Extreme climatic events such as heat or drought are key features of Earth's climatic variability and occur on a wide range of timescales . Extreme climatic events directly propagate into the terrestrial biosphere, thus affecting ecosystem functioning and land surface properties (e.g., soil moisture), which in turn triggers ecosystem–atmosphere feedback loops e.g.,. For example, drought in conjunction with severe heat reversed several years of ecosystem carbon sequestration in Europe in 2003 , and strong land–atmosphere feedbacks exacerbated the event while it was occurring .

However, ecosystem impacts of extreme climatic events are often nonlinear and interact with concurrent climatic conditions. Additionally, potential impacts can cancel each other out depending on the type and state of the ecosystem and the magnitude of the climatic event. For instance, extremely warm conditions at the beginning of the growing season during spring 2012 in the contiguous US increased ecosystem carbon uptake, which subsequently compensated for ecosystem carbon losses later during the same year's summer heat and drought. Nonetheless, warm spring conditions and corresponding earlier vegetation activity likely also contributed to exacerbating drought impacts through reduced initial soil moisture at the onset of summer drought .

Because extreme climate events have been changing in recent decades, with, for example, a general increase in the amount of warm days and the duration of warm spells and the opposite trend for cold days and spells , and are projected to continue to change , an understanding of their impacts on ecosystems is crucial. Ideally, this understanding would cover ecological processes that operate on both local and global scales. However, due to nonlinear and interacting ecosystem effects of climate extremes, differences in ecosystem responses across various growing season stages and various ways in which different ecosystem types mediate climatic extremes e.g.,, it currently remains unclear whether a global perspective on ecosystem responses to climate extremes can emerge from local-scale observations alone. Moreover, understanding ecosystem responses to climate extremes is crucial in the context of potentially increasing intensities or frequencies of climatic extremes that could lead to a positive carbon-cycle–climate feedback via a reduction in the land carbon sink.

Gross primary production (GPP), which is carboxylation rate (i.e., true photosynthesis) minus photorespiration , is strongly impacted by temperature and water stress . Besides its other main environmental drivers (radiation, humidity (i.e., vapor pressure deficit, VPD) and CO2 concentration, cf. ), temperature directly influences photosynthesis by affecting the kinetics of its two main chemical processes, namely the maximum rates of carboxylation (i.e., Vc, max; ) and electron transport (i.e., Jmax⁡; e.g., ). Both rates initially increase with rising temperature but decrease above a certain optimum temperature . Leaf (i.e., light) respiration similarly increases with temperature , which additionally reduces GPP. As a result, extremely high temperatures can severely reduce photosynthesis (and, hence, GPP) .

Soil water stress impacts photosynthesis see, e.g.,for a review by causing either ecophysiological or structural changes to the plant . For instance, a physiological reduction in photosynthesis can be caused by reductions in enzymatic activity or a reduction in mesophyll and stomatal conductance e.g.,. Structural changes reducing photosynthesis include reductions in leaf area and specific leaf area or changes in leaf geometry or orientation . Via increased tree mortality, droughts can also severely impact ecosystem-level photosynthesis long after the drought event itself e.g.,.

All these responses are highly species dependent, highlighting the need for global cross-site analyses. For example, forest species generally close their stomata much earlier compared to species from grassland or savannah ecosystems, which often keep transpiring until their water storage is depleted . In addition, anisohydric plants in general have no control over their stomata . A soil-dependent factor increasing the ecosystem's drought resilience is the rooting depth and the general availability of fine roots .

In addition, interactions between heat and drought may affect GPP. For example, drought-induced closing of the stomata and the subsequent reduction in evaporative cooling can further increase heat stress when water stress co-occurs with a high-temperature anomaly . Conversely, high-temperature impacts can be alleviated by evaporative cooling as long as enough water for transpiration is available .

Ecosystem respiration (Reco) is the sum of autotrophic respiration and the CO2 emissions arising from the heterotrophic decomposition of organic matter in soil e.g.,. Like GPP, it is affected by changing soil (and, hence, ambient air) temperatures . Rising temperatures directly increase the kinetics of microbial decomposition, root respiration and the diffusion of enzymes. Hence, soil respiration is commonly modeled as an exponential function of temperature using the van't Hoff type Q10 model or other functions of a similar shape e.g.,. Even though enzyme activity generally decreases above a certain temperature optimum , such high temperatures rarely occur in extratropical soils , so high temperatures alone are rarely an inhibiting stressor for soil respiration.

