Contrasting biosphere responses to hydrometeorological extremes : revisiting the 2010 western Russian Heatwave

revisiting the 2010 western Russian Heatwave Milan Flach1, Sebastian Sippel2, Fabian Gans1, Ana Bastos3, Alexander Brenning4,5, Markus Reichstein1,5, and Miguel D. Mahecha1,5 1Max Planck Institute for Biogeochemistry, Department of Biogeochemical Integration, P.O. Box 10 01 64, 07701 Jena, Germany 2Norwegian Institute of Bioeconomy Research, Ås, Norway 3Ludwig-Maximilians University, Department of Geography, Munich, Germany 4Friedrich Schiller University Jena, Department of Geography, Jena, Germany 5Michael Stifel Center Jena for Data-driven and Simulation Science, Jena, Germany Correspondence: Milan Flach (milan.flach@bgc-jena.mpg.de)

Abstract.Combined droughts and heatwaves are among those compound extreme events that induce severe impacts on the terrestrial biosphere and human health.A record breaking hot and dry compound event hit western Russia in summer 2010 (Russian heatwave, RHW).Events of this kind are relevant from a hydrometeorological perspective, but are also interesting from a biospheric point of view because of their impacts on ecosystems, e.g., reductions in the terrestrial carbon storage.Integrating both perspectives might facilitate our knowledge about the RHW.We revisit the RHW from both a biospheric and a hydrometeorological perspective.We apply a recently developed multivariate anomaly detection approach to a set of hydrometeorological variables, and then to multiple biospheric variables relevant to describe the RHW.One main finding is that the extreme event identified in the hydrometeorological variables leads to multidirectional responses in biospheric variables, e.g., positive and negative anomalies in gross primary production (GPP).In particular, the region of reduced summer ecosystem production does not match the area identified as extreme in the hydrometeorological variables.The reason is that forestdominated ecosystems in the higher latitudes respond with unusually high productivity to the RHW.Furthermore, the RHW was preceded by an anomalously warm spring, which leads annually integrated to a partial compensation of 54 % (36 % in the preceding spring, 18 % in summer) of the reduced GPP in southern agriculturally dominated ecosystems.
Our results show that an ecosystem-specific and multivariate perspective on extreme events can reveal multiple facets of extreme events by simultaneously integrating several data streams irrespective of impact direction and the variables' domain.Our study exemplifies the need for robust multivariate analytic approaches to detect extreme events in both hydrometeorological conditions and associated biosphere responses to fully characterize the effects of extremes, including possible compensatory effects in space and time.

