA to Codex Alimentarius should treat multi-mycotoxin contamination not as an anomaly but as the baseline against which exceptionality is judged. Because the same crops, climate drivers and contaminated commodities flow through livestock and the wider environment, the author frames this as a One Health problem, shared by human, animal and environmental health sectors. The Perspective leaves readers with a memorable reframing: the black swan was never disappearing, it was multiplying, and the flock, not the single bird, is now the more accurate description of the risk.

Subject of Research: Climate change effects on fungal ecology and the risk of multi-mycotoxin contamination of crops and food

Article Title: Beyond the black swan: climate change and the evolving risk of multi-mycotoxin contamination

Article References: Medina, A. (2026). Beyond the black swan: climate change and the evolving risk of multi-mycotoxin contamination. Frontiers in Fungal Biology, 7, Article 1981215. https://doi.org/10.3389/ffunb.2026.1981215

Image Credits: AI Generated

DOI: 10.3389/ffunb.2026.1981215

Keywords: mycotoxins, climate change, fungal ecology, food safety, Aspergillus flavus, aflatoxin, Fusarium, CO2, ecological niche, surveillance, One Health, food contamination

The researchers, led by Arnau Garcia Mesa, focused on six locations chosen to span a wide range of climatic behavior: Córdoba in southern Spain, Marrakech in Morocco, Lyon in France, Hannover in Germany, Belgrade in Serbia, and Stockholm in Sweden. These sites run the gamut from water-limited regimes, where soils are almost always dry in summer, to energy-limited regimes, where abundant moisture means the land surface and atmosphere interact very differently. For each location, the team assembled more than seven decades of data from the ERA5 and ERA5-Land reanalyses, covering daily maximum temperatures, large-scale atmospheric circulation fields, soil moisture at three depth levels, and seasonal carbon dioxide concentrations measured at NOAA’s Mauna Loa Observatory.

The machine learning framework at the heart of the study is a hybrid architecture the authors call the CombinedModel. Atmospheric data, which arrives as spatial maps of geopotential height at 500 and 200 hectopascals and sea-level pressure, is processed by a ConvNeXt convolutional neural network, a modern image-recognition architecture that had not previously been applied to driver quantification in this context. Local land-surface data and carbon dioxide values, which arrive as simple time series, are handled by a multilayer perceptron. The outputs of both branches are merged in a final classification network that decides, day by day, whether conditions resemble a hot extreme, defined as a daily maximum temperature exceeding the 90th percentile of the 1950–2000 climatology. Because extreme days are rare, the team used a weighted loss function to prevent the model from simply defaulting to predicting ordinary weather, and they trained an ensemble of twenty members to tame the randomness of network initialization.

But prediction alone was not the goal. The real innovation lies in the interpretability layer. Using SHAP values, short for SHapley Additive exPlanations, the researchers borrowed a technique from game theory that assigns each input feature a precise, additive share of the model’s output. Positive SHAP values push the model toward predicting an extreme; negative values push it away. Because the values are additive, they can be summed across grid points and variables to produce a global accounting of driver importance. The team computed these explanations only for the fifty percent most confident extreme predictions in the 2014–2024 test period, ensuring the attribution rested on predictions the network itself trusted.

The headline result is strikingly consistent: atmospheric circulation dominates everywhere. Across all six locations, circulation features accounted for 66 to 86 percent of the total explanatory power, with the geopotential at 500 hectopascals, the mid-tropospheric field that meteorologists use to spot blocking highs and ridges, contributing the most by a wide margin. This aligns with a growing body of work linking European heatwaves to persistent double jets over Eurasia, deep North Atlantic depressions, and Saharan warm-air intrusions, and it echoes a 2025 global analysis finding that 500-hectopascal geopotential is the dominant heat-extreme driver over most of the world’s land area. In other words, the large-scale shape of the atmosphere is the primary engine of a hot day, almost regardless of where you are.

The land surface, by contrast, tells a sharply regional story. Soil moisture’s share of the explanation follows a clear northward gradient: a negligible 0.2 percent in Marrakech, a moderate 9.4 percent in Córdoba, and a substantial 15.4 percent in Lyon, with similarly elevated values in the temperate and northern sites. The pattern makes physical sense. In already-arid Marrakech, summer soils are so consistently dry that year-to-year variability is small, leaving little for the land surface to modulate. In wetter transitional climates, however, a dry soil anomaly can dramatically amplify a heatwave by suppressing evaporation and shifting the surface energy balance toward sensible heating. The SHAP analysis captured this mechanism directly: for the shallowest soil layer, negative moisture anomalies were consistently associated with positive contributions to extreme-heat predictions, while deeper layers showed weaker and less robust signals, likely because neighboring soil layers are so strongly correlated that the model cannot cleanly separate their effects.

