Our analyses are based on an updated version (V1) of the leaf area index dataset LAI3g; based on the methodology developed by . The data provide global year-round LAI observations at a 15 d (bi-monthly) temporal resolution and 1/12∘ spatial resolution. The record covers the period from July 1981 to December 2017. The complete time series of LAI3gV1 was generated using an artificial neural network trained on data of the overlap period of the Collection 6 Terra Moderate Resolution Imaging Spectroradiometer (MODIS) LAI dataset (2000–2017) and the latest version (third generation) of the GIMMS group Advanced Very High Resolution Radiometer (AVHRR) normalized difference vegetation index (NDVI) data (NDVI3g). The latter have been corrected for sensor degradation, inter-sensor differences, cloud cover, observational geometry effects due to satellite drift, Rayleigh scattering, and stratospheric volcanic aerosols .

The LAI3g datasets prior to 2000 were not evaluated due to a lack of required field data . After 2000, the quality of the LAI3g dataset was assessed through direct comparisons with ground measurements of LAI and indirectly with other satellite-data-based LAI products, as well as through statistical analysis with climatic variables such as temperature and precipitation variability . Various studies used the predecessor LAI3gV0 and the related dataset of fraction of absorbed photosynthetically active radiation FAPAR; and its successor LAI3gV1 .

Leaf area index is defined as the one-sided green leaf area per unit ground area in broadleaf canopies and as one-half of the green needle surface area in needleleaf canopies in both satellite observations and models (Earth system models – ESMs – and LSMs). It is expressed in units of square meters of green leaf area per square meter of ground area. Missing values in the LAI3gV1 dataset are filled using the climatology of each 16 d composite during 1982–2017. We use the annual averaged LAI of each pixel in this study.

In addition to the GIMMS LAI3g product, we analyze the MODIS LAI record as well as four other long-term global remote sensing datasets: the Global Land Surface Satellite LAI product (GLASS LAI), the Global Mapping LAI product (GLOBMAP LAI), the NDVI product from the Land Long Term Data Record (LTDR), and a new FAPAR product from the National Oceanic and Atmospheric Administration (NOAA) Climate Data Record Program.

The MODIS LAI data analyzed in this study are based on the combined Terra and Aqua MODIS LAI products (MOD15A2H and MYD15A2H) from Collection 6 C6;. These LAI datasets are provided at an 8 d temporal resolution with a 500 m sinusoidal projection covering the entire globe. The two LAI datasets are aggregated into 16 d composites by taking the mean of all valid LAI values after an additional data quality assessment is performed for more details, please see . The data are then spatially aggregated to 1/20∘ spatial resolution and cover the period from 2000 to 2019.

The GLASS LAI dataset is based on AVHRR, MODIS, and Carbon cYcle and Change in Land Observational Products from an Ensemble of Satellite (CYCLOPES) reflectances and LAI products. The full time series was generated using an artificial neural network (general regression neural network) that has been trained on the overlap period of AVHRR, MODIS, and CYCLOPES reflectances and LAI products . We use the latest version of the GLASS LAI dataset, which covers the period from 1981 to 2018 and is provided at an 8 d temporal and a 1/20∘ spatial resolution.

The GLOBMAP LAI dataset is a reconstruction of the historical AVHRR data by a quantitative fusion with MODIS data. The algorithm inverses a geometrical optical model to establish pixel-level relationships between AVHRR and MODIS LAI for the overlapping period, which are then used to reconstruct AVHRR LAI back to the initial year of the record. We use the latest version of the GLOBMAP LAI dataset, which covers the period from 1981 to 2017 and is provided at a 15 d temporal and a 1/13.75∘ spatial resolution.

The NDVI dataset of NASA's LTDR project is based on a reprocessing of long-term AVHRR reflectances applying improved preprocessing techniques and atmospheric corrections used in the generation of MODIS datasets. The preprocessing improvements include radiometric in-flight vicarious calibration for the visible and near-infrared channels and inverse navigation to relate an Earth location to each sensor instantaneous field of view . Atmospheric corrections include corrections for Rayleigh scattering, ozone, water vapor, and aerosols. We use the recently published version 5 (v5) of the LTDR NDVI dataset, which covers the period from 1981 to 2019 and is provided at a daily temporal and a 1/20∘ spatial resolution.

The FAPAR product from the National Oceanic and Atmospheric Administration (NOAA) Climate Data Record Program provided by National Centers for Environmental Information (NCEI) is based on carefully calibrated and corrected land surface reflectances from AVHRR sensors . The algorithm relies on artificial neural networks calibrated per different land cover type using the MODIS FAPAR dataset. We use the latest version of the FAPAR dataset from 1981 until 2019, which is provided at a daily temporal and a 1/20∘ spatial resolution.

We aggregate all datasets to annually averaged values and to spatially area-weighted averages for different biomes as defined by the mask (Sect. ).

The land cover product of the MODIS sensors (MCD12C1; MODIS/Terra+Aqua Combined Land Cover Type Climate Modeling Grid (CMG) Yearly Global 0.05 Deg V006, https://lpdaac.usgs.gov/products/mcd12c1v006/, last access: 31 August 2021) is the primary source underlying the land cover map used in this study (hereafter MODIS land cover). The classes from the International Geosphere–Biosphere Programme (IGBP) in the MODIS land cover product are aggregated as follows: Tropical Forests includes Evergreen Broadleaf Forest (EBF); Temperate Forests includes Deciduous Broadleaf Forest (DBF) and Mixed Forest; Boreal Forests includes Evergreen Needleleaf Forest (ENF) and Deciduous Needleleaf Forest (DNF). Savannas includes Woody Savannas and Savannas. Shrublands includes Closed Shrublands and Open Shrublands. Croplands includes Croplands and Croplands/Natural Vegetation Mosaic. The class Others includes Permanent Wetlands, Urban and Built-up Lands, Permanent Snow and Ice, and Barren. The classes Grasslands and Water Bodies remain unchanged. The MODIS land cover product provides estimates for the time period from 2001 to 2017 for each pixel. In this study we define a representative biome map based on the most frequently occurring land cover type throughout the period of 17 years.