Introduction
One consequence of global climate change is that the intensity and frequency of heatwaves will most likely be increasing in the coming decades (Seneviratne et al., 2012).Heatwaves co-occurring with droughts form so-called compound events, for which we can expect severe impacts on the functioning of land ecosystems (e.g., primary production, von Buttlar et al., 2018) that may affect human well-being (e.g., via reduced crop yields, health impacts) (e.g., Scheffran et al., 2012;Reichstein et al., 2013;Lesk et al., 2016).Investigating historical extreme events offers important insights for deriving mitigation strategies in the future.One well-known example of a compound extreme event is the 2010 western Russian heatwave (RHW).The RHW was one of the most severe heatwaves on record, breaking temper-Published by Copernicus Publications on behalf of the European Geosciences Union.
ature records of several centuries (Barriopedro et al., 2011).It was accompanied by extensive wild and peat fires with smoke plumes about 1.6 km high at the peak of the heatwave in early August, and estimated emissions of around 77 Tg carbon due to multiple fire events (Guo et al., 2017).Carbon losses due to reduced vegetation activity were estimated to be in the same order of magnitude as losses due to fires (90 Tg, Bastos et al., 2014).The amount of emitted carbon monoxide was almost comparable to the anthropogenic emissions in this region (Konovalov et al., 2011).Approximately 55 000 cases of death have been attributed to health impacts of the RHW (Barriopedro et al., 2011).
However, to fully understand the developments and impacts of heatwaves or droughts, apart from hydrometeorological drivers, associated land surface dynamics and feedbacks need to be considered (Seneviratne et al., 2010).For instance, under persistent anticyclonic and dry conditions, land-atmosphere feedbacks are expected to further amplify the magnitude of heatwaves via enhanced sensible heat fluxes, as shown also for the RHW (Miralles et al., 2014;Hauser et al., 2016).These feedback mechanisms highlight the importance of depleted soil moisture to heatwaves.In 2010 a negative soil moisture anomaly contributed to increased temperatures (Hauser et al., 2016).It is a general observation that the combination of anticyclonic weather regimes and initially dry conditions prior to the event amplifies heatwaves in most cases (Quesada et al., 2012).
The direct impacts of such extreme events on ecosystems are manifold.Summer heat and drought typically reduce (or even inhibit) photosynthesis, hence reducing the carbon uptake potential of ecosystems (Reichstein et al., 2013).However, the magnitude of these impacts varies between ecosystems (Frank et al., 2015), and the resulting net effects are still under debate, particularly for heatwaves (Sippel et al., 2018).However, in-depth investigations of a number of individual events such as the European heatwave 2003 (Ciais et al., 2005), the 2000-2004 and 2012 droughts in North America (Schwalm et al., 2012;Wolf et al., 2016), and the RHW (Bastos et al., 2014) agree on an overall tendency towards negative impacts on the carbon accumulation potential.
The RHW has been thoroughly investigated from a hydrometeorological point of view linking the atmospheric blocking to the large-scale positive anomalies in air temperatures and negative anomalies in water availability (e.g., Barriopedro et al., 2011;Rahmstorf and Coumou, 2011).The event has also been well investigated, with an emphasis on the biospheric impacts describing the negative anomalies in ecosystem productivity and related vegetation indices (e.g., Bastos et al., 2014).However, comparing the reports of areas affected by the RHW reveals some discrepancies.Hydrom-eteorological anomalies point to much larger areas affected compared to biosphere response patterns.Figure 1 shows the zonal evolution of the RHW in both domains.We find that the spatiotemporal patterns of the temperature anomaly do not match the zonal anomaly in vegetation productivity anomalies.Thus, an integrated assessment including the hydrometeorological and biospheric domains simultaneously may further our understanding of the RHW.
The figure reveals an unusually warm period during spring and one longer heatwave during summertime (Fig. 1a).Temperature anomalies exceed more than 10 K in both spring and summer, but they lead to distinctive anomalies in gross primary productivity (GPP).Positive GPP anomalies occur during the spring event, whereas negative GPP anomalies occur during the summer heatwave.The positive GPP response in spring might be a reaction to warmer, more optimal spring temperatures (Wang et al., 2017) possibly accompanied by enough water availability.However, negative GPP anomalies in summer occur only in areas south of 55 • N (Fig. 1c), indicating that the GPP response involves far more processes than high temperatures and drought during the unique RHW.As already indicated by Smith (2011), the connection between biosphere and hydrometeorology is much more complex than just a direct one-to-one mapping.Further complicating this issue is the fact that the summer event cannot be investigated without the previous spring as both seasons are inherently related via memory effects in water availability.Increased GPP in spring due to warm temperatures can negatively influence soil moisture and thus GPP during summer (Buermann et al., 2013;Wolf et al., 2016;Sippel et al., 2017).In particular, Buermann et al. (2013) show for North American boreal forests that earlier springs are followed by reduced productivity in summer because of water constraints.
In summary, comparing these two Hovmöller diagrams shows that (1) the affected latitudinal range of the negative GPP anomaly is much smaller than the positive temperature anomaly and (2) the evolution of the summer impacts should consider potential carry-over effects of positive GPP anomalies during spring, as earlier studies showed that earlier spring onset and increased spring GPP may negatively influence soil moisture and thus GPP during summer (Buermann et al., 2013).The objective of this paper is to revisit the RHW and to investigate the GPP response during the spring event and the summer heatwave in detail by investigating spatiotemporal anomalies in hydrometeological drivers and ecological variables.
This kind of integrated assessment requires a generic methodological approach.Here, we use a multivariate extreme event detection approach that (1) does not differentiate between a positive and negative extreme event, and (2) can equally be applied on any set of time series, regardless of whether they describe the biospheric or hydrometeorological domain.We expect that we can reveal previously overlooked facets in the RHW and discuss whether our approach may fa-cilitate a broader perspective and improved interpretation of extreme events and their impacts.
2 Methods and data