In addition, the activity of soil microorganisms depends on soil moisture . Drought conditions strongly reduce soil respiration because the microbial activity causing soil respiration is dependent on the presence of water films for substrate diffusion and exoenzyme activity . In addition, low soil water status may even cause microbial dormancy and/or death . Indirectly, drought reduces microbial activity through different processes like the alteration of soil nutrient retention and availability or changes in microbial community structure . Finally, interactions between the response to temperature and water status, such as changing temperature dependency due to changing soil water status , further complicate the picture.

As described above, both heat and drought affect GPP and Reco in a similar fashion, although the amplitude and onset of this impact may differ. Hence, one important, partly unanswered question is the impact of climate extremes on the balance of these two fluxes: the net ecosystem production (NEP). Models tend to agree that drought affects GPP more strongly than Reco, but their spread is large and predictions for the C balance are uncertain . In addition, observational studies on large drought and heat events like the 2003 European heat wave or the 2000–2004 drought in North America have shown, for example, that drought may cause a much stronger reduction in GPP compared to Reco, leading to a reduction in the ecosystem's CO2 uptake.

However, it is important to understand that the tight coupling between GPP and Reco in most ecosystems complicates systematic assessments across sites. For example, heterotrophic respiration is not only a function of the environment but is also strongly driven by the availability of recently assimilated carbon . Hence, a reduction in photosynthesis may cause a lagged reduction in soil respiration in the absence of a large labile carbon stock.

The majority of studies so far focus on individual sites and predefined extreme events seefor a review and only a few have focused on comparisons of extreme-event impacts globally across sites and/or across broader regions and different ecosystems . The La Thuile dataset collected by FLUXNET consists of 252 sites of eddy covariance flux observations in a standardized way . These data provide a basis for a robust assessment of the impacts of climatic extremes on ecosystem CO2 fluxes. The opportunities arising from this trove of observations are exemplified in Fig. . The figure recalculates the impacts of the 2003 heat wave on land fluxes as estimated by using more reference years based on the data available. The general findings of , who showed a strong reduction in C uptake, are confirmed, but we now estimate a lower reduction in CO2 uptake when considering more reference years, which is consistent with , who found a similar pattern using models. Consequently, the length of today's data records and in particular the tremendous work of the numerous networks and initiatives (see Acknowledgements) who collect these data and provide them to the scientific community allow us to update previous quantifications of the CO2 flux impacts of climate extremes.

The objectives of this study are threefold: first, we want to exploit the available FLUXNET data to systematically assess if extreme events corroborate our assumptions about ecosystem behavior and to empirically describe the spectrum of extreme responses across the globe. To do so, we extract information about the occurrence of an extreme climatic event directly from the observed data, not by first assuming the occurrence of an extreme event (i.e., by identifying an extreme response of the observed ecosystem). Second, our goal is to develop an extreme-event detection framework with a focus not only on the extremeness of the climate forcing but which simultaneously takes into account the resulting extremeness of the ecosystem's response or lack thereof . Finally, we aim to bridge the gap between local site-level studies and global assessments,which most often are based on models e.g., or upscaling studies by providing some helpful benchmarks for the models and their underlying assumptions .

Our study can be outlined as a three-step process (Fig. ): first, we use consistently downscaled climate data (Sect. ) to detect climatic extreme events (Sect. ) during the growing season in a set of ecosystems. Second, we compare CO2 fluxes (Sect. ) during these extreme events with reference fluxes during comparable, non-extreme periods to quantify the impact of each extreme event (Sect. ). Third, we use site-specific information like plant functional type (PFT) or ecoclimatic zone (Geiger–Köppen climate classes) as well as climate extreme characteristics (including type and duration) to systematically assess potential causes of differences between extreme-event responses in the different ecosystems.

Measurements of CO2 flux and climate parameters collected through a network of measurement sites were used in this study. CO2 fluxes were measured using the eddy covariance technique e.g.,. The measured net carbon flux (i.e., net ecosystem exchange, NEE) was partitioned into GPP and Reco at each site. The empirical relationships used by this partitioning scheme assume similar ecophysiological conditions for any given time step (e.g., for one of the extreme events detected here) and a short reference period is used to fit these empirical functions. Environmental stress, however, could also directly impact the processes governing these empirical relationships and hence the validity of this assumption. To assess whether this could bias our analysis, we also performed all of our calculations using midday NEE as a rough estimate for GPP and averaged nighttime NEE as a proxy for Reco .