The MODIS land cover classification does not contain the biome tundra, which is why we use in addition the land cover product GLDAS-2 Noah version 3.3 that uses a modified IGBP classification scheme providing the classes Wooded, Mixed, or Bare Ground Tundra (https://ldas.gsfc.nasa.gov/gldas/GLDASvegetation.php, last access: 31 August 2021, hereafter GLDAS land cover) . Accordingly, pixels originally of the classes Shrublands, Grasslands, Permanent Wetlands, or Barren are converted to Tundra if classified as Wooded, Mixed, or Bare Ground Tundra in the GLDAS land cover product. The classes Woody Savannas and Savannas span vast areas across the globe in the MODIS land cover product. We use the GLDAS classification for these pixels but only for regions where the MODIS and GLDAS land cover products disagree. In doing so, we obtain a more accurate global land cover classification. Table S1 describes in detail how the fusion of the MODIS and GLDAS land cover products is realized.

As a last step, we integrate the MODIS tree cover product MOD44B (MODIS Terra Vegetation Continuous Fields Yearly L3 Global 250 m SIN Grid V006, https://lpdaac.usgs.gov/products/mod44bv006/, last access: 31 August 2021) to account for the underestimation of forested area in the MODIS land cover product. Areas with tree cover exceeding 10 % are formally defined as forests . Thus, we set non-forest pixels in the MODIS land cover product above 10 % tree cover to Boreal Forests in the high latitudes 50∘ N and S. For tropical forest (25∘ S–25∘ N), we increase the threshold to 20 % tree cover to allow for a realistic areal extent of savannas. The pixels in the bands 25–50∘ N and S remain unchanged because the MODIS land cover product already realistically represents the forested area in these latitudes.

Table S1 provides a detailed overview of the conflation of the MODIS land cover product, GLDAS land cover product, and the MODIS tree cover product. The final biome map (originally resolved at 0.05∘) is regridded to the different resolutions of the AVHRR sensor and the models simulations (MPI-ESM and TRENDYv7) applying a largest-area-fraction remapping scheme.

Based on the observational LAI dataset, we define various clusters for greening or browning in most biomes: North American Tundra (NAm Tundra), Eurasian Tundra (EA Tundra), North American Boreal Forests (NAm Brl F), Eurasian Boreal Forests (EA Brl F), Temperate Forests (Tmp F), Tropical Forests (Trp F), Central African Tropical Forests (CAf Trp F), Northern African Savannas and Grasslands (NAf Sv Gl), Southern African Savannas and Grasslands (SAf Sv Gl), Cool Grasslands (Cool Gl), and Australian Shrublands (Aus Sl). Some clusters require a more detailed definition of their geographical location and extent: Southern (Northern) African Savannas and Grasslands represents these vegetation types south (north) of the Equator including Madagascar. Central African Tropical Forests represents all tropical forests in Africa. Cool Grasslands refers to grasslands above 30∘ N.

MPI-ESM1.2 is the latest version of the state-of-the-art Max Planck Institute Earth System Model, which participates in the upcoming sixth phase of the Coupled Model Intercomparison Project CMIP6;. describe thoroughly the model developments and advancements with respect to its predecessor, the CMIP5 version . Here, we use the low-resolution (LR) fully coupled carbon–climate configuration (MPI-ESM1.2-LR), which consists of the atmospheric component ECHAM6.3 with 47 vertical levels and a horizontal resolution of ±200 km grid spacing (spectral truncation at T63). The ocean dynamical model MPIOM is set up on a bi-polar grid with an approximate grid spacing of 150 km (GR1.5) and 40 vertical levels. MPI-ESM1.2-LR includes the latest versions of the land and ocean carbon cycle modules, comprising the ocean biogeochemistry model HAMOCC6 and the land surface scheme JSBACH3.2 .

As opposed to the high-resolution configuration, the LR variant of the MPI-ESM includes all the important processes relevant for longer-timescale changes in the land surface, such as a thoroughly equilibrated global carbon cycle, dynamical vegetation changes, an interactive nitrogen cycle, land-use transitions, a process-based fire model (SPITFIRE), and an interactive coupling of all sub-models. Furthermore, it is possible to run this model configuration to generate 45–85 model years per real-time day with a modern supercomputer . This opens up the possibility of conducting a larger number of realizations for each experiment.

Specifically, we used the initial CMIP6 release of MPI-ESM version 1.2.01 (mpiesm-1.2.01-release, revision number 9234). The final CMIP6 version will include further bug fixes, which are expected to only slightly influence long-term sensitivities of simulated land surface processes.

We conducted historical simulations (all forcings) and three factorial experiments (all forcings except one): (a) all historical forcings except the physiological effect of CO2 (no PE; increasing CO2 does not affect the biogeochemical processes), (b) all historical forcings except the radiative effect of CO2 (no RE; increasing CO2 does not affect climate), and (c) all historical forcings except anthropogenic forcings (no CO2). All experiments were preformed in ensemble mode (six realizations per experiment) using the latest CMIP6 forcing data (1850–2013). Individual realizations were initialized from different points in time of a prolongation run of the official MPI-ESM1.2-LR pre-industrial control simulation. In doing so, we account for the influence of climatic modes (e.g., El Niño–Southern Oscillation) as a source of uncertainty in simulating long-term changes.

The simulated time series were shifted by 4 years to maximize the overlap with the observational record of 1982–2017.