Rationale
One approach to detect extreme events like the RHW could be to identify the peaks over some threshold in the marginal distribution of a variable (or its anomaly) of interest.For instance, one could identify values that deviate by more than 2 standard deviations from the variable's mean values (Hansen et al., 2012;Sippel et al., 2015).However, univariate approaches only allow us to characterize an event by, e.g., extremely high temperature anomalies, lack of precipitation, or very low soil moisture but not their compound anomaly.However, from earlier studies (e.g., Miralles et al., 2014;Hauser et al., 2016) we know that more than one variable is involved in the RHW, and a multivariate extreme event detection (i.e., a compound event, Leonard et al., 2014;Zscheischler and Seneviratne, 2017) is more feasible.Multivariate algorithms to detect extreme events are expected to offer more robust detection capabilities when accounting for dependencies and correlations among the selected variables (e.g., Zimek et al., 2012;Bevacqua et al., 2017;Flach et al., 2017;Mahony and Cannon, 2018).Multivariate extreme event detection considers all observable dimensions of the domain simultaneously.With a multivariate approach one may, for instance, detect very rare combinations of variables even if the individual variables are not extreme.In the following, we detect the anomalies in a multivariate variable space in two sets of variables describing (1) the hydrometeorological conditions and (2) the biospheric response.The workflow involves a data pre-processing to compute anomalies, a step for dimensionality reduction to not be biased by redundancies among variables.Based on the reduced data space, an anomaly score is computed that can then be used as a threshold.For various reasons, however, in practice the threshold needs to be computed across multiple spatial grid cells of comparable phenology.

Data and pre-processing
Our data set for analyzing the hydrometeorological domain includes those variables which we consider to be of particular importance for processes taking place during extreme events in the biosphere based on prior process knowledge (Larcher, 2003) and empirical analysis (von Buttlar et al., 2018).The hydrometeorological data set consists of air temperature, radiation, relative humidity (original resolution 0.71 • , all three from ERA-INTERIM, Dee et al., 2011), precipitation (original resolution 1 • , Adler et al., 2003), and surface moisture (resolution 0.25 • , http://www.gleam.eu,last access: 12 October 2018, v3.1a, Miralles et al., 2011;Martens et al., 2017).We consider surface moisture to be a hydrometeorological variable due to its importance for drought detection, although we notice that surface moisture is influenced by biospheric processes.We use gross primary productivity (GPP), latent heat flux (LE), sensible heat flux (H ) (resolution of 0.25 • , all three from FLUXCOM-RS, Tramontana et al., 2016), and the fraction of absorbed photosynthetic active radiation (original resolution 1 km, FAPAR, moderate resolution imaging spectroradiometer (MODIS) based FAPAR; Myneni et al., 2002) to describe the land surface dynamics.
The selected variables cover the spatial extent of Europe (latitude 34.5-71.5 • N, longitude: −18-60.5 • E) and are regridded on a spatial resolution of 0.25 • from 2001 to 2011 in an 8-daily temporal resolution.The temporal extent is selected as it is covered by all data sets used in the study.To check for differences in land cover types, we estimate the dominant land cover type of the European Space Agency Climate Change Initiative land cover classification on a spatial resolution of 0.25 • (original: 300 m).To check for consistency of our findings among other variables (Sect.3.2), we additionally use terrestrial ecosystem respiration (TER) and net ecosystem productivity (NEP, both originating from FLUXCOM-RS, Tramontana et al., 2016).
The actual event detection is realized on the anomalies of these data sets.To compute the anomalies, for each variable under consideration, we first estimate the seasonality as a smoothed median seasonal cycle per grid cell.We use the median instead of the mean as it is less susceptible to outliers.We then subtract these seasonal cycles from each variable and year to obtain a multivariate data cube of anomalies (Fig. 2, step 1).Small data gaps are set to zeros to ensure that they are not detected as anomalies.The gap filling is necessary for a multivariate detection approach as there are many more cases in which one variable is missing in the multivariate cube compared to a univariate data stream.

