AI reveals environmental drivers of East Coast carbon fluxes

Every forest, marsh, and meadow along the U.S. East Coast inhales and exhales carbon dioxide in a silent, ceaseless rhythm, and researchers have now taught machines to read that breathing with unprecedented precision. In a study published on 5 August 2026 in the journal Earth Science Informatics, a team at Shanghai Ocean University’s College of Oceanography and Ecological Science unveiled a machine-learning framework that reconstructs two decades of terrestrial carbon exchange along the densely populated seaboard, from northern hardwood forests to southern coastal wetlands. Using random forest models trained on long-term tower-based eddy covariance observations, the researchers predicted gross primary productivity, ecosystem respiration, and net ecosystem productivity with coefficients of determination reaching 0.88 — decisively outperforming existing global carbon-flux products such as FLUXCOM. Because these three quantities together determine whether a landscape stores carbon or releases it into the atmosphere, the achievement could sharpen regional carbon accounting at a moment when coastal ecosystems face intensifying pressure from climate change, sea-level rise, and relentless urban development.

Carbon fluxes are the currency of the terrestrial carbon cycle. Gross primary productivity, or GPP, quantifies the total carbon dioxide that plants strip from the air through photosynthesis. Ecosystem respiration, ER, counts the carbon returned to the atmosphere as plants, microbes, and soils break down organic matter. The difference between these two large opposing flows — net ecosystem productivity, NEP — reveals whether an ecosystem functions as a carbon sink or a carbon source. Pinning these terms down is far from academic: the terrestrial biosphere absorbs a substantial share of humanity’s carbon emissions each year, yet the strength of that sink varies enormously between regions and years, and some of the largest uncertainties in the global carbon budget lie in exactly the kind of heterogeneous, human-dominated landscapes that characterize the U.S. East Coast, where fragmented forests, agricultural land, tidal wetlands, and sprawling metropolitan areas intermingle within a few hundred kilometres of coastline.

The ground truth for the new models comes from eddy covariance, the gold-standard technique for measuring ecosystem-scale gas exchange. Towers rising above the canopy carry fast-response sonic anemometers and infrared gas analysers that record vertical wind speed and carbon dioxide concentration dozens of times per second. Because turbulent eddies continuously shuttle air between the surface and the atmosphere, the covariance between fluctuations in vertical wind and fluctuations in the CO2 mixing ratio yields a direct, continuous measurement of net exchange over a footprint of roughly a square kilometre. Networks such as FLUXNET and its North American branch AmeriFlux, whose contribution the authors gratefully acknowledge, have accumulated decades of such records worldwide. The catch is that a tower sees only its own footprint; extending those point measurements into wall-to-wall regional maps requires models that translate satellite observations and meteorological reanalysis into flux estimates — the very task the Shanghai Ocean University team set out to improve.

To build that translation engine, the researchers assembled a multi-scale dataset spanning roughly twenty years along the East Coast. Tower observations of carbon fluxes were fused with vegetation products from NASA’s MODIS sensors aboard the Terra and Aqua satellites and with ERA5-Land, the state-of-the-art global reanalysis dataset produced by the Copernicus Climate Change Service and the European Centre for Medium-Range Weather Forecasts, which integrates vast streams of observations into a physically consistent land-surface record. Before any machine learning began, the team screened candidate environmental variables with correlation analysis and with the Geodetector method, a spatial-statistics technique that quantifies how much of the spatial heterogeneity in a target variable can be explained by stratifying the landscape according to a potential driver. Unlike correlation alone, Geodetector captures nonlinear relationships and interactions without assuming any particular functional form, which makes it well suited to disentangling the tangled influences of climate, vegetation, and terrain on carbon exchange across sharply contrasting ecosystem types.

