Accurate greenhouse gas (GHG) emissions estimations from hydropower reservoirs are critical for ensuring that this renewable energy source effectively contributes to climate mitigation. In this paper, we critically review and compare a range of methodologies, including direct field measurements (e.g. floating chambers and eddy covariance), empirical/statistical models (like G-res and HydroCalculator), process-based simulations, satellite remote sensing, machine learning (ML) techniques, and hybrid modelling frameworks that integrate these components. Our analysis evaluates each approach against key criteria: accuracy and uncertainty, scalability and transferability, and data requirements and transparency. Direct measurements remain the gold standard for site-specific validation; however, they are limited by spatial and temporal coverage and demand substantial resources. Empirical models offer simplicity but struggle to capture dynamic environmental drivers, often leading to under- or overestimation of emissions. Process-based models provide critical mechanistic insights but require extensive input data and computational resources. While satellite observations and ML enhance spatial and temporal coverage and predictive capability, explainable AI can overcome the “black-box” nature of ML Hybrid approaches that combine in-situ data, remote sensing, ML, and process-based elements show the most significant promise for an accurate and scalable emissions estimation. As the European Union and other regions work to meet stringent climate targets, robust reservoir GHG accounting is essential for guiding investments and driving mitigation actions in genuinely low-carbon hydropower. Our findings highlight the necessity of integrated monitoring networks, open-access data, and interdisciplinary collaboration to develop next-generation tools that bridge precise measurement with large-scale modelling for informed climate and energy policy. • Critical review of GHG estimation methods for hydropower reservoirs conducted. • Traditional and emerging approaches are systematically analysed and compared. • Evaluation focuses on accuracy, scalability, uncertainty and data needs. • Hybrid models with ML and remote sensing show highest potential. • Research gaps and future directions identified for policymakers and stakeholders.

Hurricanes have significant consequences for ecosystems, potentially disrupting the carbon cycle at both local and regional scales and releasing carbon back into the atmosphere through storm-associated impacts on vegetation and agricultural areas. The present work analyzes the interactions amongst terrestrial carbon fluxes, rainfall, and land cover for three significant hurricanes: Harvey (Texas), Irma (Florida), and Maria (Puerto Rico). This study utilized net ecosystem exchange (NEE) data derived from the Soil Moisture Active Passive (SMAP) NASA satellite mission, which provides global estimates of soil moisture and carbon flux, and analyzed these data for coastal climate zones during the hurricane season. The results were validated using eddy covariance tower-based in-situ CO 2 flux observations during hurricane landfall. Results showed that southern Texas (Harvey) experienced the highest amount of carbon release (0.33 megatons), followed by Florida (Irma) (0.03 megatons) and Puerto Rico (Maria) (0.02 megatons). The land cover products, such as the National Land Cover Dataset (NLCD) and the Copernicus Global Land Service (CGLS), showed overall reductions in land cover in Florida (-1.02%), Texas (-0.97%), and Puerto Rico (-0.46%). Furthermore, vegetation cover changes were estimated using MODIS-derived enhanced vegetation index (EVI), showing major changes over Puerto Rico (-3.81%) and southeast Texas (-2.94%), while normalized difference vegetation index (NDVI) showed more moderate reductions over Puerto Rico (-3.06%), southeast Texas (-1.12%), and Florida (-0.16%). These reductions indicate short-term vegetation stress and decreased photosynthetic activity, which may temporarily reduce carbon uptake, leading affected regions to transition from carbon sinks to temporary carbon sources. These findings highlight hurricanes as significant drivers of short-term carbon emissions and vegetation change. This study enhances understanding of hurricane-associated disturbances in the carbon cycle by examining spatial and temporal variations in carbon fluxes during extreme weather events. • The impacts of hurricanes Harvey, Irma, and Maria (2017) on terrestrial carbon fluxes were assessed. • Carbon release during and after landfall was measured using SMAP-derived NEE and eddy covariance CO 2 flux data. • Hurricane Harvey resulted in the largest carbon emission (0.33 megatons), followed by Irma (0.03 megatons) and Maria (0.02 megatons). • Vegetation loss, derived from MODIS NDVI and land cover change products, was greatest in Texas (6,199.3 km 2 ), then Florida (492.92 km 2 ), and Puerto Rico (14.53 km 2 ). • Vegetation declines led to reduced photosynthetic activity, temporarily turning affected areas from carbon sinks into carbon sources. • Findings highlight hurricanes as significant short-term drivers of carbon emissions and ecosystem disturbance across coastal regions.