Land surface models (LSMs) or dynamic global vegetation models (DGVMs) simulate key physical and biological key processes of the land system in interaction with the atmosphere. LSMs provide a deeper insight into the mechanisms controlling terrestrial energy, hydrological, and carbon cycles, as well as the drivers of phenomena ranging from short-term anomalies to long-term changes . Here, we analyze the most recent TRENDY ensemble (version 7) comprising 13 state-of-the-art LSMs which vary in their representation of ecosystem processes. All models simulate vegetation growth and mortality, deforestation and regrowth, vegetation and soil carbon responses to increasing atmospheric CO2 levels, climate change, and natural variability . Some models simulate an explicit nitrogen cycle (allowing for potential nitrogen limitation) and account for atmospheric N deposition Table A1 in. Most LSMs include the most important components of land use and land-use changes, but they are far from representing all processes resulting from direct human land management Table A1 in. A more detailed description of the TRENDYv7 ensemble, model-specific simulation setups, and references can be found in Table A4.

We use output from five simulations: all forcings (S3), physiological effect of CO2 only (S1), radiative plus physiological effect of CO2 (S2), land-use changes only (S4), and the control run (S0; no forcings – fixed CO2 concentration of 276.59 ppm and fixed land-use map, loop of mean climate and variability from 1901–1920). The forcing data consist of observed atmospheric CO2 concentrations; observed temporal patterns of temperature, precipitation, and incoming surface radiation from the CRU JRA-55 reanalysis ; and human-induced land cover changes and management from an extension of the most recent Land-Use Harmonization (LUH2) dataset .

In this study, we only analyze output for the period 1982–2017 (matching the observational record) from models providing spatially gridded data for all five simulations. A few models provide LAI at the level of plant functional type (PFT). We calculate the average value of all LAI values on the PFT level multiplied by their land cover fraction for each grid cell. All model outputs were spatially regridded to a common resolution of 1∘ based on a first-order conservative remapping scheme .

The design of factorial simulations in TRENDYv7 and by the MPI-ESM are conceptually different. The MPI-ESM simulations were conducted using the counterfactual approach; i.e., all forcings are present except the driver of interest. TRENDYv7 provides simulations with different combinations of drivers as described above. To obtain comparability, we have to make the assumption that the absence of a specific driver has the same effect, in absolute values, as its sole presence. Thus, we process the output of the simulations S1, S2, S3, and S4 to obtain the counterfactual setup as described above for the MPI-ESM. This approach neglects possible synergy effects from simultaneously acting forcings. Also, it has to be noted that these simulations are only to some extent comparable between the two ensembles. For instance, in the MPI-ESM we can specifically determine the impact of the radiative effect of CO2, whereas TRENDYv7 uses observed atmospheric fields including changes induced by other drivers, such as non-CO2 greenhouse gases.

For certain clusters, some models show unreasonable LAI changes and/or extreme inter-annual variability. To reduce the influence of these extreme models on the overall analysis, we apply a two-step filtering method for each cluster beforehand. Models are excluded from the analysis if they exceed 3 times the inter-annual variability in observations and/or show a drastic change (of either sign) of more than 250 % between the start and end of the observational period. Further, we apply a weighting scheme based on the performance of the all-forcings run for each cluster. We calculate quartic weights based on the distance between the simulated and observational estimate. These weights are applied when calculating the multi-model average and standard deviations for the factual and counterfactual runs.

Global monthly means of atmospheric CO2 concentration are taken from the GLOBALVIEW-CO2 product (for details see 10.3334/OBSPACK/1002) provided by the National Oceanic and Atmospheric Administration Earth System Research Laboratory (NOAA ESRL).

Areas of significant change in LAI are estimated using the non-parametric Mann–Kendall test, which detects monotonic trends in time series. In this study, we set the significance level to p≤0.1. An alternative statistical test for trend detection Cox–Stuart test; yields approximately the same results. The trends are calculated either for time series on the pixel level or for area-weighted large-scale aggregated time series (e.g., biome level).

Either we define greening (browning) as a positive (negative) temporal trend, or for better comparison among models and observations as well as for a better global comparison across diverse biomes, we express these trends relative to the initial LAI level at the beginning of the observational record (average state from 1982–1984), denoted as Λ (% per decade).

The calculation of yearly net changes in leaf area balances the effects from both statistically significant browning and greening grid cells. For each cell, we multiply the estimated trends by the respective grid area. The net change is the sum of all grid cells, where areas of insignificant change are set to zero.

Models fairly accurately reproduce global patterns of vegetation greening; however, the fraction of browning is considerably underrepresented. Yet, we can only consider pixels with significant negative trends in LAI, in observations and models alike, and test models with respect to driver attribution of browning trends. Thus, the attribution of browning trends in this paper exclusively refers to browning pixels only.

Models reveal biases in comparison to observations. To obtain informative results in the attribution analysis, we process the simulations to match the mean and variance of the observational time series. Assuming additive and multiplicative biases in simulations, we apply the following corrections: 1b=σoσaf,2a=x‾o+b×x‾af, and3yi=a+b×xi, where x‾o represents the mean value and σo the standard deviation of the observational times series. x‾af and σaf are analogous to the all-forcings simulations. All simulated time series xi are scaled using Eq. (3), where i∈Ω={factual runs,counterfactual runs}. This processing step does not affect the nature of simulated trends.

The causal counterfactual approach is anchored in a formal theory of event causation developed in computer science . Recently, a framework for driver attribution of long-term trends in the context of climate change has been introduced and is increasingly gaining in popularity . Through the use of this method we can ascertain the likelihood that a certain external forcing has caused an observed change in the Earth system. More precisely, we address the question of interest in a probabilistic setting; i.e., what is the probability that a given forcing (e.g., radiative effect of CO2) has caused an observed long-term change in the system (e.g., greening of the Arctic)?