Carbon dioxide emerged as the third actor, and its role was revealing. The CO2 feature carried more than ten percent of the explanatory weight at most sites, and its importance was highest precisely where the yearly count of extreme days showed the strongest warming trend. By including CO2 explicitly, the framework controlled for the long-term anthropogenic signal without having to detrend the data, allowing a cleaner attribution of the physical drivers operating on daily timescales. The researchers stress that the explainability outputs represent the model’s interpretation rather than a direct measurement of reality, but the fact that the identified contributions reproduce well-established physical relationships lends the approach considerable credibility.

Robustness checks formed a major part of the study. When the team redefined extremes using the 80th percentile threshold, the same architecture recovered the same picture: circulation still dominated at 64 to 84 percent, and the northward gradient in land importance persisted, though soil moisture’s relative share shrank, suggesting that land-atmosphere feedbacks are not fully activated in milder heat events. A 95th-percentile model, hampered by too few extreme samples, proved unreliable and was set aside. The team also swapped reanalysis data for observations, using the E-OBS gridded temperature dataset and drought indices computed from it as soil-moisture proxies. The Standardized Precipitation Evapotranspiration Index, which incorporates evapotranspiration and thus better represents the surface water balance, reproduced the original results closely, while the precipitation-only Standardized Precipitation Index captured less of the land signal, a finding with practical implications for applying the framework to climate models, where soil layer definitions differ.

Two case studies brought the method down to the scale of individual disasters. For the August 2021 heatwave that pushed temperatures in Córdoba province to a scorching 47.6 degrees Celsius, the model attributed the event overwhelmingly to circulation, with positive SHAP values appearing precisely over the Iberian Peninsula where positive 500-hectopascal geopotential anomalies were located, a remarkable demonstration of spatial locality in the network’s reasoning. Intriguingly, CO2 ranked as the second most important feature even for this single event, hinting at a detectable anthropogenic fingerprint in an individual heatwave. The July and August 2018 heatwave in Hannover told a different story: the shallow soil-moisture layer played a much larger role, with SHAP values at times rivaling the circulation signal, consistent with the wet transitional regime of northern Germany, where the exceptional 2018 drought turned abundant soils into amplifiers of extreme heat.

The authors see their framework as more than an attribution exercise. Because it is data-driven yet physically interpretable, it could be applied to climate model output, to subseasonal-to-seasonal prediction systems, or to new regions, helping to diagnose where models capture or miss key feedbacks and to target climate risk assessments where they matter most. Future work, they note, should test whether the results hold across different network architectures and whether adding humidity fluxes or aerosol information sharpens the picture further. For now, the study offers a compelling demonstration that artificial intelligence, when forced to explain itself, can do more than forecast the weather; it can tell us, in numbers, why the weather turned deadly.

Subject of Research: Drivers of summer heat extremes in Europe and North Africa quantified with explainable machine learning

Article Title: Quantifying atmospheric and land drivers of hot temperature extremes through explainable Artificial Intelligence

Article References: Garcia Mesa, A., Palma, L., Donat, M., Materia, S., Gràvalos Talló, B., & Marcos Matamoros, R. (2026). Quantifying atmospheric and land drivers of hot temperature extremes through explainable Artificial Intelligence. Weather and Climate Dynamics, 7(3), 1709-1731. https://doi.org/10.5194/wcd-7-1709-2026

Image Credits: AI Generated

DOI: 10.5194/wcd-7-1709-2026

Keywords: heatwaves, explainable AI, SHAP values, atmospheric circulation, soil moisture, land-atmosphere coupling, climate change, CO2, machine learning, temperature extremes, ERA5, Europe

The new dataset, described in a preprint under review for the journal Earth System Science Data, addresses a problem that has quietly plagued atmospheric science for years. The Global Atmosphere Watch program, coordinated by the World Meteorological Organization, maintains monitoring stations from the pristine summits of remote islands to densely populated continental interiors. These stations deliver some of the most precise CO2 measurements on Earth, but their hourly records are far from continuous. Gaps arise from flask sampling regimes, instrument interruptions, and the exclusion of data during quality-control screening. Those missing hours are not merely cosmetic blemishes; they can systematically bias daily, monthly, and annual mean estimates of CO2, mask the natural rhythm of day-night concentration swings, and hide short-lived events such as pollution episodes or wildfire plumes.