Wirelessly controlled modular automatic chambers for greenhouse gas flux monitoring in natural and agricultural ecosystems.

The optimal temperature of photosynthesis (Topt) generally increases with plant growth temperature. Changes in Topt are associated with changes in the maximum carboxylation capacity at 25°C (Vcmax25) and the maximum electron transport rate at 25°C (Jmax25). The ratio between Jmax25 and Vcmax25 declines with warming. Accurate representation of leaf-level photosynthetic responses to temperature is essential for realistic projections of the terrestrial carbon cycle and its response to ongoing climate changes. However, many land surface models incorporate thermal acclimation through empirical approaches and through assigning distinct but static parameter values to plant functional types (PFTs). Eco-evolutionary optimality (EEO) approaches provide a simpler way of modelling photosynthesis without recourse to PFTs. Here, we use the subdaily P model, an EEO-based model of photosynthesis that explicitly separates the instantaneous and acclimated responses of photosynthetic parameters to temperature to investigate how optimal temperature changes with growth temperature, as represented by leaf or air temperature. We show that the simulated responses are consistent with observations from both controlled experiments and eddy covariance flux tower data. We show that changes in Topt, and in the assimilation rate at Topt, are caused by changes in carboxylation capacity and electron transport rate that follow directly from the hypotheses underlying the model.

Impacts of surface heterogeneity on energy partitioning in paddy ecosystems: A dual eddy covariance study

ABSTRACT Precise quantification of gross primary productivity (GPP) at fine resolution is crucial for regional carbon cycle analysis. However, existing GPP products frequently fail to capture fine-scale variability due to an inherent trade-off between satellite spatiotemporal resolution and the coarse-scale design of most existing GPP models. To address this issue, a machine learning-based downscaling and correction framework was designed to generate high-accuracy monthly GPP at 30-m resolution (DS-FC-GPP) through fusing multiple remote sensing model-based products along with eddy covariance (EC) data. Moderate resolution imaging spectroradiometer (MODIS) and global land surface satellite (GLASS) GPP products were simultaneously downscaled from 500 m to 30 m resolution using a fine-resolution vegetation index time series. Subsequently, a unified fusion and correction framework was established based on EC data. The analysis revealed that DS-FC-GPP exhibited enhanced spatial patterns with finer details and maintained consistency with the original products when aggregated back to coarse resolution. In the spatial validation, the leave-one-site-out validation achieved a coefficient of determination (R 2) of 0.828 with a mean absolute error (MAE) of 39.497 gC·m−2·month−1, while the leave-one-region-out scheme yielded a comparable performance with an R 2 of 0.812 and an MAE of 42.506 gC·m−2·month−1. The consistently strong performance across both validation strategies demonstrates the robustness of the proposed framework. Moreover, DS-FC-GPP better captured intra-annual peaks and valleys, maintaining stable accuracy throughout the year. The transfer experiment over China indicated that the DS-FC-GPP framework exhibits robust generalization and transferability. The present work highlights the potential of multisource data fusion and downscaling to promote fine-resolution GPP estimation, offering valuable insights into terrestrial carbon cycle processes.