In the following, we highlight the key ideas and relevant concepts of causal theory. A detailed description and formal derivations can be found in , , and . We define the cause event (C) as “presence of a given forcing” (i.e., the factual world that occurred) and the complementary event (C‾) as “absence of a given forcing” i.e., the counterfactual world that would have existed in the absence of a given forcing;. Further, we define the effect event (E) as the occurrence of a long-term change (here, greening or browning) and the complementary event (E‾) as the non-occurrence of a long-term change (i.e., no persistent vegetation changes). In making use of numerical models, we can conduct factual runs comprising all forcings (i.e., historical simulations) as well as simulate counterfactual worlds by switching off a forcing of interest (i.e., all forcings except one). Based on an ensemble of simulations, in a multi-model and/or multi-realization setup, we derive the so-called factual (p1) and counterfactual (p0) probability, which read p1=P{E|do(C)} and p0=P{E|do(C‾)}, respectively . More precisely, p1 describes the probability of the event E in the real world where forcing C was present, whereas p0 refers to the probability of the event E in a hypothetical world where forcing C was absent. The notation do(⋅) means that an experimental intervention is applied to the system to obtain the probabilities .

The three distinct facets of causality can be established based on the probabilities p1 and p0: 4PN=max⁡1-p0p1,0,5PS=max⁡1-1-p11-p0,0, and6PNS=max⁡p1-p0,0.

PN refers to the probability of necessary causation, where the occurrence of E requires that of C but may also require other forcings. PS refers to the probability of sufficient causation, where the occurrence of C drives that of E but may not be required for E to occur. PNS describes the probability of necessary and sufficient causation, where PN and PS both hold . In other words, PNS may be considered the probability that combines necessity and sufficiency. Thus, the main goal is to establish a high PNS that reflects and communicates evidence for the existence of a causal relationship in a simple manner .

To obtain the PNS, we follow the methodology described in detail in and derive cumulative distribution functions (CDFs) for the factual and counterfactual worlds, denoted D0 and D1, respectively. Assuming a Gaussian distribution, PNS follows as 7PNS=max⁡{D1(μ1,Σ)-D0(μ0,Σ)}, where μ1 and μ0 refer to the mean response of all factual and all counterfactual runs, respectively. Σ denotes the overall uncertainty and is estimated based on all simulations, comprising factual, counterfactual, and centuries-long unforced (pre-industrial) model runs for details see. Finally, the maximum of the PNS determines the sought probability of causation . We express probabilities using the terminology and framework defined by the IPCC .

More than 3.5 decades of satellite observations (1982–2017, Sect. ) reveals that 40 % of Earth's natural vegetation shows statistically significant positive trends in LAI (Mann–Kendall test, p<0.1; Table ), concurrent with a 65 ppm increase in atmospheric CO2. However, more and more browning clusters are beginning to emerge in all continents (14 %; Table ). Analyzing earlier versions of three shorter-duration (1982–2009) LAI datasets, reported a considerably smaller browning fraction of less than 4 % and greening percentages ranging from 25 % to 50 % for all vegetation (i.e., including agriculturally dominated regions). The higher browning proportion in the extended record analyzed in this study indicates an intensification of leaf area loss in recent years. In the following, we take a closer look at different major biomes and their changes in LAI.

A global map of statistically significant trends in LAI (denoted Λ, Sect. ) for natural vegetation reveals greening (Λ>0) and browning (Λ<0) clusters across the globe (Fig. ). Temperate forests (56 % for Λ>0) and Eurasian boreal forests (53 % for Λ>0) exhibit extensive regions of increasing LAI and, thereby, contribute the largest fraction to the enhancement of leaf area on the planet (Table ). The global belt of tropical forests, on the other hand, while showing a net greening (28 % for Λ>0), also features widespread browning areas (16 % for Λ<0). In particular, the central African tropical forests contain large areas of pronounced negative trends (25 %) for Λ<0). North American boreal forests exhibit the largest fraction of browning vegetation (31 %) for Λ<0), resulting in an annual net loss of leaf area (Tables  and ). The picture of Earth's forests is generally in line with results based on other data sources. For instance, reported a net gain in global forested area, with net loss in the tropics compensated for by a net gain in the extra-tropics.

Tundra in North America is primarily greening (46 % for Λ>0 versus 7 % for Λ<0), whereas in Eurasia, browning is intensifying (35 % for Λ>0 versus 20 % for Λ<0), especially in northern Scandinavia and on the Taymyr Peninsula in northern Russia. Grasslands in cool arid climates, mainly comprising the Mongolian and Kazakh Steppe, as well as the Australian shrublands, stand out as prominent greening clusters (40 % and 49 %, respectively, for Λ>0). Although these biomes show strong positive trends, they are characterized by a low level of LAI. The African continent, which is still dominated by natural vegetation, reveals a distinct change in leaf area. A greening band of savannas and grasslands in the northern regions of sub-Saharan Africa and a greening cluster in southern Africa border the browning regions of equatorial Africa (Fig. ). Overall, the response of LAI to rising CO2 is somewhat homogeneous for some biomes (widespread browning of the tropical forests and dominant greening of the temperate forests) but divergent for others (tundra and boreal forests show a “North America–Eurasia” asymmetry, interestingly, in that they show changes of reversed sign; Fig. ).