Throughout the rest of the paper, we refer to NEP instead of NEE (i.e., NEP = GPP - Reco=(-1)⋅NEE) because NEP is centered on the ecosystem (i.e., positive NEP equals CO2 uptake) and facilitates a more intuitive interpretation together with the component fluxes GPP and Reco.

Eddy covariance measurements are continuously taken at various sites across the globe by individual research teams and are collected and consistently processed by the FLUXNET network . For this analysis, we used the FLUXNET La Thuile dataset, which consists of a total of 252 sites. We used additional data from the European eddy fluxes database cluster (http://www.europe-fluxdata.eu/) for site years collected since the creation of the La Thuile dataset in 2007. Both networks consistently filter the submitted data for potential outliers. The half-hourly measurements supplied by the data providers are consistently gap-filled via marginal distribution sampling (MDS) , i.e., by filling missing values with measurements taken under similar meteorological conditions, and aggregated to daily mean values. For this analysis we used only daily aggregates and excluded data for days with less than 85 % original measurements or high confidence gap-filled data.

To be able to compare flux measurements during a potential extreme event with fluxes during non-extreme conditions during comparable stages of the phenological cycle in other years, we selected 102 sites with time series longer than 3 years. In addition, we removed four sites where the correlation between downscaled climate data and measured site meteorology was too low (R2<0.6) (Sect. ) and four sites where water availability (Sect. ) could not be calculated due to missing data. Finally, we excluded 25 managed and disturbed sites where disturbances such as fire and thinning would have resulted in biases in the calculations of the non-extreme reference data in years before or after the disturbance. This resulted in a subset of 69 sites (Table ) out of the original 252 La Thuile sites, with a total of 433 site years of data (i.e., years with data available for more than 75 % of all days). These sites span 11 PFTs including grasslands, wetlands and forest type ecosystems (Table ) and all major Geiger–Köppen climate zones (Table ) (i.e., first category zones A–E), as well as half of the 24 Geiger–Köppen subzones (i.e., the secondary categories).

To be able to identify extreme events over sufficiently long and consistent time periods for all sites, compared to the much shorter time periods where actual measurements were available, we used downscaled climate data for the extreme-event detection. We used daily air temperature and, for the calculation of the water availability (see below), global radiation and precipitation from ERA-Interim data at a 0.5∘ spatial resolution (i.e., the area of 1 pixel ≤ (55 km)2).

Multiple linear regression models of the nearest nine grid boxes (i.e., the grid box with the tower and its direct neighbor pixels) were fitted to FLUXNET site-level meteorology measurements. The resulting models were used to predict site-level values for a time period of 30 years between 1983 and 2012. The resulting time series were then used to detect climate extremes. The correlation between downscaled and site-level data for air temperature was R2>0.9 for nearly 90 % of the sites. Sites with R2 < 0.6 (≈ 5 % of the sites, mainly tropical evergreen broad-leaved forests) were removed from the analysis due to the low quality of the downscaling.

To consistently quantify the amount of soil water available to the plant, a water availability index (WAI) was calculated. This index was based on the water balance between precipitation and evapotranspiration and was calculated as a simple two-layer bucket model see Supplement 3 infor detailed equations, etc.. At each time step, the soil is recharged with water by precipitation up to a maximum value defined by the storage capacity (125 mm). Losses of water by evapotranspiration are taken as the minimum of either potential evapotranspiration or supply-limited evapotranspiration. Potential evapotranspiration is estimated based on from net radiation (also taken from the reanalysis data) using a Priestley–Taylor coefficient of 1.26. Potential evapotranspiration is then finally scaled with smoothed fAPAR (fraction of absorbed photosynthetically active radiation) (from MODIS, Moderate-resolution Imaging Spectroradiometer). Supply-limited evapotranspiration is calculated following and is simply defined as a fraction (i.e., 0.05, the median of the values determined by ) of current WAI. Assuming that both water recharge (i.e., precipitation) and water loss (i.e., evapotranspiration) operate from top to bottom, WAI was computed for a simple two-layer model, where the storage capacity of the upper layer was set to 25 mm and that of the lower layer to 100 mm. Only WAI of the lower layer was used in the subsequent analysis and scaled to 0–1 (by dividing by the maximum capacity of 100).