To repair the record, the team, including Tan Mi, Xueting Pu, Xi Zheng, and colleagues, developed a two-stage reconstruction framework that blends direct observations with independent model-based estimates of atmospheric CO2. The approach integrates the GAW measurements themselves with CO2 fields from CarbonTracker and the Copernicus Atmosphere Monitoring Service, two widely used modeling systems that simulate how carbon dioxide is emitted, transported, and absorbed across the globe. Alongside these, the framework draws on anthropogenic emission inventories, vegetation data, meteorological variables, and spatiotemporal predictors — a rich feature set that allows the reconstruction to capture the physical processes driving CO2 variability at each location rather than simply interpolating between sparse points.

The first stage of the framework tackles the easier problem: filling in missing hours on days when at least one valid observation exists. For these partial days, the researchers employed hour-pair models, which learn the statistical relationships between concentrations at different times of day, to reconstruct the absent hours and produce complete daily profiles. This intra-day gap filling achieved impressive validation statistics, with a coefficient of determination of 0.96 and a root-mean-square error of 3.65 parts per million. In practical terms, the reconstructed values track the true hourly concentrations closely enough to preserve the diurnal cycle — the daily breathing of the biosphere as plants photosynthesize by day and respire by night, and as human activity pulses through morning and evening rush hours.

The second stage confronts the harder challenge: long stretches of time when a station produced no valid observations at all. Here, the team trained a globally trained initial prediction model, which learns from the full network of stations how CO2 behaves across seasons, latitudes, and land-cover types, and then applied station-specific post-corrections to tailor the predictions to each site’s local conditions. This full-sequence reconstruction achieved a coefficient of determination of 0.93 with a root-mean-square error of 4.76 parts per million. Across all stations, the mean filling rate reached 68.9 percent, meaning that on average more than two-thirds of the hourly record was reconstructed rather than directly observed, with a mean reconstruction uncertainty of 2.80 parts per million — a figure that gives users an honest sense of where the data are measured and where they are modeled.

Perhaps the most consequential finding is what happens when the repaired records are fed into the standard global analysis. Using the World Data Centre for Greenhouse Gases framework, the researchers compared global monthly mean CO2 concentrations computed from their reconstructed dataset against those derived from the original, gappy GAW records. The reconstructed values were, on average, 0.54 parts per million higher, with an uncertainty of plus or minus 0.21 parts per million. That may sound like a rounding error in a gas that now exceeds 420 parts per million globally, but in the precision world of greenhouse gas accounting, half a part per million is far from trivial. It suggests that incomplete temporal coverage has been measurably skewing global monthly estimates, potentially affecting the trend lines and seasonal cycle analyses that underpin carbon budget assessments worldwide.

The dataset also proves its worth at the scale of individual events. When compared with grid-cell simulations from CarbonTracker and CAMS, the reconstructed records retained stronger station-scale diurnal variability — the fine-grained local texture that coarse global models tend to smooth away. More dramatically, the reconstruction captured CO2 enhancements during the 2023 Canadian wildfire period, one of the most extreme fire seasons in modern history, when smoke and combustion gases drifted across the continent and beyond. A dataset that can resolve such transient carbon pulses opens a window on processes that monthly means and coarse models simply cannot see, from the immediate atmospheric fingerprint of a burning forest to the hourly interplay between urban emissions and local weather.

The implications reach well beyond record-keeping. Regional carbon-emission inversion — the family of techniques that work backward from atmospheric concentrations to estimate where and how much carbon is being emitted and absorbed — depends critically on accurate, well-characterized concentration data. Inversions are only as good as the observations that constrain them, and gappy records force researchers to either discard valuable information or make assumptions that propagate uncertainty into their emission estimates. A temporally continuous hourly dataset with quantified uncertainty gives inversion scientists a far richer constraint, potentially sharpening estimates of regional carbon fluxes and improving the evaluation of emission reduction policies. The dataset equally supports long-term assessments of global and regional surface CO2 changes and fine-scale analyses of variability that were previously impractical.

The release comes at a moment when the demand for granular carbon data is accelerating. Governments, cities, and corporations are increasingly committing to net-zero targets, and verifying progress toward those targets requires knowing not just annual averages but the dynamics of CO2 in the atmosphere at fine temporal and spatial resolution. Satellites such as those in the Orbiting Carbon Observatory family provide global coverage but with their own gaps and retrieval uncertainties, while ground stations provide the precision anchor points against which satellite data are calibrated. A complete, hourly, quarter-century ground-based record strengthens that entire observational chain, offering benchmark data for model evaluation, satellite validation, and the detection of subtle changes in the global carbon cycle.