Leaf area loss occurs primarily in densely vegetated biomes (i.e., forests), which outweighs leaf area gain in rather sparsely vegetated regions (i.e., grasslands). For instance, vigorously greening areas of circumpolar tundra result in a leaf area gain of 8.74×103 km2yr-1, which is almost outbalanced 4-fold by a leaf area loss of 34.31×103 km2yr-1 in the browning regions of the tropical forests (Table ). To assess the responses of different biomes to rising CO2 in more detail, we iteratively calculate statistically significant LAI trends for different time windows with an advancing initial year (i.e., 1982, 1983, …, 2000) but fixed final year (2017). Although the estimated trends become less robust with shorter time series, this analysis allows us to test for weakening or strengthening responses to further rising CO2. We see that the fraction of significantly browning regions is increasing over time, reaching a maximum for a time window starting in 1995. The greening fraction evolves in the opposite manner. The estimates are represented as fractions of the total area of significant change because the latter inherently decreases as a result of the Mann–Kendall test for shorter time windows. Thus, the average annual net leaf area gain of 150.51×103 km2yr-1 for the entire observational period (1982–2017) decreases with advancing initial year, approaching zero for the period 1995 to 2017, and rebounding to ∼40×103 km2yr-1 for the period 2000 to 2017 (black line in Fig.  inset). To obtain comparability between different time windows, the net leaf area gain estimates were scaled to the total area of significant change derived for 1982–2017 (unprocessed estimates for period 2000–2017 are listed in Table S2 in the Supplement). reported a global greening proportion of ∼33 % (21 % for AVHRR; Table S2) and a browning proportion of only 5 % (13 % for AVHRR; Table S2) analyzing the MODIS record including anthropogenic vegetation (2000–2017). On a global scale, LAI trends from MODIS and AVHRR agree over 61 % of the vegetated area . analyzed in detail the AVHRR and MODIS LAI trends for different climate zones, vegetation classes, and latitudinal bands and found general agreement between the two satellite-based sensors. Inconsistencies arise mainly in humid tropical regions (e.g., absence of intensive browning in central African tropical forests in the MODIS record) and partially in the northern high latitudes . In Fig.  we present a detailed comparison of different remote sensing datasets at the global scale, and we elaborate further on the discrepancies among the estimates in Sect.  for a similar analysis, also refer to.

The intensification of browning during the second half of the AVHRR observational period (2000–2017) results in a reversal of the sign in terms of net leaf area change in some biomes (e.g., tropical forests, North American boreal forests, and Eurasian tundra; Table S3 in the Supplement). Critically, the tropical forests display the sharpest transition from a substantial net gain of 24.11×103 km2yr-1 (Table ) to a comparably strong net loss of leaf area (-18.42×103 km2yr-1; Table S3). To address the temporal development of positive and negative changes in leaf area in more detail, we calculate time series of area-weighted averages of LAI (Fig. a). We find that browning of natural vegetation occurs at a considerably higher level of LAI (on average ∼1.85) than greening (on average ∼1.32). Throughout the observational period, these two time series of opposite trends converge towards an LAI of 1.6 (Fig. a). This convergence of greening and browning is evident in terms of not only their LAI level (Fig. a) but also their proportions (inset in Fig. ). The time series of anthropogenic vegetation on the other hand, aggregated for positive and negative Λ separately, are both confined to a comparably low LAI level (on average between 1 and 1.25). We next investigate the global LAI distributions of negative and positive changes and their development over time. Comparing distributions of the earlier years (1982–1984) with those of the more recent years (2015–2017) reveals that browning primarily occurs at a high (5–6) and a medium (1–2.5) level of LAI (Fig. b). Greening, however, occurs almost entirely at low levels of LAI between 0–1.5. As a consequence, the global area-weighted averages of the browning and greening regions are approaching one another (dashed versus solid vertical lines in Fig. b), as also depicted by the time series (Fig. a). Overall, these results suggest a homogenization of Earth's natural vegetation in terms of LAI texture with rising CO2. This homogenization becomes prominent when we compare the distributions of negative and positive Λ over time using a Q–Q plot (quantile–quantile; Fig. c). The relationship between the quantiles is skewed to the left at higher LAI (positive Λ on x axis, negative Λ on y axis) because browning is prevalent in high-LAI regions. Over time, the quantiles of the greening and browning distributions approach the 1–1 line (representing identical distributions), emphasizing their convergence.

Thus far, we have described the diverse long-term changes in natural vegetation across all continents and throughout the satellite era. We next investigate the underlying mechanisms driving these greening and browning trends and use the fully coupled MPI-ESM and the TRENDYv7 ensemble of observation-driven LSMs (Sect.  and ). First, we ask if these models capture the observed behavior of natural vegetation under rising CO2. The MPI-ESM reproduces the observed browning of high-LAI regions and the greening of low-LAI regions; however, the levels of LAI do not match the observations (Fig. a and b). Figure c and d compare global maps of the observed and the simulated levels of annual average LAI. Overall, the MPI-ESM is consistent with observed patterns, with the strongest spatial variations in tropical forest regions. Historical simulations of TRENDYv7 (here 13 models) also show pronounced changes in vegetation but exhibit diverse behavior among the models (results not shown for brevity). Seven LSMs reproduce observed converging trends of greening and browning, whereas the other six models show divergent trends. All TRENDYv7 models are driven with identical atmospheric forcing fields; hence, these six models most likely lack or incorrectly represent key processes of ecosystem functioning. In general, simulated greening patterns are comparable to observations , but browning, especially in the North American boreal forests, is underestimated .

Hereafter, we use changes in annual average LAI relative to the baseline period 1982–1984 (Sect. ) for better comparability between biomes, various simulations, and the observed signal. Time series of relative LAI changes from historical simulations (multi-model average for TRENDYv7 and multi-realization average for the MPI-ESM) are comparable to observations at the global scale (Fig. a and b; temporal correlations are low due to high internal variability in the signal).