The WAI does not account for local soil or vegetation specific properties such as soil texture or rooting depth. such that the WAI may be interpreted as a “climatological water availability metric”. The results are sensitive to the fixed value of storage capacity, which influences the timing and magnitude of extreme drought events. For example, a larger (smaller) storage capacity value would tend to result in a later (earlier) extreme-drought detection. We are confident, however, that these changes would not strongly bias the qualitative and global patterns of the flux impacts investigated in this analysis.

Extreme events were defined and detected in the following stepwise procedure:

identification of single extreme data points (i.e., days) crossing the upper and lower 5th percentile threshold

A percentile-based approach was used to define the upper and lower 5th percentile of the original distribution as extreme (subscript max and min; cf. Table 1). Due to the strong seasonal cycles of air temperature and WAI at most outer tropical sites, this definition resulted in extreme events mainly being detected in summer and winter and represents a means for capturing extreme conditions beyond an actual value with direct physiological meaning.

However, from an ecosystem physiological perspective, an extreme climatic event can also occur outside the maximum or minimum period of the year (e.g., during spring or fall for temperature). To detect such extreme events, air temperature time series were deseasonalized by subtracting a mean annual cycle (MAC) to yield anomalies. The MAC was computed as the daily average of all 30 years and smoothed with a 2-week moving average. The upper (and lower) 5th percentiles of these anomalies were defined as extreme (subscript anom, max and anom, min) (Table  for all extreme-event notations used). Such anomaly extremes were only detected for air temperature because seasonally varying sensitivity to water availability is not expected.

After the identification of single extreme time steps (i.e., days), contiguous extreme time steps were concatenated into extreme events. Additionally, two successive but not contiguous extreme events were subsequently treated as one single long extreme event if the non-extreme period between them was shorter than 20 % of the combined length of the two extreme events together. This prevented short-term fluctuations in temperature (WAI did not usually fluctuate so quickly) below the extreme threshold during one long period of high temperature from separating this period into smaller extreme events and allowed for a more realistic assessment of the extreme-event duration (see below).

To differentiate between the effects of univariate extremes and the possibility of different impacts of simultaneous extremes of heat and drought, the following types of extreme events were differentiated: (1) single variable extreme events irrespective of the possible extremeness of other variables (denoted T/WAImin⁡), (2) single variable extreme events without other variables being extreme (denoted, for example, as Tmax⁡,s or WAImin⁡,s) and (3) concurrent extremes , i.e., coupled extreme events with multiple variables being extreme (Tmax⁡+WAImin⁡) (Table  for an overview).

Finally, all extreme events were described by characteristics such as duration and type (see above) to identify which of these factors influence the type and magnitude of possible impacts. At this first stage we did not consider several other ecosystem specific important factors which influence the ecosystem's response to climatic extremes such as site history and detailed species composition e.g.,. Such an analysis should be generally possible at future stages (Sect. ); however, the relevant information first has to be gathered across all sites in a standardized and comparable way.

To identify those events that actually have a physiological impact among all the detected climatic extreme events, a consistent quantification of the actual impact on the ecosystem was required.

To do so, differences between the mean of the fluxes during the extreme event and comparable reference periods were computed see, e.g.,for a similar approach. These reference periods were defined to be non-extreme, identical days of the year (DOY) from all other available years. For the reference period, the mean was computed from a moving-average smoothed time series (i.e., 14-day moving-average filtering computing the median) to minimize the influence of stochastic fluctuations. During the actual extreme event, however, non-smoothed data were used to compute these means.

Δf=f‾extr-f‾ref=1n∑k=ii+n-1fk-1ny∑k=jfk

Here, f denotes the respective CO2 flux (NEP, GPP or Reco), i denotes the first day of one particular extreme event of length n, j denotes the identical (and not extreme) days of the year (DOY) in all other years, and y is the number of reference years.

As the amplitudes of Reco and GPP differ significantly between highly productive and less productive ecosystems, all analyses were done for original (Eq. ) and for z-transformed time series: Δz=z‾extr-z‾ref=1n∑k=ii+n-1zk-1ny∑k=jzk, with zk=fk-f‾σ^(f) for all k.

Even though extreme events outside the growing season, such as extreme-frost periods in winter, can have impacts on the ecosystem's carbon fluxes, such impacts would be lagged in many cases (i.e., visible during the following growing season). Because only instantaneous responses were investigated with our framework, it was necessary to exclude such extreme events from the analysis. To identify the growing season, a spline function was used to smooth the time series of GPP. In the first step, all smoothed values above the 25th percentile were considered to be the growing season. Subsequently, in each year these periods were extended at the beginning and end of the detected period by identifying the first day when the smoothed series dropped below the 5th percentile.