Feature extraction and anomaly detection
We use a multivariate anomaly detection algorithm proposed by Flach et al. (2017) and apply it separately to two sets of variables for the biosphere and hydrometeorology.The method expects a multivariate set of anomalies and projects them to a reduced space via principal component analysis, retaining a number of principal components that explain more than 95 % of the variance (Fig. 2, step 3b).This procedure accounts for linear correlations in the data only by removing redundancies among the variable anomalies.
We compute an anomaly score via kernel density estimation (KDE, Parzen, 1962;Harmeling et al., 2006) in the reduced anomaly space (Fig. 2, step 4).KDE showed very good performance among different other options to detect multivariate anomalies in previous experiments (Flach et al., 2017).One strength of KDE is that it considers nonlinear dependencies among dimensions (Fig. 3).The anomaly scores are transformed into normalized ranks between 1.0 (very anomalous, data point in the margins of the multivariate dis-  tribution) and 0.0 (completely normal, data point in the dense region of the multivariate distribution; Fig. 2, step 5).In this univariate index of compound extremes, it is legitimate to use a classical threshold that can be intuitively analyzed.However, to avoid an equal spatial distribution of event occurrences we do not apply this multivariate anomaly detection per pixel, but rather by region.

Spatiotemporal segmentation
The spatiotemporal segmentation aims to identify spatial areas of comparable phenology, climate, and seasonality.To identify these regions, we follow the methodology described by Mahecha et al. (2017) and extend it to the multivariate case.The main idea is that the (now spatial) principal components of the mean seasonal cycles can be used for classifying regions according to their characteristic temporal dynamics.
The procedure for extracting spatial segments of similar grid cells works as follows (for a detailed description, see Mahecha et al., 2017).
1. We estimate the median seasonal cycle in each grid cell and of each variable individually and standardize the median seasonal cycles to zero mean and unit vari-ance to get the cycles comparable across different units (Fig. 4, step 1).
2. To remove the effect of different phasing (similar but only lagged seasonal cycles), we sort the median seasonal cycles according to a variable showing a strong seasonality, which is temperature in our case.Thus, we memorize how to bring temperature in a sorted increasing or decreasing order (the "permutation" of temperature) and apply the same permutation to the other median seasonal cycles (Fig. 4, step 2).We prepare the data for dimensionality reduction by concatenating the seasonal cycle of all variables to a matrix seasonal cycles × space.We apply a principal component analysis (PCA) to reduce the dimension of the concatenated median seasonal cycles.
3. We select locations (grid cells) of similar phenology and climate by dividing the orthogonal principal component subspace into equally sized bins (Fig. 4, step 3).We used N PC = 4 components in this step, explaining 71 % of variance.The bins are sufficiently small compared to the length of the principal components to ensure a fine binning of very similar phenology and climate.4. We compute the multivariate anomaly score in an overlapping moving window for all grid cells that fall into one of the bins (the central bin and the neighboring bins, Fig. 4, step 4).
A final detail to consider is the effect of changing seasonal variance (temporal heteroscedasticity).These patterns lead to detecting extreme events predominantly during the highvariance seasons (i.e., summer times).To avoid seasonal biases in the extreme event detection, we additionally apply the entire anomaly detection scheme to seasonally overlapping moving windows across years.Within the spatiotemporal segmentation procedure, we ensure that the number of observations is at least 198 (9 time steps × 11 years, at least one spatial replicate).To reunify the spatiotemporal segments, we assign the normalized anomaly scores temporally to the time step in the center of the temporal moving window and spatially to the grid cell in the central bin of similar climate and phenology.

Statistics of extreme events
We assume that 5 % of the data are anomalous in each overlapping spatiotemporal segment and convert the anomaly scores into binary information.However, the main results of compensation effects are not sensitive to this threshold selection (Appendix Table A1, varying the threshold between 1 % and 10 %).To compute statistics based on the www.biogeosciences.net/15/6067/2018/Biogeosciences, 15, 6067-6085, 2018 spatiotemporal structure of each extreme event, we follow an approach developed by Lloyd-Hughes (2011) and Zscheischler et al. (2013) and compute the connections between spatiotemporal extremes if they are connected within a 3 × 3 × 3 (long × lat × time) cube.Each connected anomaly is considered as a single event (Fig. 2, step 6).In this way, we observe event-based statistics, i.e., affected area (km 2 ), affected volume (km 2 days −1 ), centroids of the area, and histograms of the single variable anomalies stratified according to different ecosystem types (land cover classes).Furthermore, we observe the response of individual variables to the multivariate event by computing the area weighted sum of the variable during the event in which the variable of interest is positive relative to the seasonal cycle (res + ) or negative (res − ), respectively.For many biospheric variables, one expects a mainly negative response to hydrometeorological extreme events like heatwaves or droughts (Larcher, 2003;von Buttlar et al., 2018).Thus, we define compensation of a specific variable to be the absolute fraction of res + from res − .The balance of a variable is the sum of res + and res − .Centroids of res + and res − are computed as the average of the affected longitudes, latitudes, and time period, weighted with the number of affected grid cells at this longitude, latitude, and time period, and its respective anomaly score.They are used to compute the spatial and temporal distance between res + and res − .Affected area, volume, response, and centroids take the spherical geometry of the Earth into account by weighting the affected grid cells with the cosine of the respective latitude.

