Spatio-temporal fields of land–atmosphere fluxes derived from data-driven models can complement simulations by process-based land surface models. While a number of strategies for empirical models with eddy-covariance flux data have been applied, a systematic intercomparison of these methods has been missing so far. In this study, we performed a cross-validation experiment for predicting carbon dioxide, latent heat, sensible heat and net radiation fluxes across different ecosystem types with 11 machine learning (ML) methods from four different classes (kernel methods, neural networks, tree methods, and regression splines). We applied two complementary setups: (1) 8-day average fluxes based on remotely sensed data and (2) daily mean fluxes based on meteorological data and a mean seasonal cycle of remotely sensed variables. The patterns of predictions from different ML and experimental setups were highly consistent. There were systematic differences in performance among the fluxes, with the following ascending order: net ecosystem exchange (R² < 0.5), ecosystem respiration (R² > 0.6), gross primary production (R²> 0.7), latent heat (R² > 0.7), sensible heat (R² > 0.7), and net radiation (R² > 0.8). The ML methods predicted the across-site variability and the mean seasonal cycle of the observed fluxes very well (R² > 0.7), while the 8-day deviations from the mean seasonal cycle were not well predicted (R² < 0.5). Fluxes were better predicted at forested and temperate climate sites than at sites in extreme climates or less represented by training data (e.g., the tropics). The evaluated large ensemble of ML-based models will be the basis of new global flux products.

Climate simulation-based scenarios are routinely used to characterize a range of plausible climate futures. Despite some recent progress on bending the emissions curve, RCP8.5, the most aggressive scenario in assumed fossil fuel use for global climate models, will continue to serve as a useful tool for quantifying physical climate risk, especially over near- to midterm policy-relevant time horizons. Not only are the emissions consistent with RCP8.5 in close agreement with historical total cumulative CO ₂ emissions (within 1%), but RCP8.5 is also the best match out to midcentury under current and stated policies with still highly plausible levels of CO ₂ emissions in 2100.

Multi‐year lags in tree drought recovery, termed ‘drought legacy effects’, are important for understanding the impacts of drought on forest ecosystems, including carbon (C) cycle feedbacks to climate change. Despite the ubiquity of lags in drought recovery, large uncertainties remain regarding the mechanistic basis of legacy effects and their importance for the C cycle. In this review, we identify the approaches used to study legacy effects, from tree rings to whole forests. We then discuss key knowledge gaps pertaining to the causes of legacy effects, and how the various mechanisms that may contribute these lags in drought recovery could have contrasting implications for the C cycle. Furthermore, we conduct a novel data synthesis and find that legacy effects differ drastically in both size and length across the US depending on if they are identified in tree rings versus gross primary productivity. Finally, we highlight promising approaches for future research to improve our capacity to model legacy effects and predict their impact on forest health. We emphasise that a holistic view of legacy effects – from tissues to whole forests – will advance our understanding of legacy effects and stimulate efforts to investigate drought recovery via experimental, observational and modelling approaches.

Soil is the largest organic carbon (C) pool of terrestrial ecosystems, and C loss from soil accounts for a large proportion of land‐atmosphere C exchange. Therefore, a small change in soil organic C (SOC) can affect atmospheric carbon dioxide (CO₂) concentration and climate change. In the past decades, a wide variety of studies have been conducted to quantify global SOC stocks and soil C exchange with the atmosphere through site measurements, inventories, and empirical/process‐based modeling. However, these estimates are highly uncertain, and identifying major driving forces controlling soil C dynamics remains a key research challenge. This study has compiled century‐long (1901–2010) estimates of SOC storage and heterotrophic respiration (Rh) from 10 terrestrial biosphere models (TBMs) in the Multi‐scale Synthesis and Terrestrial Model Intercomparison Project and two observation‐based data sets. The 10 TBM ensemble shows that global SOC estimate ranges from 425 to 2111 Pg C (1 Pg = 10¹⁵ g) with a median value of 1158 Pg C in 2010. The models estimate a broad range of Rh from 35 to 69 Pg C yr⁻¹ with a median value of 51 Pg C yr⁻¹ during 2001–2010. The largest uncertainty in SOC stocks exists in the 40–65°N latitude whereas the largest cross‐model divergence in Rh are in the tropics. The modeled SOC change during 1901–2010 ranges from −70 Pg C to 86 Pg C, but in some models the SOC change has a different sign from the change of total C stock, implying very different contribution of vegetation and soil pools in determining the terrestrial C budget among models. The model ensemble‐estimated mean residence time of SOC shows a reduction of 3.4 years over the past century, which accelerate C cycling through the land biosphere. All the models agreed that climate and land use changes decreased SOC stocks, while elevated atmospheric CO₂ and nitrogen deposition over intact ecosystems increased SOC stocks—even though the responses varied significantly among models. Model representations of temperature and moisture sensitivity, nutrient limitation, and land use partially explain the divergent estimates of global SOC stocks and soil C fluxes in this study. In addition, a major source of systematic error in model estimations relates to nonmodeled SOC storage in wetlands and peatlands, as well as to old C storage in deep soil layers.