High-temperature extremes without particularly low water availability (i.e., Tmax⁡, Tanom, max, Tmax, s and Tanom, max, s) had only small or virtually zero impacts on GPP (Fig. ), which is consistent with earlier findings (e.g., for the European heat wave 2003; ). This averaged effect can be partly explained by the specific response of different ecosystem types (see Sect. ). Heat extremes in general tended to have no or only a small negative impact on observed rates of GPP in most cases. Even though GPP has been shown to have clear temperature optima and decreases at high temperatures due to enzyme inhibition , such conditions (i.e., temperatures well above 30 ∘C) are experienced only rarely in the (mostly temperate and Mediterranean) sites investigated. Other studies also confirm the small impact of heat alone on GPP . Only for very long and pronounced extreme events was a clear negative impact on GPP observed (discussed in detail in Sect. ).

In our analysis, water scarcity events (i.e., WAImin⁡ and WAImin, s) in general showed a reduction in GPP and Reco, which, due to compensation of these component fluxes, led to no discernible changes in NEP on average over the considered FLUXNET sites (Fig. ). In contrast, events in which low water availability coincided with heat led to a very strong reduction in GPP but a lesser reduction in Reco and as a consequence to the strongest reduction in carbon uptake (see Sect.  for a more detailed discussion). Such a strong effect of droughts (compared to high temperatures alone) on GPP and the generally decreasing effect of drought on GPP is consistent with other studies e.g., where water stress directly forces plants to close their stomata to limit transpiration, reducing photosynthesis. Similarly, found that water availability is a much bigger control on the interannual variability in GPP (IAV, which is controlled to a large degree by extreme events) compared to a smaller temperature control on a global level.

In contrast to the small response of GPP to heat, however, Reco generally increased during most high-temperature extreme events (Fig. ). As a consequence, NEP decreased, which represents reduced carbon uptake of the ecosystem. Rising temperatures in general lead to an increase in the microbial degradation of biomass , which explains rising Reco rates during short periods of high temperatures as observed in other studies . An additional factor could be higher radiation inputs, which result in increased photodegradation in relatively open non-forest ecosystems.

Compared to temperature, soil respiration as the main component of Reco is regulated much more strongly by soil water availability . Droughts in general in our study led to a similar reduction in Reco compared to GPP. The reason for this could be the inhibition of soil microbial processes due to moisture limitation. Additionally, a decrease in GPP also results in a coupling of the two fluxes and, hence, also leads to a reduction in Reco . The compensating effect of drought-induced reductions in both GPP and Reco resulted in small or negligible changes in NEP, which also has been demonstrated at local e.g., and global levels.

In contrast to the single-factor extreme events discussed above, concurrent heat and drought extremes (Tmax⁡+WAImin⁡) led to a much stronger reduction in GPP in most cases. By contrast, Reco was not so strongly (or not at all) reduced. This resulted in the strongest NEP (i.e., C sink) reduction in any extreme event (Fig. ).

Several studies have found a lower drought sensitivity of Reco compared to GPP whereas we observed comparable or even slightly greater reductions during drought-only (WAImin, s) extremes on the global scale. The strong drought extremes investigated in these studies, however, usually coincided with heat extremes and are hence more comparable to our concurrent heat and drought extremes (Tmax⁡+WAImin, s) where we also see a nearly negligible mean effect on Reco (compared to GPP).

NEP is the sum of the opposing fluxes of GPP and Reco, and hence, the direction and amplitude of its change is always determined by the sum of the extreme-event impacts on the gross fluxes. For heat extremes, the general increase in Reco adds to slight decreases (or no change) in GPP, leading to a generally reduced rate of net carbon uptake. For only drought (and no heat) extremes, the reductions in both GPP and Reco seem to roughly cancel each other out, leading to no strong effects on NEP (again, as a FLUXNET average). However, during the concurrent heat and drought extremes, Reco is less strongly reduced than GPP (and also compared to only drought extremes), leading to strong reductions in net carbon uptake compared to non-extreme conditions. Part of this effect can be explained by the compensating and opposite effects of heat and drought on Reco .