The dataset, publicly available through Zenodo, arrives as a preprint currently open for community discussion and peer review at Earth System Science Data, a journal specializing in the publication of high-quality observational datasets. That review process will subject the reconstruction methods to external scrutiny, but the framework’s validation statistics and its demonstrated ability to reproduce both long-term means and short-lived events give the approach a strong foundation. For a field that has spent decades carefully measuring a changing atmosphere, the message of this work is at once sobering and empowering: the gaps in our records were large enough to matter, and now, with a combination of machine learning, physical modeling, and meticulous validation, they can be filled. As the planet’s carbon trajectory becomes ever more central to policy and public life, tools like this reconstruction ensure that the scientific record keeps pace with the atmosphere it is trying to understand.

Subject of Research: Reconstruction of temporally continuous hourly CO2 concentrations at Global Atmosphere Watch stations from 2000 to 2024

Article Title: A global, temporally continuous, hourly CO2 concentration dataset at GAW monitoring stations for 2000–2024: Enabling fine-scale carbon dynamics analysis

Article References: Mi, T., Pu, X., Zheng, X., Nie, G., Liu, X., Zhang, H., Grieneisen, M. L., Zhan, Y., Wang, M., & Yang, F. (2026). A global, temporally continuous, hourly CO 2 concentration dataset at GAW monitoring stations for 2000–2024: Enabling fine-scale carbon dynamics analysis. https://doi.org/10.5194/essd-2026-554

Image Credits: AI Generated

DOI: 10.5194/essd-2026-554

Keywords: CO2, GAW stations, carbon cycle, gap filling, CarbonTracker, CAMS, greenhouse gases, dataset reconstruction, diurnal variability, carbon emissions, wildfires, WDCGG

The challenge the team confronted is a familiar one in Earth science: sparse and biased data. Ground-based monitoring stations provide exquisitely accurate measurements of CO2 at the surface, but they are unevenly distributed, clustered heavily in North America, Europe, and East Asia while leaving large swaths of Africa, South America, and the oceans nearly unmeasured. Satellite instruments such as OCO-2, which measures column-averaged CO2 rather than surface concentrations, offer global coverage but observe only in cloud-free conditions and record the total amount of gas through the atmosphere rather than the concentration where it interacts with ecosystems and human populations. Numerical transport models can bridge these gaps, but they run at coarse resolutions and depend on uncertain estimates of emissions and atmospheric mixing.

To overcome these limitations, the researchers developed a LightGBM-based prediction model, a gradient boosting decision tree algorithm known for its efficiency with large datasets. The model ingested an unusually rich set of input variables: OCO-2 satellite retrievals of column CO2, ground observations from monitoring stations, meteorological and climatic variables drawn from the ERA5 reanalysis, a vegetation index derived from MODIS, and anthropogenic indicators including nighttime lights, population density, road networks, and fossil fuel emission inventories. By learning the statistical relationships between these predictors and the ground truth of station measurements, the model could estimate daily near-surface CO2 concentrations at locations and times where no direct measurement exists.

The performance figures are striking. The model achieved a correlation coefficient of 0.89 and a root-mean-square error of 3.5 parts per million against independent validation data, a level of accuracy the authors note surpasses many regional transport models. That precision matters because the differences being resolved are subtle: the global annual mean near-surface CO2 concentration over 2015 to 2021 came out at 408.71 plus or minus 2.98 parts per million, rising at an average rate of 2.66 plus or minus 0.27 parts per million per year. Those numbers align closely with estimates from the World Meteorological Organization, providing confidence that the machine learning reconstruction captures real atmospheric behavior rather than statistical artifacts.

Perhaps the most consequential finding is geographic. High growth rates in near-surface CO2 were concentrated in Southeast Asia, the South China Sea, West Africa, and the Amazon, with the steepest single-year increase occurring in 2016, a year influenced by an exceptionally strong El Nino event that suppressed tropical carbon uptake and fueled widespread fires. The Amazon result is particularly sobering. The team found that near-surface CO2 in that region grew faster than the column-averaged values measured higher in the atmosphere, a divergence that reflects dynamics unique to the surface layer, where the weakening of the forest carbon sink and fire emissions leave their most direct fingerprint. Long-term studies have documented a declining capacity of mature Amazon forests to absorb carbon, and the new dataset provides daily, spatially explicit evidence of how that decline manifests in the air itself.

South Asia emerged as another standout region. As a zone of intense and rising carbon emissions, it exhibited both higher and more variable near-surface CO2 concentrations than most other parts of the world. The daily resolution of the dataset allowed the researchers to distinguish persistent elevated concentrations from short-lived spikes, information that monthly or annual products simply cannot deliver. This variability matters for emissions verification: a region whose concentrations swing widely requires different monitoring and policy responses than one with a stable but high baseline.