We use the framework of counterfactual causal theory to attribute changes in LAI to a given driver in a probabilistic setting . Note that the causal relationships in this approach are determined based on the predictions of the models for the all-forcings (also referred to as factual) and factorial (also referred to as counterfactual) runs, and therefore the causality results may reflect biases or misrepresentations in the models. Based on the factual and counterfactual runs, we derive the probability of causation that combines the necessity and sufficiency of each factor (PNS; see Sect.  for details). When aggregated to area-weighted global averages (i.e., Earth greening trend), the observed estimate (∼ 1.08 % per decade) and the factual MPI-ESM estimate (∼ 1.14 % per decade) are comparable, whereas the multi-model average of the TRENDYv7 ensemble is an overestimate (∼ 1.79 % per decade; Fig. c). Omitting CO2-induced climate change (no radiative effect of CO2, no RE) does not have a strong effect in the MPI-ESM (∼ 1.04 % per decade); i.e., the estimate does not differ considerably from the factual run. The TRENDYv7 models indicate that the positive trend in LAI can be explained by climate change to some extent (∼ 1.21 % per decade). However, the PNS values for the radiative effect of CO2 are generally rather low (Fig. d), implying that the probability of the radiative effect of CO2 acting as a sufficient and necessary causal driver of the globally aggregated LAI trend signal is rather low. The opposite is the case when the physiological effect of CO2 (no PE) is excluded. Both model setups agree that almost no positive trend in LAI is present in a world without the CO2 fertilization effect (∼0.18 % per decade for MPI-ESM and ∼0.08 % per decade for TRENDYv7; both estimates are lower than the overall uncertainty estimate of ∼0.49 % per decade). Note that the term “overall uncertainty” here refers to a broader concept of uncertainty that includes several components, such as climate variability; inter-model variability; variability between realizations; and, if applicable, variability in observations, adapted from the approach introduced by see also Sect.  for details.

As a consequence, a high PNS can be established: the physiological effect of CO2 has in the case of the MPI-ESM likely (68 %) and in the case of TRENDYv7 very likely (91 %) caused the positive trend of global LAI in recent decades (Fig. d). This result is in line with , who reported that 70 % of global greening is attributable to CO2 fertilization. Removing both effects of CO2 results in slight negative trends, probably due to land-use practices (e.g., deforestation; Fig. c).

Natural vegetation shows a slight negative trend for the period 2000–2017 (ca. -0.4 % per decade; Fig. e). This estimate is within the range of the overall uncertainty and, thus, should be interpreted with caution. Note that the net change in leaf area is still positive when considering only significantly changing pixels (inset in Fig. ). To provide confidence in this result, we analyze three additional remote sensing datasets for LAI (MODIS-LAI, GLASS-LAI, and GLOBMAP-LAI) as well as for the normalized difference vegetation index (LTDR-NDVI) and for the fraction of absorbed photosynthetic active radiation (NCEI-FAPAR), both proxies for leaf area changes (see Sect.  for details). We calculate time series of changes relative to the average baseline value from 1982–1984 to obtain comparability between the conceptually different estimates for changes in natural vegetation (Fig. a–c). Next, we compare the trends for the entire observational period (1982–2017/2018, Fig. d) with trends of the more recent past (2000–2018, Fig. e). Three of the additional four long-term datasets show a weakening of vegetation greening for the second half of the observational period in accordance with the GIMMS LAI3g dataset (GLASS-LAI and especially LTDR-NDVI also show a reversal of the sign to a negative trend). In contrast, the dataset GLOBMAP-LAI depicts a substantial strengthening of the positive trend in LAI for the recent decades. However, the dataset also shows a suspicious jump in the year 2001, which could be an artifact related to problems in the fusion of the AVHRR and MODIS data . Furthermore, GLOBMAP-LAI generally shows the largest discrepancy among all datasets when compared to ground measurements . The shorter-term record of MODIS-LAI depicts a stable moderate greening trend for the time span of 2000–2019. Since the MODIS record cannot provide any information on the state of the vegetation in the 1980s and 1990s, we cannot assess whether MODIS would also depict a slowdown of the overall greening trend over this time period. Note that the comparability of relative trends between the long-term remote sensing products (baseline period 1982–1984) and short-term MODIS-LAI (baseline period 2000–2002) is limited. Overall, the analyses of the different remote sensing datasets support to a large extent the findings drawn from the GIMMS LAI3g dataset. For these reasons, as well as for reasons described earlier in the introduction, we focus on the GIMMS LAI3g dataset in these analyses, but we note that the single-product-centric view may imply some additional uncertainties besides the general uncertainty associated with AVHRR-based datasets.

Models reproduce the flattening of the trend and even the reversal in the sign only when the physiological effect of CO2 is excluded or with a complete absence of CO2 forcing (Fig. e). A recent study by suggests, by analyzing various observational datasets, that global CO2 fertilization has declined in recent years and highlights that land surface models are not reproducing the magnitude of this decline, mainly due to the underrepresentation of nutrient limitation. While these results are consistent with ours, we are not convinced that one can infer a decline in or saturation of the CO2 fertilization effect from these observational datasets. Rather, we argue that countervailing effects associated with the radiative effect of increasing CO2 (climatic changes, e.g., increase in atmospheric dryness and changes in water availability), as discussed below in more detail, become more pronounced and increasingly reduce vegetation productivity.

Overall, driver attribution at the global scale, as described above and also in , neglects the heterogeneity of natural vegetation and the possibility that divergent responses of different natural biomes might cancel each other out. To account for this omission, we identify 11 clusters of significant change and derive probabilities of causation for each driver across different vegetation types (Fig. ).

Forests in temperate climates exhibit a strong positive trend in LAI (∼ 2.53 % per decade), which is also seen in the models, albeit slightly overestimated (∼ 3.18 % per decade for MPI-ESM and ∼2.69 % per decade for TRENDYv7; Fig. S2 in the Supplement). The physiological effect of CO2 is the main driver with a high PNS (85 % for the MPI-ESM, 80 % for TRENDYv7; Fig. ). The trends are slightly weaker when only analyzing the second half of the observational period, but the overall result does not change. Observed warming might have additionally contributed to enhanced vegetation growth e.g., growing season extension;; however, it is not identified as an important driver by models. Most temperate forests are in developed countries and, thus, have been managed in a sustainable manner for several decades . It is conceivable that some of the positive trends in LAI could be attributed to forest management or regrowing forests ; however, this is not captured by the models (i.e., trends are negative when complete CO2 forcing is absent; Fig. S2).