While our analysis confirms a crucial impact of dryness on the individual carbon fluxes GPP and Reco, it also shows that drought extreme events in which dryness coincides with Tmax⁡ extremes have a disproportionately large negative impact on the net carbon balance (i.e., compare also Fig.  lowest panels on the right side), which is consistent with model results . The combined effect of dryness and heat might be interpreted in a process-oriented way in that dryness acts primarily to reduce GPP, while heat increases Reco, thus both leading to a severe reduction in net ecosystem carbon sequestration. Hence, we conclude that an assessment of combinations of extreme climate variables, in particular heat and drought , is indeed crucial for understanding ecosystem impacts .

Extreme-event duration is an important factor that influences ecosystem impacts . In our study, with increasing duration of the extreme climatic event, the impact on GPP generally emerged more clearly (Fig. ). For Tmax⁡ extreme events there was a threshold at a duration of > 27 days at which GPP strongly decreased by approximately 1–2σ. This effect was also visible for Reco, albeit less pronounced. With increasing duration, the response in Reco was reversed: for short heat events (i.e., with a duration of less than 18 days), Reco increased with respect to normal conditions by up to 2σ, whereas for events that last longer than a month, the response of Reco was predominantly negative.

During concurrent Tmax⁡ and WAImin⁡ extremes, GPP and Reco were reduced only for extreme events longer than 18 days. For all other extreme types and for the other fluxes, no clear relationship between extreme length and impact was observed (Fig. ).

The impact of extreme climate events on GPP ranged from a neutral impact (heat lasting less than 1 week, not coinciding with dryness) to severe impacts (if temperature extremes persisted for more than 1 month). The reversal from positive impacts for short durations to negative impacts for long extreme events in the case of Reco might be interpreted as an initial pulse of microbial activity in the soil, which is reduced after some time when the supply limitation of respiration (i.e., GPP effects) kicks in. Hence, these findings highlight that event duration is a critical parameter that might qualitatively affect the directionality of the response and thus lead to highly nonlinear ecosystem responses. These duration effects are often not explicitly considered in the analysis of climate extreme effects on ecosystems e.g.,. Future research should address the question of whether such nontrivial patterns can be reproduced in model simulations.

Most climate extreme indices for temperature consider only relatively short temperature extremes, such as monthly maximum values of temperature or the count or percentage of days that exceed an absolute or relative threshold. Furthermore, currently used climate extreme indices are based on univariate metrics . Our empirical analysis shows that ecosystem impacts of climate extremes critically depend on the duration of an extreme event and the coincidence of several climate variables. Hence, most critical/negative ecosystem impacts are seen on timescales of 2–3 weeks to a few months see also and when heat coincides with dryness.

Compared to the differences between the means of the impacts discussed above, the spread of the impacts is rather large (Fig. ). One reason for this is that differences between ecosystems are hidden by the global (i.e., averaged) focus investigated and discussed above. Figure  shows the extreme-event impacts for the different extreme-event types separated for the different PFTs, Fig.  shows this for different Geiger-Köppen climate classes, and Fig.  shows this for the combination of the two factors.

The clearest differences between impacts for ecosystems in particular climate zones appeared in the open shrublands (OSH) of the polar climate zone (ET) (Fig. ). Both GPP and Reco were increased by more than ∼ 1σ during Tmax⁡ extremes (Fig. ). A stronger increase in GPP led to a slight overall increase in NEP (i.e., a C gain). No drought extremes occurred during the investigated growing seasons in these ecosystems.

A similar but smaller (∼ 0.3σ) GPP increase during heat extremes occurred in the cold arid (BSk), mostly GRA and OSH (Fig. ), ecosystems (Fig. ). (For an explanation of ecosystem and climate class abbreviations, please see Tables A1 and A2 in Appendix A.) Here, however, Reco was increased by a similar magnitude, resulting in only a slight increase in NEP. Again, drought extremes did not occur during the investigated growing seasons. In contrast, in the warm arid (i.e., BSh climate zone) and exclusively ENF ecosystems, GPP experienced moderate decreases during the Tmax⁡ extremes. In combination with an increase in Reco comparable to the impact in the warm steppe climates (BSh), this resulted in a general NEP decrease.

The ecosystems in the mostly North American and continental European and Asian “snow” climate zones (Dfa, Dfb, Dfc) experienced mean increases in Reco during heat extremes of around 0.5σ (Fig. ). GPP, however, showed almost no changes averaged over the whole Dfc (i.e., cold summer) climate zone during heat extremes but with this being the result of a reduction in its open shrublands (OSH) and opposing increases in the wetlands (WET) of this climate zone (Fig. ). In hot and warm summer ecosystems of this climate zone (Dfa and Dfb), GPP was slightly increased. As a consequence, this resulted in a relatively strong NEP decrease in Dfc ecosystems but only in a moderate decrease in Dfa and Dfb climates. For drought extremes, however, only the summer hot Dfa cropland (CRO) ecosystems showed reductions in Reco and, to a lesser extent, in GPP during drought extremes.

