In ecological studies, drought is better characterised by soil moisture anomalies i.e. agricultural drought , than atmospheric drought indices. We therefore base our drought assessment on two complementary soil moisture datasets. The first is the observation-based soil moisture dataset obtained from SoMo.ml , used as reference in this study, and the second, for comparison with SoMo.ml, is given by ERA5 volumetric soil water content .

The SoMo.ml data are generated using a long short-term memory neural network model trained with meteorological forcing from ERA5 and land surface characteristics as inputs and global in situ soil moisture measurements as target variables. The data cover soil moisture in the first 50 cm of the soil and are available at 0.25∘ lat/long resolution and daily time steps for the period 2000–2019. We remapped the fields to the finer resolution of the MODIS grid and aggregated the data to monthly means. We then subtracted the mean seasonal cycle and long-term linear trend and divided this by the corresponding standard deviation to obtain standardised soil moisture anomalies (SManom).

Monthly temperature and volumetric soil water content from the ECMWF ERA5 Reanalysis were obtained from the Copernicus Climate Change Service at 0.25∘ lat/long resolution at monthly time steps, selected for the period 2000–2019 (common with SoMo.ml), and remapped to the finer resolution of the MODIS grid using conservative remapping. Standardised anomalies were calculated as described for SManom for ERA5 temperature and soil moisture fields (Tanom, SManomERA5). Soil moisture anomalies from ERA5 in layers 1–2 (top 28 cm) are used for comparison of drought conditions with those estimated by SoMo.ml, although the two datasets are not fully independent.

We used the 16 d enhanced vegetation index (EVI) from the Moderate Resolution Imaging Spectroradiometer (MODIS) sensor from the MOD13C1 Climate Modeling Grid (CMG) product. The MOD13C1 CMG provides continuous cloud-free spatial composites from 1 km data projected on a 0.05∘ lat/long grid and was selected for the period 2001–2019. Standardised EVI anomalies (EVIanom) were calculated following the same approach as for climate variables. The standardisation allows comparing the relative magnitude of anomalies for pixels with distinct temporal variability patterns and with vegetation productivity simulated by LSMs, which have different physical units.

We used land cover distribution in 2018 from the ESA Climate Change Initiative land cover (CCI-LC). The data are originally provided in land cover classes at 300 m spatial resolution and were converted to fractional cover at 0.05∘ lat/long resolution for forest, grassland, and crop classes using the CCI-LC user tool.

We used isohydricity fields from global satellite measurements from at 1∘ lat/long resolution. Anisohydric plants (low isohydricity) show weak regulation of stomatal opening and prioritise carbon assimilation over water savings during droughts. High isohydric plants show strong stomatal regulation of productivity and thereby preserve water at the cost of carbon assimilation during drought.

We use soil available water capacity (AWC) from and , which used the Land Use and Cover Area frame Statistical survey (LUCAS) topsoil database to map soil properties at continental scale.

Standardised anomalies of gross primary productivity (GPPanom) and soil moisture (SManom) were estimated by the mean of seven land surface models and dynamic global vegetation models (for simplicity referred to as LSMs) between 1979–2019 from an extension of simulations: a baseline simulation for comparison with observations and a factorial simulation to quantify the individual impact of summer 2018 and its legacy effects, when compared to the reference simulation. A detailed description of the models used and the simulation protocol is provided in Appendix .

First, all model outputs were remapped to a common 0.25∘ grid, and the multi-model ensemble mean was calculated for the common period with MODIS (2001–2019). The variables were then deseasonalised, detrended, and standardised as was done for the other variables in the study.

We use the observation-based SoMo.ml as a reference dataset to define agricultural drought conditions. Regions with average SManom below -1σ during summer (June, July, and August, JJA) are considered drought-affected areas during the DH events. Then, a regional domain affected by both DH18 and DH19 events is selected to evaluate the impacts of two consecutive DH events. Within this region most pixels had negative SManom and the majority registered SManom<-1.5σ, but they differ in the magnitude of agricultural drought in DH19. This allows for comparing responses across pixels for different combinations of stress between DH18 and DH19. Since we are interested in evaluating how recovery from DH18 affected impacts of DH19, we limit our analysis to pixels with negative EVIanom in DH18.