The response of tropical forests to rising CO2 is more complex. The signal over the entire observational period is slightly positive (∼0.3 % per decade); however, it is within the range of overall uncertainty. Therefore, no robust driver attribution is possible (Figs.  and S3 in the Supplement). TRENDYv7 models show strongly opposing responses of LAI to the different effects of CO2: LAI decreases when the physiological effect is omitted but increases when the radiative effect is omitted. The MPI-ESM shows qualitatively the same responses but in a less pronounced way (Fig. S3). For the second half of the satellite record, the observed trend switches sign to a strong negative trend (ca. -1.4 % per decade). The models reproduce this tendency, but the multi-model average of the TRENDYv7 ensemble is still positive. During the same time period, the opposing reactions to CO2 in the factorial runs are more strongly marked (Fig. S3). These results suggest that browning caused by CO2-induced climate change is compensated for by greening affiliated with the CO2 fertilization effect at the biome level. Based on these findings, we hypothesize that the physiological effect of CO2 is strong in models and outbalances the negative effect of climate change in the tropical forests . As a consequence, the all-forcings simulations fail to reproduce the observed patterns of strengthening vegetation browning in the tropics . Because of the demonstrated limited predictive power of the models in simulating the vegetation response to climatic changes, we also rely more heavily on the published literature in the following discussion of our results.

The Amazonian tropical forests are frequently afflicted by severe droughts. During the satellite era most of these droughts were strongly modulated by the El Niño–Southern Oscillation (ENSO). For example, the droughts of 1982–1983, 1987, and 1991–1992 ; 1997 ; and 2015–2016 . The causes of the droughts in 2005 and 2010, however, were not related to ENSO but rather to a warm anomaly in sea surface temperatures in the tropical North Atlantic . Whereas the ENSO-driven droughts peak in northern hemispheric winter, thus during the wet season, the non-ENSO droughts happened during the dry season (July–September), when tropical ecosystems are more vulnerable to negative rainfall anomalies.

These intense and frequent droughts have diverse impacts on tropical ecosystems , the most prominent being an increase in wildfires and tree mortality. Recently, perennial legacy effects have been identified which lead to persistent biomass loss in the aftermath of severe droughts . For instance, some regions were still recovering from the impact of the megadrought of 2005 when the next major drought began in 2010 . found that these extreme events are also capable of disrupting hydrological mechanisms, which can lead to long-lasting changes in the structure of Amazonian ecosystems. Such droughts and associated wildfires are predicted to increase in frequency and intensity as a consequence of the ENSO-related amplification of heat waves but also due to the projected warming of the tropical North Atlantic .

In addition to these episodic disturbances, long-term changes in climate also affected the tropical forests in the Amazon region. Rising surface air temperatures have considerably increased the atmospheric water vapor pressure deficit (VPD), which has a negative effect on vegetation growth . Moreover, we find that precipitation has steadily decreased during the dry season (July–September, Figs. S4 and S5 in the Supplement) based on the latest version of the ECMWF reanalysis for the last 40 years ERA5;. This rainfall deficit and the identified lengthening of the dry season exacerbate vegetation water stress during dry seasons and favor conditions for wildfires. The slight increasing trend in wet-season precipitation (February–April) most likely cannot compensate for the water loss and its impact during the dry season (Fig. S4). Overall, the intensification of the dry season and the recurring droughts cause long-term browning trends , in line with our results of intensified browning of Amazonian forests (Fig. S5).

African tropical forests have been experiencing a long-term drying trend since the 1970s . In contrast to South America, the steady decline in rainfall is seen during both dry and wet seasons (Fig. S4). The origin of this decreasing trend in year-round rainfall is still under debate. Precipitation in equatorial Africa is expected to increase under climate change , so it is hypothesized that this trend is associated with the Atlantic Multidecadal Oscillation and/or changes in the West African Monsoon system . Long-term drying in rainforests could also be connected to the physiological effect of rising CO2. Recently, it has been demonstrated that the reduction in stomatal conductance and transpiration induces a drier, warmer, and deeper boundary layer, resulting in a decline in local rainfall . Regardless of what the causes may be, this long-term water deficiency most likely has led to the most pronounced cluster of vegetation browning in Earth's tropical forests (∼174×103 km2 net loss of leaf area in the time period of 2000–2017). No robust attribution is possible with the set of models analyzed in this study, since they fail to capture this substantial decrease in leaf area in the all-forcings runs (Fig. S6 in the Supplement). In the case of the TRENDYv7 models, this finding is particularly noteworthy as they are driven with observed precipitation changes: the spatial patterns of negative trends in LAI and dry-season precipitation in the central African tropical forests coincide to a large extent (Fig. S4).

Interestingly, the MODIS record does not exhibit this browning cluster , though it has been reported in other independent observational datasets . Also, atmospheric CO2 inversion studies have identified negative trends in carbon uptake for this region , which corroborates our results based on the LAI3g dataset.

Although we exclude direct anthropogenic land cover changes (Fig. S1, Table S1) as well as abrupt changes (Mann–Kendall test for monotonic trends, Sect. ), the LAI trend maps nevertheless show characteristic deforestation patterns, e.g., the so-called “arc of deforestation” in the Amazon region Fig. S5;. Hence, deforestation practices may explain some part of the observed gradual browning of the Amazon and African tropical forests .

In Oceania, however, deforestation appears to be a crucial driver of the observed browning in the pristine tropical forests. Significant negative trends align strongly with patterns of drastic deforestation during recent decades, described in detail by in comparison to Fig. . As opposed to central Africa and the Amazon region, climate changes are unlikely to be the key driver of browning regions in Oceania. There, precipitation, although highly variable in the dry season, appears to increase (Fig. S4) and the increase in VPD is rather minor in tropical forests .