That screening produced a winning recipe of eight predictors: T2M, the two-metre air temperature; VPD, the vapour pressure deficit that measures how thirsty the atmosphere is; SSRD, the downward solar radiation that powers photosynthesis; EVI, the enhanced vegetation index that tracks canopy greenness from space; LSWI, the land surface water index that reflects vegetation and soil moisture; LAI, the leaf area index describing how much photosynthetic surface the canopy exposes; EVAVT, an evaporation-related variable; and DEM, the digital elevation model that encodes terrain and elevation. Each predictor represents a distinct physiological lever. Temperature and radiation set the pace of the enzymatic machinery of photosynthesis and respiration; VPD governs whether leaf stomata stay open or clamp shut to conserve water; and the satellite-derived vegetation indices record the structural and phenological state of the canopy itself. When this eight-factor combination was fed into the models, it delivered the most accurate and stable flux estimates of any input set the team tested, a result that held across ecosystem types and across all three carbon fluxes.

With the inputs fixed, the researchers staged a head-to-head contest between four widely used machine-learning algorithms: random forest (RF), artificial neural network (ANN), support vector regression (SVR), and extreme gradient boosting (XGBoost). Each brings a different inductive bias. Random forests grow hundreds of decision trees on random subsets of the data and average their predictions, a bootstrap-aggregation strategy that suppresses overfitting. Neural networks stack layers of weighted neurons that can approximate highly nonlinear functions but demand careful tuning and abundant training data. Support vector regression fits a margin-tolerant function in a high-dimensional kernel space, while XGBoost builds trees sequentially, each new tree correcting the residual errors left by its predecessors. Under identical input combinations, the random forest emerged as the consistent champion for GPP, NEP, and ER alike. The outcome suggests that the ensemble’s robustness to noisy inputs and its resistance to overfitting on heterogeneous, multi-decadal observations outweigh the theoretical flexibility of the more elaborate architectures for this regional estimation problem.

The verification statistics are the study’s headline numbers. The trained models achieved coefficients of determination — R² values — of 0.88 for gross primary productivity, 0.81 for ecosystem respiration, and 0.55 for net ecosystem productivity against independent observations. Benchmark models from FLUXCOM, the leading international initiative that upscales eddy covariance data into continuous global flux products, achieve R² values of only 0.61, 0.57, and 0.28 for the same three fluxes. In practical terms, the new models explain roughly 88 percent of the observed variance in photosynthetic carbon uptake where the existing standard explains about 61 percent, and they nearly double the explained variance for the net sink term. NEP remains the hardest target for any modelling approach, and for an instructive reason: it is a small residual obtained by subtracting two enormous numbers, so even modest independent errors in GPP and ER compound into a large relative error in their difference. Even so, leaping from 0.28 to 0.55 transforms a product of marginal utility into one capable of resolving meaningful spatial and interannual variation.

Perhaps the most consequential scientific insight lies in the variable-importance analysis. Across every ecosystem type examined, the enhanced vegetation index ranked as the single most important driver of carbon-flux variability, elevating vegetation growth status above temperature, radiation, and atmospheric dryness as the master control on carbon exchange along the East Coast. The interpretation becomes intuitive once spelled out: photosynthesis and respiration are both carried out by the canopy and the organisms it sustains, so the state of the vegetation — how green it is, how much leaf area it displays, how its phenology unfolds through the seasons — effectively sets the stage on which all the climatic actors perform. Climate variables still matter, and the prominence of vapour pressure deficit echoes a growing body of evidence that atmospheric drought is an increasingly dominant brake on carbon uptake as the planet warms. But the finding suggests that satellite observations of greenness, already among the most widely available remote-sensing products, carry exceptional predictive power for regional carbon monitoring.

Beyond the headline accuracy, the framework hands researchers and policymakers a two-decade, multi-scale portrait of how carbon moves through one of the most economically and ecologically important regions of North America, resolving spatiotemporal patterns across ecosystem types that coarse global products tend to blur. Such maps can inform forest management and coastal restoration decisions, provide benchmarks against which Earth system models can be evaluated, and support the emerging carbon-accounting infrastructure that depends on credible, spatially explicit baselines. The authors note that the underlying data will be made available from the corresponding author upon reasonable request, and they report that the work proceeded without dedicated external funding. Limitations remain, most notably the difficulty of the net flux term and the dependence of any upscaled product on the density and quality of tower coverage. What the study ultimately delivers is a methodological template: pair rigorous driver screening with carefully benchmarked machine learning, ground everything in decades of direct flux measurement, and the breathing of entire landscapes becomes not just measurable at a handful of towers, but legible across an entire coastline, every day of the year.