Extreme events in western Russia in 2010
We identify two multivariate extreme events in the set of hydrometeorological variables in western Russia 2010, based on the spatiotemporal connectivity.The two extreme events are separated by approximately 1 week of normal conditions towards the end of May.
Both multivariate hydrometeorological anomalies partly overlap with a multivariate anomaly in the set of biosphere variables (biospheric spring event and biospheric summer event).Of specific interest is that the area affected by anomalous hydrometeorological summer conditions is remarkably larger than the one detectable in the biospheric variables (biospheric summer event, 2.4 × 10 6 vs. 1.1 × 10 6 km 2 , Table 1).This fact already indicates that biosphere responses are more nuanced than the hydrometeorological events and do not simply follow the extent of the hydrometeorological anomaly.
As indicated, e.g., also by Smith (2011), a hydrometeorological extreme event does not necessarily imply an extreme response.

Hydrometeorogical events
As GPP is a key determinant of ecosystem-atmosphere carbon fluxes, we focus on the gross primary productivity (GPP) response to the multivariate hydrometeorological anomaly: we find that the GPP response is entirely positive during the short-lasting hydrometeorological spring event (+17.8Tg C, Table 1), while it is mainly negative during the summer event (+8.8, −49 Tg C, Table 1).A part of the GPP summer losses (18 %) associated with the RHW in the southern region are instantaneously reduced by over-productive vegetation in the higher latitudes, which are hit by the extreme event.Please note that the carbon balance in summer accounts for the GPP response to the same hydrometeorological extreme event, namely the RHW, which leads to contrasting responses in adjacent regions.If we estimate the annually integrated effect of the anomalies, another 36 % of the carbon losses are compensated during spring in higher latitudes.We did not find extreme events after summer, which implies a fast recovery of vegetation activity after summer.Integration over the spring and summer events thus equals the annual integration.Overall, we find that 54 % of the negative GPP anomalies are compensated either because of the positive spring anomalies or across ecosystems hit by the same event during summer.These compensation effects reduce the negative carbon impact of integrated annual hydrometeorological events from −49.0 to −24 Tg C in total (Table 1).We want to emphasize that the negative impact of the RHW in terms of GPP is just reduced, and still negative in total.

Biospheric events
Moving the focus to the multivariate biosphere events (biospheric spring and biospheric summer event), which overlap with the hydrometeorological events, we find that GPP responses based on the biospheric spring event are almost entirely positive (+33.8Tg C), and based on the biospheric summer event almost entirely negative (−82.6 Tg C).If we consider the annually integrated effect of the anomalies, spring carbon gains are estimated to offset 41 % of the subsequent carbon losses in summer (56 days earlier) in the higher latitudes (514 km distance of the centroids, Table 1).To further examine these findings, we check for these kinds of compensation effects among different variables and another GPP www.biogeosciences.net/15/6067/2018/Biogeosciences, 15, 6067-6085, 2018 Table 1.Statistics of the extreme events, based on their spatiotemporal connectivity structure: affected area, affected volume, positive and negative GPP response (res +/− ) to the event, compensation of the negative response (comp.), as well as average spatial and temporal distance between the parts of the events with positive and negative responses.event area (km 2 ) volume (km 2 days −1 ) GPP comp.res + GPP res data set in the following section.Note that the data set of biosphere variables includes GPP itself.Computing the responses based on the extent of the biospheric event is nevertheless useful, as an extreme event in the biosphere variables is not exclusively restricted to extreme conditions in the hydrometeorological conditions (Smith, 2011).

