AI Maps Daily Global CO2 at Ground Level With Unprecedented Detail

Carbon dioxide is invisible, well-mixed, and yet profoundly uneven in the ways it accumulates near the ground, where people actually live and breathe. A research team led by scientists at Shandong University in China has now built the most detailed daily picture yet of near-surface CO2 across the entire planet, using satellite observations from NASA’s OCO-2 mission and a machine learning framework to fill in the vast gaps that satellites and ground stations leave behind. The resulting dataset, published in Frontiers of Environmental Science & Engineering, covers every day from 2015 through 2021 at a spatial resolution of 0.5 degrees by 0.625 degrees, offering researchers and policymakers a powerful new lens on how the greenhouse gas driving climate change behaves in the lowest layer of the atmosphere.

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