Bacterial growth efficiency (BGE), defined by the balance between bacterial production (BP) and respiration (BR), is a key indicator of microbial metabolism and carbon fluxes in freshwater ecosystems. Although bottom-up controls on BGE are well recognized, the relative importance of top-down processes, particularly viral lysis and grazing, remains poorly understood. To evaluate the influence of mortality pathways on bacterially mediated carbon fluxes, we conducted a seasonal time-series study in Lake Cournon, sampling both the euphotic zone and bottom waters. BGE exhibited marked seasonal variability (14-53%) that exceeded depth-related differences, reflecting shifts in the balance between BP and BR associated with bacterial biosynthesis and carbon mineralisation. No significant relationships were observed between BGE and bottom-up controls, suggesting a stronger role for mortality-driven processes. Viruses negatively affected BGE, likely through preferential lysis of highly active high-nucleic-acid bacterial populations that contributed substantially to bacterial productivity. Viral lysis appeared to stimulate BR more than BP in the remaining bacterial community through the release of dissolved organic matter, resulting in lower BGE at the community level. In contrast, size-selective grazing by nanoflagellates exerted a positive influence on BGE, likely by recycling nutrients and stimulating the growth of active bacterial populations. Overall, our findings demonstrate that the mode of bacterial mortality can strongly regulate bacterial carbon metabolism and alter the pathways of organic carbon flow in freshwater ecosystems.

Abstract Potential evapotranspiration (PET), defined as the evapotranspirative flux from a region under fully saturated conditions, is a critical variable in hydrologic modeling, water stress assessment, and understanding ecosystem responses to climate. The widely used Priestley–Taylor method provides a simple, low-data requirement approach for estimating PET. However, it uses a fixed coefficient (α PT = 1.26) in most applications, but this oversimplification neglects biome-specific variability, limiting its accuracy across diverse environments. Although numerous studies have attempted to derive dynamic characterizations of α PT , most estimates are developed for either obtaining reference evapotranspiration (ET 0 ) or actual evapotranspiration (ET). Consequently, these approaches do not provide α PT that accounts for ecosystem-specific aerodynamic and plant conductance constraints required to fully represent true ecosystem-level PET. In this study, we utilized 3,128 site-years of eddy covariance data from 246 FluxNet sites worldwide to optimize α PT values across a broad range of biomes. Results showed significant spatial and seasonal variability in α PT , with higher values in forests and winter months and lower values in savannas’ summer. Temperature and radiation emerged as key drivers of this variability. Using the influencing variables, we next derived functional equations to estimate α PT based on key bio-environmental variables. These equations yielded demonstrable improvements in PET estimates, and can be directly incorporated into land surface and hydrologic models, and generation of remote sensing products. Furthermore, PET derived using our simplified functional equations of α PT , when applied to obtain the Evaporative Stress Index (ESI) and ET, yielded improved estimates of both ET and plant stress. Overall, these findings offer a more ecologically representative approach to PET estimation using the Priestley-Taylor method, with broad implications for hydrologic modeling and drought assessment.

Rapid mid-winter and early spring warming events are emerging as a key but under-recognized driver of carbon loss from cold-region ecosystems. In mountain peatlands, their influence remains largely unknown. Here, we quantify the impact of chinook wind events-warm, dry downslope winds that can raise air temperatures by over 20°C within hours-on ecosystem respiration in a montane peatland on the eastern slopes of the Canadian Rockies. Using eddy covariance carbon dioxide (CO2) flux measurements and meteorological data, 13 chinook events were identified between February and May 2021 and segmented each into pre-, during-, and post-event phases. A generalized additive mixed model accounting for temporal autocorrelation showed that CO2 emission rates increased significantly during chinook events and remained elevated afterward. CO2 emissions during snow-covered events in February-April were most likely driven by physical degassing from melting snow and thawing surface peat, whereas snow-free events in April and May likely reflected enhanced microbial activity in thawing peat. Our findings demonstrate that regularly occurring cold-season warming events can trigger substantial but short-lived CO2 releases from mountain peatlands, revealing a climate-sensitive carbon loss pathway likely to intensify as snowpack duration shortens and freeze-thaw regimes shift in mountain regions world-wide.