Temperate and summer hot and dry (Csa, mainly Mediterranean) ecosystems experienced the strongest GPP reductions (0.3σ), with particularly strong impacts in the forest and savannah ecosystem compared to grasslands and open shrublands (Fig. ), during heat extremes, whereas Reco in general was not impacted, resulting in an NEP decrease during heat extremes (Fig. ). During drought periods, these Csa sites were among the ecosystems with the strongest reductions in Reco for all forest and savannah ecosystems but not the open shrublands which experienced increases in respiration (Fig. ) and to a lesser extent in GPP. In contrast, temperate summer dry ecosystems with only warm summers (Csb) did not experience such strong reductions in GPP and even increases in Reco during heat extremes and a smaller decrease in GPP during drought extremes (compared to Csa). Most other ecosystems in humid temperate climate zones (Cfa and Cfb) showed impacts consistent with the general patterns (i.e slight GPP and stronger Reco increases during heat extremes, a reduction in both fluxes during drought and a smaller reduction in Reco during concurrent heat and drought) which is in line with other research i.e.,.

The few equatorial winter dry (Aw) woody savanna ecosystems under investigation experienced slight reductions in GPP during heat extremes. They were one of the few climate zones where Reco was slightly reduced during Tmax⁡ extremes. Due to the few sites and short time series, drought extremes did not occur here often enough to reliably investigate their impacts.

Whether temperature or water availability governs an ecosystem's response to extreme events is mainly dependent on whether the ecosystem is located in a temperature- or water-limited environment . This explains the strong increases in both GPP and Reco during high-temperature extremes in the open shrublands of the temperature-limited polar ET climate zone compared to all other climatic zones (Fig. ). Similar results have been found by . In addition, the detected extreme events are at relatively low temperatures below 20 ∘C, which are probably well below a possible heat stress for the affected plants and still in the range where increasing temperatures increase both GPP rates and the decomposition processes which govern Reco. An additional factor could have been the increased sunlight during the extreme events (which may have caused the heat extreme in the first place) in these energy-limited regions.

Temperature extremes at sites in the arid steppe climates (BSh and BSk) have comparatively small impacts, probably because most heat extremes occur during dry periods with very low biological activity (Fig. ). For one BSk site, however, the period of high temperatures and high fluxes coincides with GPP increases during these extreme events, causing the general mean GPP increase in this climate class compared to the BSh sites.

For the one available tropical Aw site, very small seasonal temperature changes between ≈ 30 and 32 ∘C are observed (Fig. ). As a result, our extreme-detection framework detects all extreme events during the slightly hotter rainy season at the beginning and end of the year (in the Southern Hemisphere). Still, such small temperature differences are unlikely to cause visible physiological impacts, which is demonstrated by the nearly nonexistent mean impact on GPP, Reco and NEP in this climate zone. However, station density in tropical ecosystems is very low compared to temperate Northern European or North American sites so this may also be a consequence of the small amount of extreme events detected.

The approach presented in this paper is based on a global, empirical characterization of the impacts of climate extremes on ecosystem–atmosphere carbon fluxes, which has several advantages but also limitations for addressing global ecological questions. Classical extreme-event research has often focused on events where the response was already known a priori to be strong and has possibly neglected several comparable climatic periods with similar conditions but with smaller or even opposite impacts. In contrast, all periods are included in our analysis because we did not select our extreme events a priori. Our results show that comparable extreme events can lead to contrasting impacts, which depend on ecosystem type or extreme-event timing.

In addition, this research is one of the few global and cross-site/ecosystem investigations of extreme climate impacts on (measured) CO2 fluxes. We try to extend the sometimes limiting (but still valuable) focus on particular sites and compare such responses globally. This allows for a holistic picture with which such local site observations can be compared.