Compensation in other data sets and variables
The annually integrated compensation effect in GPP is highly consistent among different variables.For instance, NEP (excluding fire) shows such a kind of compensation, but also FAPAR and LE (Table 2).Sensible heat flux, on the other hand, is high during the hydrometeorological summer event (biospheric summer event) as well as the hydrometeorological spring event (biospheric spring event), as expected for strong positive temperature anomalies.However, some of the remote sensing data products might be affected by high fire induced aerosol loadings during the heatwave that affect atmospheric optical thickness (e.g., Guo et al., 2017;Konovalov et al., 2011).Exploring an almost entirely climatedriven GPP product (FLUXCOM RS + METEO, Jung et al., 2017), we also find the integrated compensation effect, although much less pronounced (Appendix Fig. B1).Thus, we are confident that the observed compensation effect is not related to the optical thickness during the RHW.

Influence of vegetation types
In Fig. 6 we present the histograms of GPP anomalies for different land cover classes (forests, grasslands, and crops) based on the hydrometeorological spring event and hydrometeorological summer event (biospheric spring event and biospheric summer event, respectively, Fig. C1) to highlight two aspects: first, during the spring event (hydrometeorological spring or biospheric spring), forests react almost entirely with positive GPP anomalies (Fig. 6a).Forests in this region are energy-limited, so the timing of the extreme event leads to hydrometeorological conditions (e.g., positive temperature anomalies in spring, more incoming radiation accompanied by enough water availability) which are favorable for vegetation productivity, as absolute spring temperatures are still below the temperature optimum of GPP (Fig. 8a, Wolf et al., 2016;Wang et al., 2017).Second, during the hydrometeorological summer event, we observe positive to neutral GPP responses in forests, whereas crops and grasslands react strongly negatively (Fig. 6b).The positive vs. negative GPP responses almost entirely reflect the map of dominant vegetation types (forest vs. agricultural ecosystems, Fig. 7).However, different vegetation types exhibit a transition from higher latitudes (predominantly forest ecosystems) to lower latitudes (dominated by agricultural ecosystems).Thus, the different responses of vegetation types might be confounded by the fact that absolute temperatures also follow a latitudinal gradient (Fig. 1b).Absolute temperatures for agricultural ecosystems are higher and far beyond the temperature optimum of GPP (Fig. 8c).Additionally, agricultural ecosystems are drying out in summer (low soil moisture, Fig. 8c).In contrast, forestdominated ecosystems at higher latitudes experience temperatures just slightly above the temperature optimum of GPP, accompanied by high soil moisture (Fig. 8b).The response of forest ecosystems partly reflects a latitudinal gradient: forest ecosystems in the lower latitudes react positively to the spring temperature anomaly and then tend to react more negatively to the summer heatwave than forest ecosystems in higher latitudes.Forest ecosystems in higher latitudes are still productive in terms of GPP during the peak of the heatwave (Fig. 9).We find negligible anomalies in autumn for both ecosystems, which implies a fast recovery after the heatwave.
To disentangle the variable importance of the different confounding factors, we run a simple linear regression model which tries to explain GPP as a function of the hydrometeorological driver variables (temperature, precipitation, radiation, and surface moisture, including their anomalies and absolute values), as well as vegetation type, duration and latitude (Appendix D).We use an algorithm following Chevan and Sutherland (1991) which extracts the independent contribution of the variable importance related to this particular variable regardless of the model complexity or dependencies among variables.The model reveals from a statistical point of view that vegetation type and the latitudinal gradient are the most important variables explaining GPP during the summer event, followed by the hydrometeorological drivers.Access to deeper water and soil type as well as nonlinear feedbacks are factors which are not represented in the model but might explain the high importance of latitude.Apart from vegetation type being important for the GPP response, underlying water use efficiency (calculated according to Zhou et al. ( 2014) is consistently higher in forest-dominated ecosystems compared to agriculture-dominated ecosystems (Appendix Fig. E1a), and higher evaporative fraction in forest ecosystems during the peak of the heatwave (Appendix Fig. E1b).