Changes in climate are altering plant growth patterns and associated phenological events like the Start of Season (SOS), End of Season (EOS), and Length of the Growing Season (LGS). However, there is limited research quantifying the impact of these changes on key vegetation-atmospheric interaction processes such as the carbon and water cycles. This study uses 914 site years of data across 132 flux tower sites in the FLUXNET2015 dataset to explore the relationships between carbon sequestration, expressed by Gross Primary Productivity (GPP), and multiple phenological variables, including LGS, changes in SOS (ΔSOS), and changes in EOS (ΔEOS). LGS explains 23% of the variability in GPP across all sites. Significant correlations were found in deciduous broadleaf forests (R² = 0.5) and evergreen needleleaf forests (R² = 0.44), while ecosystems such as shrublands, savannas, and wetlands displayed weaker connections. Changes in the SOS also affected GPP, with an earlier SOS increasing the total annual GPP. Deciduous Broadleaf Forests (R² = 0.54), Evergreen Needleleaf Forests (R² = 0.5), Grasslands (R² = 0.47) showed a significant negative association between ΔSOS and ΔGPP, whereas Croplands showed weaker correlations. Conversely, EOS variations had little impact on GPP. Upscaled to global vegetated land area these relationships suggest that each additional day in the growing season could increase carbon uptake by 1.035 Gt C yr- 1, while an earlier SOS by 0.93 Gt C yr- 1 and a one-day delay in EOS by approximately 0.65 Gt C yr- 1. These findings underscore the need to account for seasonal shifts and phenological changes in global carbon models.

Abstract Evapotranspiration (ET) and photosynthesis are key processes in ecosystem functioning on which soil moisture (SM) has an important influence. Eddy covariance measurements and machine learning (ML) increasingly enable flux prediction in ungauged regions. With numerous SM estimation methods available, each representing different spatiotemporal scales, understanding how data choice influences ML predictions is important. Our study examines how different SM data affect ML predictions of ET and photosynthesis. At semi‐arid to arid sites, we found that in situ near‐surface SM enhances ML predictions of ET. For photosynthesis, SM memory, indicative of deeper SM control, shows the highest predictive power, improving predictions by up to 30% at the driest sites. These contrasting responses reveal that ET and Gross Primary Productivity (GPP) are likely governed by distinct SM mechanisms: spatial scale matching for ET and temporal depth for GPP. Our study demonstrates that process‐guided feature engineering can improve ML predictions where root‐zone SM observations are often unavailable.

Abstract. Multirotor drones (part of the category of small Uncrewed Aerial Systems [sUAS] or small Uncrewed Aerial Vehicles [sUAV]) are used in atmospheric research to make measurements of the lower atmosphere, and their use is poised to increase in the future. New drone atmospheric sensing opportunities, such as ride-along applications and drone swarms, are emerging. These opportunities, which may not allow room for specialized shielding or aspiration equipment, together with increased drone usage, necessitate the characterization of the performance of unshielded sensors mounted to drones if the accuracy of such observations is to be understood. In this work, we characterize the accuracy of thermodynamic measurements, specifically temperature and water vapor mixing ratio, based on the sensor mounting position onboard multirotor drones. To assess the influence of the drone mechanics on the measurements, ninety-eight individual drone flights with eight distinct thermodynamic sensor positions were performed next to an instrumented flux tower and a tethersonde carrying identical sensors, where the tower and tethersonde measurements are assumed as truth. The flights were at least nine minutes in length, and nine of the flights were conducted at night. At the best position, absolute daytime temperature errors were between −0.83 and +0.61 K at the 95 % confidence interval, while nighttime temperature errors were smaller, ranging from −0.28 and +0.48 K. Water vapor mixing ratio errors are within −0.22 and +0.66 g kg−1. We conclude that measurements in field campaigns are more accurate when sensors are placed away from the main body of the drone and are sufficiently aspirated, such as a position near, but not directly under, a spinning propeller.