Following the extreme summer in central Europe in 2018, mild temperatures and strong soil moisture deficits remained until January 2019, when SManom returned to normal conditions (Figs.  and ). In central Europe, June 2019 was extremely hot, but July and August 2019 were milder (Fig. , ), and soil moisture deficits became very pronounced in July (Fig. ). In this region, except for during April 2019, the months preceding summer were not particularly dry and were even slightly wetter than average in February, March, and May, the latter also being colder than average. Therefore, the DH18 and DH19 constitute more a sequence of two compound events than a single drought. The areas experiencing severe dry and hot conditions in both summers correspond to a region covering central and eastern Europe and southern Sweden. This region is our study domain and indicated by the rectangle in Fig. ). Both DH events led to vegetation browning, though negative EVIanom was more widespread in DH18 than DH19. Within the study region, 79 % of the area showing negative EVIanom in DH18 (EVIanomDH18) also registered negative EVIanom in DH19 (EVIanomDH19).

The spatial distribution of the clusters resulting from the unsupervised classification based on (EVIanomDH18, EVIanomDH19) pairs and corresponding centroids are shown in Fig. a and b, as well as the corresponding (SManomDH18, SManomDH19) and (TanomDH18, TanomDH19) (Fig. c and d). The four clusters aggregate pixels according to different impacts in DH18 and DH19. One cluster, covering 20 % of the area, includes pixels with moderate impacts in DH18 and further browning in DH19, being therefore referred to as (CDecline) (dark brown, EVIanomDH19 below the 1:1 line in Fig. a). This cluster is associated with mixed cover of forests (10 %–40 %, dominated by needle-leaved forest) and grasslands (15 %–60 %), (Fig. ). Cluster CHighV (high vulnerability, covering 15 % of the area) corresponds to pixels experiencing strong impacts in both events and is associated with high grassland and cropland fractions and low tree cover. Pixels with strong impacts in DH18 and weakly negative EVIanomDH19, i.e. partial recovery in DH19 (CPRecov, 21 % of the area), are mainly dominated by croplands. Finally, a group of pixels shows moderate EVIanomDH18 and positive EVIanomDH19 (CGreening, 44 %), corresponding mostly to mixed forest–grassland pixels (30 %–65 % of forest, dominated by needle-leaved forest).

All clusters align along proportional DH18 : DH19 values of SManom and Tanom, with predominantly negative SManom and positive Tanom in both DH events but alleviation of soil moisture deficits and heat stress in DH19 compared to DH18 (Fig. ). The two recovery clusters (CPRecov and CGreening) correspond to pixels with less severe drought conditions and milder temperatures in DH19, and CGreening corresponds to pixels where dry and hot conditions in DH18 were also more moderate. CHighV corresponds to pixels experiencing drier and hotter anomalies in both summers and accordingly shows stronger impacts. Cluster CDecline, however, shows increasing browning in DH19 in spite of drought and heat stress alleviation (Fig. ). The distributions of climate anomalies for each cluster overlap each other and, in some cases, the 1:1 line, indicating that the intensity of the hazards (temperature, drought) cannot account for the resulting impacts alone.

All clusters show significant positive linear relationships between summer EVIanom and SManom and negative linear relationships with Tanom in 2001–2017 (Fig. ). The relationships include the two extreme summers of 2003 and 2015, which had comparable Tanom and SManom to DH18 and DH19 in most clusters. However, the long-term sensitivities estimated are robust even if these summers are excluded.

The results correspond to a general summer water-limited regime, especially in clusters CDecline, CHighV, and CPRecov, which show stronger sensitivities to Tanom and SManom (slopes in Fig. ), and higher variance explained by both models (R2 0.58–0.68 for SManom and 0.49–0.55 for Tanom). For these clusters, EVIanom is below the 95 % confidence interval of the long-term linear relationships for DH18 (CPRecov and CHighV) and DH19 (CDecline and CHighV). SManom and Tanom in DH18 and DH19 are generally similar to those of 2003, but DH18 was drier than 2003 in CPRecov and CHighV.

These departures may be related with seasonal legacy effects from the warm spring in DH18 and/or the onset of non-linear responses to heat and drought. To account for these modulating effects, we model long-term (2001–2017) EVIanom–climate relationships using spring and summer SManom and Tanom as predictors using random forest regression (see Sect. ). The model is able to predict 48 %–90 % (median and maximum out of bag score across pixels) of the pixel-level temporal variability of summer EVIanom in 2001–2017 (Fig. ). Analysis of the variable importance shows that the model estimates summer water limitation and negative legacy effects from spring warming (Fig. ), consistent with a summer water-limited regime and process-based modelling studies .

As in the linear case, the RF model estimates less negative or more positive EVIanom in DH18 and DH19 than observations (Fig. ). The residuals are below the range of the training period for the high-impact clusters: CDecline and CPRecov in DH19 and DH18, respectively, and CHighV in both (Fig. c). In CGreening, residuals are predominantly positive (i.e. observed EVIanom more positive than predicted), but still partly overlap with the range of residuals in the training period (Fig. ).