Our global results highlight the importance of drought events for the ecosystem carbon cycle. Hence, a reliable estimate of water availability is crucial for the identification of climatic extreme events. As soil water measurements at FLUXNET sites differ strongly between sites in quality, depth and duration, we chose to use the modeled WAI for better between-site comparability and consistency e.g.,. Even though we see responses of the fluxes to decreasing WAI, the detailed investigation of individual drought events (e.g., the 2003 heat wave: Fig. ) highlighted the possible sudden decrease in the fluxes to gradual changes in WAI, emphasizing the need for a reliable estimate of WAI. At this stage, WAI was not optimized for the individual sites and represents a purely hydrometeorological variable rather than a direct measure of ecosystem-specific water stress.

We applied the 95th (or 5th) percentile threshold to define extreme events throughout our study to allow for a comparable extreme definition for all ecosystems. Importantly, this approach has as few a priori assumptions as possible (compared to identifying extreme events via somewhat subjective expert knowledge or by identifying extreme events using extreme responses) and allowed us to thoroughly test such assumptions. However, this approach also has some limitations. First, enforcing this extreme definition always leads to a fixed number (i.e., 5 %) of extreme days per site. For long enough and strongly varying time series, this approach yields actual extreme events. However, for shorter time series or sites with weakly varying climate (e.g., tropical sites), this method may lead to a false positive extreme-event identification of non-extreme conditions. The WAI extreme detection is probably more strongly affected by this problem. For the fairly smooth time series with long periods of low and only slightly varying WAI at several sites (see for example the IT Ro1 WAI time series of 2003 in Fig. ), this approach probably led to rather arbitrary breaks between extreme and non-extreme time spans caused by only very small WAI differences. A more flexible data-driven approach to determine site-specific extreme thresholds may be helpful for alleviating this problem in future approaches. For WAI in particular, an ecosystem and soil-type-specific threshold may lead to improved results. Finally, future approaches should take additional extreme-strength indicators like amplitude or occurrence into account when defining the extreme threshold. One also has to note that FLUXNET sites are not necessarily well placed to capture extreme events .

We used changes in the CO2 fluxes to quantify the impact of the extreme events. Such changes, however, can only be defined relative to an undisturbed reference period. Due to the strong seasonal cycles at many of the investigated sites, we used fluxes from other years but identical periods (in the year) as these reference values. However, shifts in the phenological cycle between years could bias these reference values, especially during stages of steep phenological changes at the beginning and end of the growing season. We used smoothed data from multiple years to attenuate this effect. A promising future improvement would be to synchronize each yearly cycle with a reference by shifting it in time until a maximum agreement is reached. For short extreme events, the impact could alternatively be calculated with regard to the fluxes before and or after the extreme.

In addition to the methodological modifications and improvements outlined above (Sect. ) there are several promising methodological extensions and possibilities.

For strongly fluctuating time series such as air temperature, our method of defining individual days as extreme and subsequently joining them into concurrent extreme events often resulted in the identification of several successive but interrupted events. These were then analyzed and treated independently, which may neglect their cumulative impact e.g., on the ecosystem. We alleviated this effect by joining large extreme events with small gaps in between, but our choice of when to join the extreme events and when to treat them separately was rather ad hoc. Such problems could be solved by applying a moving-window-based approach when detecting the extreme events, which takes into account the “extremeness” of a defined period before each individual day. In particular, this approach could improve the results for the multivariate extreme events where the fluctuations in temperature led to many small, fragmented extreme events.

In addition, our method for defining multivariate extreme events is (intentionally) simple and suffers from some restrictions. By independently identifying extreme events in each climate forcing (i.e., temperature and WAI), we may miss out potentially differing impact thresholds in situations when both forcings are extreme. A true multivariate extreme-detection methodology, possibly also including other variables such as vapor pressure deficit or radiation, could overcome this limitation e.g.,.

One important aspect of extreme-event impacts on ecosystems not covered by the approach presented here are lagged or carry-over (i.e., memory) effects e.g.,. These are impacts which persist even after the end of the actual extreme or occur only after the event or during subsequent growing seasons. In addition, extreme events outside of the growing season (i.e., frost events during winter) are not investigated here. We chose to focus on instantaneous effects and neglect such lagged aspects because only the direct and unambiguous connection of possible impacts to one unique extreme event ensured a large enough sample size to apply the assumption-free approach and test all possible extreme-event sizes and types for impacts. However, focusing on a subset of long and pronounced extreme events, an identical approach could be used to assess non-instantaneous effects. An additional interesting aspect would be to examine the effect of the size of the time span between the extreme-event onset and the flux response for Reco and GPP (i.e., the size of the “lag”) and a possible difference between the two fluxes e.g.,.