Storage change ( S ) is an important component of the mass balance equation and quantifies the accumulation or depletion of matter in the studied control volume, under the eddy covariance (EC) sensor. The quantification of S is required to estimate surface fluxes. This study compared four methods for calculating S of CO 2 , based on EC and profile measurements at a Danish temperate forest ICOS site (DK-Sor). The 12-heights sequential sampling system quantified in- and above-canopy S . Its design and physical averaging properties were thoroughly described. Two vertical configurations of the profile system were analyzed: (i) top-tower, (ii) full profile (incorporating all levels), along with two alternative calculation methods based on top-tower EC data alone, including the method proposed in an often used software. Results showed that the deviations between the S methods had a seasonal course and that the top-tower profile was on average 21% lower than the full profile method. The choice of S method also impacted the surface flux estimations on an annual scale, with relative differences in net ecosystem exchanges of up to 8%, represented by 22 g-C m −2 yr −1 . The S methods impacted the friction velocity threshold determination, leading to a variation in the amount of data retained during low-turbulence filtering. The full profile retained the most data. Lastly, the tailor-made calculation from EC concentration measurements were shown to fit the top-tower profile measurements closer, compared to the EddyPro-calculated S . These results highlight the importance of accurate storage change measurements in high and dense forest sites. • Tall and dense forest sites require a profile storage change system. • The uncertainty of top-tower storage measurements shows a seasonal course. • Storage change calculation methods impact the net annual CO 2 ecosystem exchanges. • Accurate storage change estimation decreases the u ∗ threshold and retains more data.

• Mass spectrometry based eddy covariance quantifying NH 3 emission. • NH 3 emissions up to 1.4 μgN m -2 s -1 during application of solid fraction of manure. • Only ∼2.25 % of loss of TAN as NH 3 . Ammonia (NH 3 ) fluxes were measured at a research farm in Southern Finland from 16 March to 22 May 2023 using the eddy covariance technique with a benzene chemical ionization mass spectrometer. NH 3 emissions (0-0.02 µgN m -2 s -1 ) remained low for the primarily studied field until its fertilization with the solid fraction from solid-liquid separation of cattle manure on 27 April, which led to fluxes up to 1.4 µgN m -2 s -1 . The manure was incorporated into the soil by harrowing the next day, after which the fluxes declined rapidly and remained below 0.1 µgN m -2 s -1 . The NH 3 emission factor (EF) from the solid fraction was 2.25% of applied total ammoniacal nitrogen (TAN) for the whole campaign, with 85% of the emissions during 27 April to 22 May occurring within 30 h. This EF lies at the low end of wide range of values reported for unincorporated solid cattle manure in literature (2.5–70%) depending on manure type and management. The comparatively low emissions in this study likely reflect the characteristics of the separated solid fraction and rapid post-application incorporation. Long-term background NH 3 concentrations and fluxes were generally low and broadly linearly dependent on temperature, while no clear dependence on windspeed was observed. The measured background concentrations (0.3±0.2 µg m -3 ) were similar to as reported for a nearby urban-background site in Finland (0.2±0.3 µg m -3 ) but much lower than values observed at agricultural hotspots in Europe (>10 µg m -3 ).

Wetland ecosystems, crucial for carbon sequestration and coastal hazard mitigation, have experienced tremendous losses in land surface area over the last century, primarily due to land reclamation. This has led to increased rates of land subsidence in regions with high levels of reclamation, causing heightened vulnerability in these areas under anticipated scenarios of climate induced sea-level rise. This study integrates multi-sensor satellite remote sensing (optical, thermal, and active microwave) with spatially explicit eddy covariance flux measurements to model gross primary productivity (GPP) in restored wetlands of California's Sacramento-San Joaquin Delta. GPP is a crucial process in this context, as it impacts the potential for wetlands to act as land carbon sinks and has been shown to reverse land subsidence in restored wetlands of previously reclaimed areas. Still, there remain gaps in understanding how vegetation vigor, wetland composition and structure, and environmental conditions individually and interactively impact carbon assimilation in these ecosystems. This research aims to understand how complementary remotely sensed signals from multiple satellite platforms across the electromagnetic spectrum combine to improve classical, optically based GPP models, while determining the relative importance of certain biotic and abiotic environmental conditions that regulate GPP. Using a Bayesian generalized additive modeling framework, we evaluated how vegetation vigor (NDVI), canopy structure and biomass density (microwave backscatter), and land surface temperature affect wetland GPP at 10-m spatial resolution over a five-year period. Our results reveal a strong hierarchical and complementary influence of these variables, with the highest GPP occurring in warm, well-watered, and densely vegetated conditions. The model explained on average 66% of GPP variability and provides a scalable, open-access framework for assessing carbon fluxes in wetland landscapes. These findings offer valuable insight into planning restoration, monitoring restoration outcomes, carbon accounting, and identifying coastal adaptation strategies for valuable blue carbon ecosystems. • We develop a wetland GPP model using eddy covariance and remote sensing data fusion. • Hierarchy of wetland GPP drivers revealed through remotely sensed observations. • Wetland GPP shows greatest sensitivity to interaction of drivers. • Hierarchical model enables informed upscaling of wetland GPP to landscape level.