Discussion
In this paper we show that the hydrometeorological extreme events affecting western Russia in spring and summer 2010 do not directly map to the observed vegetation responses.Positive to neutral GPP responses prevail in higher latitudes during summer, whereas strong negative impacts on GPP can be found in lower latitudes.We interpret this effect by different water management strategies of forest vs. agricultural ecosystems (Teuling et al., 2010;van Heerwaarden and Teuling, 2014) that meet a general latitudinal temperature gradient.Apart from a more efficient water usage of forestdominated ecosystems, access to deeper soil water might be another reason for ecosystem-specific responses (Fan et    2017; Yang et al., 2016).Note that the latitudinal temperature gradient alone might explain differences in the response within ecosystems in summer and between spring and summer, but does not sufficiently explain differentiated GPP responses in summer among different ecosystems (predominantly forest vs. agricultural ecosystems).
Another important aspect is that the combination of the anomalous spring and the unique heatwave in summer might be inherently connected via land surface feedbacks.Buermann et al. (2013) showed that warmer springs going hand in hand with earlier vegetation activity negatively affect soil moisture in summer, and thereby vegetation activity.It is a general observation that warm and dry springs enhance summer temperatures during droughts, which suggests the presence of soil-moisture temperature feedbacks across seasons (Haslinger and Blöschl, 2017).In the case of the Russian heatwave 2010, soil moisture was one of the main drivers (Hauser et al., 2016), hand in hand with persistent atmospheric pressure patterns (Miralles et al., 2014).Thus, we suspect that the spring event is connected to the summer heat-wave in 2010, if not setting the preconditions for a heatwave of this unique magnitude.
The integration of the carbon balance over spring and summer might be justified by assumed connections between spring and summer as outlined before.However, we would like to note that common annual integration and assessment of compensatory effects on the carbon balance over events during the growing season equal the integration over spring and summer for this particular case, as we did not find any events after summertime.The absence of events after the summer heatwave implies a fast recovery of the ecosystems.
Compensations of parts of the negative impacts on the carbon balance during hydrometeorological extreme events have been reported in earlier studies.On the one hand, Wolf et al. (2016) report that a warm spring season preceding the 2012 US summer drought reduced the impact on the carbon cycle.Yet on the other hand, the increased spring productivity amplified the reduction in summer productivity by spring-summer carry-over effects via soil moisture depletion: higher spring productivity leads to higher water con- sumption in spring.The high water additionally consumed during spring reduces the water availability in summer and thereby affects productivity during the following summer.However, it remains unclear whether this observation was a singular case or whether it could become a characteristic pattern to be regularly expected in a warmer world.In this study, we provide some evidence for presumed comparable effects.In contrast to the discussion in Wolf et al. (2016), enhanced productivity does not exclusively occur temporally, i.e., spring partly compensates for summer losses, but rather spatially adjacent forest ecosystems are reducing the negative impact of agricultural ecosystems on the carbon balance.Spatially adjacent ecosystems partly compensating carbon losses due to drought or heatwaves have been observed earlier, e.g., in mountainous ecosystems that respond differently than lowlands during the European heatwave 2003 (Reichstein et al., 2007).
Following up on compensatory effects, Sippel et al. ( 2017) use ensemble model simulations to disentangle the contribution of spring compensation vs. spring-summer carry-over effects on a larger scale.They show that in general warm springs compensate for parts of summer productivity losses in Europe, whereas spring-summer carry-over effects are constantly counteracting by enhancing summer losses.Also, Mankin et al. (2017Mankin et al. ( , 2018) ) note that increased spring productivity with spring-summer carry-over effects can be observed in Earth system models.We can confirm the general finding that spring partly compensates for summer productivity losses in observations for our case study on the RHW.Without using model simulations it is difficult to quantify spring-summer carry-over effects via soil moisture depletion.In the case of the RHW only very few areas are anomalously productive in terms of GPP in spring and unproductive in summer as well.Thus, we suspect that exclusively temporal spring-summer carry-over effects play a rather small role for the RHW.However, we also emphasize that longer-term effects, such as effects in subsequent years through species changes (Wagg et al., 2017), have not been considered in the present study and likely remain hard to quantify beyond dedicated experiments.
The RHW is among the best studied extreme events in the Northern Hemisphere.However, the enhanced productivity of northern forests which diminishes the negative carbon impact of the RHW as reported in this study has only received marginal attention so far.For instance, Wright et al. (2014) mention positive NDVI anomalies in spring 2010, but then focus largely on productivity losses in the Eurasian wheat belt.Similarly, Bastos et al. ( 2014) focus on a spatial extent of the biosphere impacts that only partly includes forest ecosystems at higher latitudes.Our estimation of carbon losses due to decreased vegetation activity (82 Tg C) is comparable to the one of Bastos et al. (2014) (90 Tg C).Similar to the results of our study, Yoshida et al. (2015) report reductions in photosynthetic activity in agriculture-dominated ecosystems during the RHW, but only small to no reductions in forest ecosystems during summertime.However, their interpretations focus on the summer heatwave.Nevertheless, re-evaluating impact maps (published, e.g., in Wright et al., 2014;Yoshida et al., 2015;Zscheischler et al., 2015) in the light of our findings suggests that their evidence supports the presence of contrasting responses, differing among ecosystems during the RHW.When it comes to extreme events, the general tendency in many existing studies is naturally to focus on negative impacts as they are of particular interest for society (Bastos et al., 2014;Wright et al., 2014;Yoshida et al., 2015;Zscheischler et al., 2015).