Understanding why concentrations vary where they do required opening up the machine learning model itself. Using interpretability techniques rooted in Shapley value analysis, the team quantified the influence of each environmental driver across different climate zones. The results revealed a striking regional divide in the physics and biology controlling surface CO2. In tropical regions, temperature exerted a strong positive influence on near-surface concentrations, consistent with enhanced ecosystem respiration in warm conditions. In arid regions, by contrast, evaporation and soil type emerged as the dominant positive factors, suggesting that dryland soils and moisture dynamics play an underappreciated role in modulating how much CO2 lingers near the ground.

These insights translate directly into mitigation thinking. The authors identify soil improvement and large-scale afforestation as potential strategies for reducing CO2 levels in the high-concentration areas their maps reveal, since healthier soils and expanding forests can shift the local carbon balance toward uptake. While no amount of tree planting can substitute for cutting fossil fuel emissions, the dataset makes it possible to target such nature-based interventions at the specific landscapes where surface concentrations are rising fastest, a level of precision that has been impossible until now.

The daily cadence also unlocks a dramatic new capability: detecting short-term CO2 surges caused by large-scale wildfires. The analysis showed that major fires elevated surface CO2 by up to 3.45 parts per million, signals that could be traced in the daily maps as flames swept through fire-prone regions. Because wildfire emissions are notoriously difficult to verify, and because fire activity is increasing in many parts of the world under climate change, this capability supports regional emissions verification and near-term carbon assessment in ways that static inventories cannot. The data underlying the study have been made publicly available, and the framework is designed to be extendable, meaning the same approach could be updated with newer satellite generations and longer records. As the world races to track its progress under the Paris Agreement, a daily, high-resolution, ground-level view of the planet’s most important greenhouse gas may prove to be one of the most valuable tools yet developed.

The OCO-2 mission, launched by NASA in 2014, was designed primarily to track sources and sinks of carbon dioxide by measuring sunlight reflected off the planet’s surface in narrow spectral bands sensitive to the gas. Its retrievals, however, represent the average concentration through the entire atmospheric column, which is why translating them into estimates of the near-surface layer required the kind of statistical bridging this study provides. By pairing column measurements with ground stations that sample air close to the surface, the machine learning framework effectively learned how to downscale and translate between these two very different observational perspectives.

The choice of LightGBM reflects practical considerations as much as scientific ones. Gradient boosting decision trees handle nonlinear interactions among predictors without requiring assumptions about the underlying relationships, and LightGBM’s histogram-based approach makes training feasible on the enormous volume of data involved in daily global mapping. The team also employed seasonal-trend decomposition, a well-established statistical technique, to separate long-term trends from seasonal cycles in the concentration records, allowing the growth rate estimates to be computed on a cleaner signal.

The 2016 peak in concentrations deserves particular attention. The El Nino conditions of 2015 to 2016 brought drought to tropical Asia and the Amazon, reduced photosynthetic carbon uptake, and intensified biomass burning, producing the largest annual rise in atmospheric carbon dioxide on record at that time. That the dataset captures this episode in near-surface detail, particularly over fire-affected regions, serves as an independent check on its fidelity to known atmospheric events.

Validation against independent station measurements, rather than the data used for training, lends credibility to the reported accuracy figures. The close agreement between the estimated global growth rate of 2.66 parts per million per year and values derived from satellite-based analyses of column carbon dioxide further suggests the reconstruction is consistent with established observational records.

Making the underlying dataset openly accessible is a meaningful contribution in itself. Researchers studying regional carbon budgets, ecosystem responses, or urban emissions can now overlay daily near-surface concentrations with their own data, potentially accelerating work that previously depended on sparse station networks or coarse model output.

Subject of Research: Global daily near-surface CO2 mapping using OCO-2 satellite data and machine learning

Article Title: Estimation, variations, and impact factors of high-resolution global daily near-surface CO2 during 2015–2021 based on OCO-2 and machine learning

Article References: Liu, R., Wang, X., Ren, Y., Tao, C., Ji, S., Gao, Z., Jiang, Y., Ren, S., Fang, L., Chen, J., Zhang, Q., Wang, G., & Wang, Q. (2026). Estimation, variations, and impact factors of high-resolution global daily near-surface CO2 during 2015–2021 based on OCO-2 and machine learning. ENGINEERING Environment, 20(10), Article 148. https://doi.org/10.1007/s11783-026-2248-z

Image Credits: AI Generated

DOI: 10.1007/s11783-026-2248-z

Keywords: CO2, OCO-2, machine learning, LightGBM, remote sensing, carbon cycle, climate change, wildfires, Amazon, greenhouse gases, satellite data, environmental science