Scientists want to know everything, everywhere, and all the time. This is particularly true in Earth science, where we seek to understand processes that span from the molecular to the planetary scale in how the world works, how it affects us, and how we impact it-especially the water cycle. Evapotranspiration (ET) was the last component to be measured in closing the water cycle: for decades, closing the water budget meant adding up all the measurable components, then inferring ET as the residual. Early measurements relied on water loss from pans and weighing lysimeters, followed by sensors inserted into plants to monitor sap flow and leaf chambers capturing transpiration. Scaling up to ecosystems became possible through eddy-covariance flux towers and further across landscapes through proximal sensing with drones, aircraft, and, ultimately, with satellites. While enormous progress has been made to measure or estimate ET everywhere and all the time, no single approach has yet achieved both simultaneously. Flux towers help with all the time, but not everywhere. Satellites can do everywhere, but not all the time (except, in part, for geostationary satellites, though with insufficient spatial coverage and resolution). A new advent of smallsat constellations is moving us to everywhere and all the time in detail, though we are only in the beginning of that era. This paper discusses the evolution and revolution of Earth observation for ET, as we advanced from the first Landsat and development of ET models through the progression of increasingly higher spatiotemporal resolution across international space agencies and commercial industry with increasing ET model sophistication, cloud computing, and machine learning. We continue to march ahead towards ET everywhere, all the time, and use that knowledge to better manage water and sustain our planet.

Measurement of methane fluxes (FCH4) from natural systems, such as wetlands, has lagged far behind carbon dioxide fluxes. Short and fragmented wetland FCH4 data limit our ability to assess its long-term dynamics and potential climate feedbacks. Extrapolating short-term FCH4 records to recent decades remains challenging for both process-based models and data-driven machine learning (ML) approaches. Here, we develop a knowledge-guided ML framework that integrates eddy covariance (EC) FCH4 observations, field warming experiments, and biogeochemical knowledge to reconstruct the long-term FCH4 budgets and trends. Focusing on the 11 longest EC monitoring sites in the AmeriFlux network, we found considerable variability in multi-decadal trends of wetland FCH4, with increases up to 14% per decade from 2000 to 2024. We also found that the strength of these increasing trends declines from high to low latitudes, highlighting the vulnerability of northern wetlands. This work presents novel and robust reconstructions of long-term wetland FCH4, offering critical benchmark datasets for bottom-up ecosystem models and advancing fundamental understanding of wetland biogeochemistry.

The lack of energy balance closure in Eddy-Covariance (EC) measurements is a well-known, still unresolved challenge in micrometeorology, with energy balance closure (EBC) rates typically ranging between 60% and 80%. While numerous hypotheses have been proposed to explain this imbalance, the relative contributions of neglected energy storage terms, data quality and flux processing options remain insufficiently disentangled. Using standardized ICOS and NEON datasets, we show that a significant portion of the observed energy imbalance can be attributed to overlooked or inconsistently handled energy components and turbulent flux quality control. Using data drawn from 84 sites, we show that comprehensive energy accounting-including soil heat flux, storage terms (soil, air, biomass), photosynthetic energy demand, and strict quality filtering of turbulent fluxes-improved EBC by 16% on average, with site-specific gains up to 40%. However, we also identify a persistent residual imbalance that is unlikely to be resolved through methodological refinements or additional measurements alone, pointing to fundamental physical processes that are not accounted for in the standard measurement and processing. We argue that this unresolved imbalance should be explicitly acknowledged and bounded, rather than implicitly absorbed into correction schemes, and we outline practical guidance for diagnosing and interpreting EBC in standardized flux networks. This perspective evaluates methodological advances and residual uncertainties, providing an actionable framework for the appropriate use of EC energy fluxes in carbon, water, and climate research.