Conclusions
We re-analyzed biospheric and hydrometeorological conditions in western Russia in 2010 with a generic spatiotemporal multivariate anomaly detection algorithm.We find that the hydrometeorological conditions and the biospheric responses exhibit two anomalous extreme events, one in late spring (May) and one over the entire summer (June, July, August), covering large areas of western Russia.For the summer event, we find that the spatially homogeneous anomaly www.biogeosciences.net/15/6067/2018/Biogeosciences, 15, 6067-6085, 2018 pattern (characterized by high solar radiation and temperature, low relative humidity, and precipitation) translates into a bimodal and contrasting biosphere response.Forest ecosystems in higher latitudes show a positive anomaly in gross primary productivity, while agricultural systems decrease their productivity dramatically.
If we consider the annually integrated effect of the anomalous hydrometeorological conditions in 2010, we find that forest ecosystems reduce the negative impact of the productivity losses experienced in agricultural ecosystems by 54 % (36 % during spring, 18 % during summer).Please note that the annually integrated impact of the 2010 events on the carbon balance stays strongly negative.Our findings do not alleviate the consequences of extreme events for food security in agricultural ecosystems.
From a methodological point of view, this study emphasizes the importance of considering the multivariate nature of anomalies.From this study, we learn that it is insightful to consider the possibility of both negative as well as positive impacts and to assess their annually integrated statistics.Although the integrated impact of gross primary production on the hydrometeorological conditions in 2010 is strongly negative, it is important to notice the partial compensatory effects due to differently affected ecosystem types, as well as timing of the extreme events.

Figure 3 .Figure 4 .
Figure 3. Illustration of the multivariate anomaly detection algorithm with two variables.The data have (a) linear dependencies (multivariate normal) and (b) a nonlinear dependency structure.Univariate extreme event detection (peak-over-threshold in the marginal distribution of a variable) does not follow the shape of the data, whereas algorithms assuming a multivariate normal distribution (Hotelling's T 2 , Lowry and Woodall, 1992) are suitable for case (a); kernel density estimation (KDE) gets the shape of the data in both cases (a) and (b); 5 % extreme anomalies are outside the shaded areas (region of "normality") for all three algorithms.

Figure 5 .
Figure5.Left column: temporal duration of the (a) hydrometeorological spring event, (c) hydrometeorological summer event, and biospheric events (e, g).Right column: corresponding GPP response, i.e., the sum of deviations from the seasonal cycle during the event for the (b) hydrometeorological spring event, (d) hydrometeorological summer event, and biospheric events (f, h).While the GPP response during the hydrometeorological spring event is entirely positive (more productive than usual, b), GPP response during the hydrometeorological summer event differs between higher latitudes (> 55 • N, short-lasting, positive) and lower latitudes (long-lasting, negative).

Figure 6 .Figure 7
Figure 6.Histogram of GPP anomalies (reference period: 2001-2011) for different land cover classes based on the spatiotemporal extent of (a) the hydrometeorological spring event and (b) the hydrometeorological summer event.Bars denote the sum of all vegetation classes.

Figure 8 .
Figure 8. Temperature optimality for GPP in (a) forests during spring, (b) forests during summer, and (c) crops during summer.Contour lines enclose 75 % of the data points.

Figure 9 .
Figure 9. Temporal evolution of the GPP anomaly (reference period: 2001-2011) for (a) agricultural ecosystems and (b) forest ecosystems, colored according to the latitude.

Table 2 .
Negative responses to the RHW are partly compensated based on the integrated biospheric or hydrometeorological events in 2010.The finding is consistent over different variables and data sets.