The boreal forests show strong positive trends in Eurasia (∼2.69 % per decade for observations, ∼3.48 % per decade for MPI-ESM, and ∼2.08 % per decade for TRENDYv7), which can mostly be attributed to amplified warming of the temperature-limited northern high latitudes (PNS = 71 % for TRENDYv7; PNS = 44 % for the MPI-ESM; Fig. S7 in the Supplement). North American boreal forests exhibit a negative response to the effects of rising CO2, which has amplified over the last 2 decades (ca. -0.95 % per decade, 2000–2017). Models do not reproduce the dominant browning pattern (Fig. S8 in the Supplement), which is most likely connected to inadequate representation of disturbances . Several studies have proposed that browning has occurred as a consequence of droughts, wildfire, and insect outbreaks in the North American boreal forests . showed that the frequency of wildfires is strongly related to the dynamics of large-scale climatic patterns (Pacific Decadal Oscillation, El Niño–Southern Oscillation, and Arctic Oscillation) and, thus, cannot be tied conclusively to anthropogenic climate change. However, there is also evidence that the residing tree species suffer from drought stress induced by higher evaporative demand as the temperature rises . Moreover, models lack a representation of the asymmetry in tree species distribution between North America and Eurasia, which could explain their divergent reactions to changes in key environmental variables . Further observational evidence for the browning of North American boreal forests and the associated decline in net ecosystem productivity can also be inferred from CO2 inversion products .

Tundra ecosystems also reveal a dipole-type development between North America and Eurasia but with a reversed sign. Hence, North American tundra is strongly greening (∼4.23 % per decade for observations, ∼ 4 % per decade for MPI-ESM, and ∼4.51 % per decade for TRENDYv7), which is virtually certain (PNS = 99 % for TRENDYv7) and about likely as not (PNS = 51 % for the MPI-ESM) caused by warming (Fig. S9 in the Supplement). The trend decreases for the period 2000–2017, which could be linked to the warming hiatus in the years 1998–2012 . This is in line with the observed slowdown in tundra greening due to short-term cooling after volcanic eruptions .

Eurasian tundra shows a positive trend for the years 1982–2017 but a reversal in trend sign for the years 2000–2017 (Fig. S10 in the Supplement). Models exhibit some evidence of a strengthening browning signal but fail to capture the full extent of the emerging browning clusters seen in observations. If we only consider the grid cells that show significant browning in observations and models, we are able to conduct a robust driver attribution. According to the TRENDYv7 ensemble, the browning cluster in Eurasian tundra can very likely be attributed to CO2-induced climate change (PNS = 93 %; PNS = 47 % for the MPI-ESM). These results are in line with studies showing that tundra ecosystems are susceptible to warm spells during the growing season and to frequent droughts . The asymmetry between Eurasia and North America can be explained by changes in large-scale atmospheric circulation. Eurasia is cooling through increased summer cloud cover, whereas North America is warming through more cloudless skies . Also linkages between regional Arctic sea ice retreat, subsequent increasing ice-free waters, and regional Arctic vegetation dynamics have been postulated .

Non-forested greening clusters beyond the high northern latitudes coincide with semi-arid to arid climates . The northern sub-Saharan African savannas and grasslands have greened extensively in recent decades (∼ 4.63 % per decade; Fig. S11 in the Supplement), which is reproduced by the observation-driven TRENDYv7 models (∼4.55 % per decade) and is likely caused by climatic changes (PNS = 68 %). No robust attribution is feasible based on the MPI-ESM simulations. However, it is noteworthy that the fully coupled Earth system model points to climate change as having a negative effect in these regions, thus not reproducing the observed increase in rainfall (Fig. S11). This provides evidence for the hypothesis that African precipitation anomalies are not induced by rising CO2 but rather follow a multidecadal internal climatic mode .

The overall uncertainty in LAI changes is high in the southern African grasslands and savannas, and thus, no robust long-term change can be identified (Fig. S12 in the Supplement). It has been shown that shrublands in the more southern regions are greening in response to increased rainfall . In general, the literature suggests that greening and browning patterns in arid climates are mainly driven by precipitation anomalies . Close resemblance arises when comparing the spatial patterns of precipitation trends throughout the satellite era with significant changes in vegetation in arid environments, especially in the African continent. Decreased rainfall in arid South America coincides with strong browning clusters . This is in disagreement with the expected strong manifestation of CO2 fertilization in water-limited environments .

Australian Shrublands show a persistent positive LAI trend (∼3.84 % per decade), intermittently perturbed by climatic extreme events e.g., strong anomalous rainfall with subsequent extensive vegetation greening in 2011, Fig. S13 in the Supplement;. Models reproduce the steady greening of Australia, but no robust driver attribution is feasible due to high overall uncertainty. However, both model setups point to the physiological effect of CO2 as the dominant driver (Fig. S13). These results are congruent with recent studies that show CO2 fertilization has enhanced vegetation growth by lowering the water limitation threshold.

Grasslands in the cool arid climates exhibit persistent positive trends (∼2.03 % per decade, Fig. S14 in the Supplement). Simulated estimates are in the range of the observations (∼2.33 % per decade for MPI-ESM and ∼1.81 % per decade for TRENDYv7). Our analysis suggests that the positive response of cool arid grasslands to rising CO2 can be explained by the physiological effect of CO2 (PNS = 85 % for TRENDYv7; PNS = 88 % for the MPI-ESM). These ecosystems are dominated by C3-type plants , which are susceptible to CO2 fertilization , thus consistent with our results. In the warm arid areas, C4-type grasses dominate , which are less sensitive to the physiological effects of CO2 . As discussed above, vegetation changes there are mostly driven by precipitation anomalies, although CO2 fertilization might also contribute to a limited extent .