Increasing climate variability is impacting the carbon cycle in unprecedented ways, demanding an understanding of the conditions leading to extreme carbon flux states. In this study, we sought to determine the strength and influential timescales of key environmental drivers governing unusually high daily gross primary productivity (GPP) and ecosystem respiration (Reco) across diverse ecosystems in the western USA. We obtained CO2 flux data from 14 AmeriFlux eddy covariance flux tower sites (11-22 years/site) across the western USA to understand the drivers of extreme CO2 sinks (extreme GPP) and sources (extreme Reco). To account for seasonality, we defined extreme sink and source states as those GPP and Reco values exceeding a site-level 95th quantile spline regression. We assembled environmental covariates computed across multiple timescales (day, week, month, year), including temperature, vapor pressure deficit, short-wave radiation, soil moisture, and precipitation, along with site-level characteristics (e.g., mean annual precipitation [MAP]). We used random forest classification models to evaluate the importance of different covariates for identifying extreme versus "nominal" fluxes. Patterns of large, but infrequent precipitation events over the preceding month were key positive drivers of the probability of both extreme sinks and sources, and the importance of such precipitation "packaging" was 16.5 times greater under water limitation (low MAP). We also found climate timescales interacted hierarchically: historical, long-term climate and conditions over the past month and/or year establish baseline ecosystem resource limitation (water or light limitation), within which short-term interactions-such as compounding events and event duration-operate to influence extreme fluxes. Preceding-month temperature and preceding-year variation in shortwave radiation had climate-specific influences, particularly in high MAT sites and in Mediterranean sites, respectively. Our results indicate the evolution of extreme carbon fluxes with climate change, including the likelihood of increased occurrence of extreme carbon flux states in response to increasing hydroclimate volatility.

ABSTRACTGross primary productivity (GPP) is the largest term in the global carbon budget but cannot be directly observed. We present a knowledge‐guided machine learning (KGML) framework that partitions eddy covariance‐measured net ecosystem exchange (NEE) into gross primary production (GPP) and ecosystem respiration (RECO) with partitioned water vapor fluxes and CO2 flux source areas from 36 U.S. National Ecological Observatory Network (NEON) towers. The KGML is guided by hard physical constraints that enforce mass balance and ‘soft’ theoretical expectations including optimal stomatal response to vapor pressure deficit (VPD) and links between GPP and transpiration (T) through stomatal function. The model achieves strong physical consistency (NEE R2 = 0.99) while capturing expected ecophysiological relationships including GPP‐T coupling (R2 = 0.58) and stomatal responses to light and VPD. Compared to conventional partitioning methods, KGML infers lower GPP and RECO estimates on average, with the largest negative biases occurring at low light levels (0–200 μmol photons m−2 s−1). These differences likely reflect a combination of mechanisms including light‐induced respiration suppression consistent with the Kok effect, stomatal‐transpiration coupling constraints, and dynamic allocation between respiration components. The flux differences vary across plant functional types (PFTs), where forested ecosystems (deciduous broadleaf, evergreen needleleaf, and mixed), savannas and grasslands show the largest negative annual GPP deviations (−10% to −18% versus nighttime partitioning), while croplands and open shrublands show moderate negative deviations (−5% to −10%). The lower GPP estimates by PFT are closer to those inferred by Keenan et al. that explicitly considers limitations on RECO from the Kok effect. We discuss implications for our understanding of ecosystem and global carbon cycle processes, as well as ways to further benefit from the full information content of eddy covariance observations by combining physics with knowledge of biological processes.