Pixels with high tree cover tend to show less negative or more positive residuals than pixels with low tree cover (<0.4 %) in both DH events (Fig. ), but in DH19 the range of residuals in high tree cover pixels is larger than DH18, including pixels with strongly negative values. The partial rank correlation of the spatial distribution of EVIanom residuals with respect to different explanatory variables is shown for pixels with high and low tree cover in Fig. . Given the large number of pixels, all correlations are significant.

In DH18, Tanom in spring (Tanomspr, + for high and low tree cover) and summer SManom (SManomsm, - for high tree cover and + for low tree cover) show the strongest relationships with EVIanom residuals. In DH19, EVIanomyr-1 (+), Tanomspr, and Tanomsm (-) show strong correlations, with consistent sign for both high and low tree cover pixels. DH19 residuals of pixels with high tree cover show strong correlation with SManom with opposite signs in spring (+) and summer (-) and with AWC (-). In DH19, pixels with low tree cover show negative correlation between IsoH and EVIanom residuals.

To test whether the importance of EVIanomyr-1 is particular to DH19 or if it reflects long-term inter-annual legacy effects of anomalies in vegetation activity, we fit a second temporal RF model where EVIanomyr-1 is used as an additional predictor (Figs.  and ). Including vegetation condition in the previous summer improves the predictive power of the long-term RF model (72 %–97 % out-of-bag score, compared to 48 %–90 % for the model trained with climate drivers only). Even though the residuals for the training period are considerably reduced relative to the climate-driven model, the residuals for DH18 and DH19 are comparable.

The GPP from the LSM multi-model ensemble mean matches the differences in impacts between clusters in DH18 well (Fig. a and b) and the temporal evolution of GPP anomalies during the 2018 growing season (April to September, Table ), with correlations with EVIanom of 0.74–0.90. Even though the root-mean-squared error (RMSE) is comparable in the two growing seasons, the correlations of GPPanom with growing-season EVIanom are much lower in DH19 (-0.09 to 0.43). GPPanom by LSMs is above average in spring and early summer 2019 for all clusters, and anomalies in DH19 are either more positive or less negative compared to EVIanom.

LSMs simulate a stronger attenuation of drought compared to the observation-based SManom, albeit with consistent relative differences in SManom between clusters (compare Figs.  and ). LSMs simulate the temporal evolution of SManom well in the two growing seasons, with high correlation with both SoMo.ml and SManomERA5 (correlations of 0.81–0.98). The RMSE for simulated SManom is generally lower than that of GPPanom.

The sensitivity of GPPanom to simulated SManom and to Tanom (Fig. ) is consistent with that of EVIanom in all clusters (Fig. ), although for CPRecov and CGreening the LSMs estimate non-significant negative relationships between GPPanom and Tanom. The deviations of GPPanom from the linear response for CHighV and CPRecov in DH18 are correctly captured by LSMs but this is not the case for DH19 in CDecline.

For three clusters covering 56 % of the pixels negatively impacted by DH18, the extremely low EVIanom in response to DH18 and DH19 could not be predicted from EVI–climate relationships in 2001–2017 (Figs.  and ). These departures reveal increased sensitivity to dry and hot conditions and can be a sign of increased ecosystem vulnerability to such events. However, it should be noted that while we focused on pixels that were negatively impacted by DH18, some pixels in the regional domain selected showed greening, even in DH18 (Fig. ). These regional asymmetries result in partial regional compensation of the DH18 impacts, as shown in .

In both DH18 and DH19, higher tree cover fraction is associated with more positive or less negative residuals (Fig. ), indicating that trees were more resistant to DH than grasses and crops. The predominance of crops and grasslands in CHighV, which had strong negative residuals in both events, and of high tree cover in CGreening also support this effect. Trees can better cope with drought with their deeper rooting depth and through the use of carbon reserves to support activity under stress conditions . Moreover, some trees and grasses with stronger stomatal regulation can buffer the drought progression and its impacts by avoiding hydraulic failure . This is reflected in the small but positive relationship between isohydricity and EVIanom residuals in pixels with high tree cover.

Increased vulnerability may be explained by modulating effects of global change on vegetation condition (e.g. “hotter droughts”, , Fig. ) and, in the case of DH19, it may be further linked to inter-annual legacies from the impact of DH18. The first should be expressed by relationships between EVIanom residuals and climatic variables. The latter are more difficult to assess without comprehensive data about different competing factors, e.g. defoliation or damage from embolism , higher susceptibility to diseases and pests due to reduced health or increased hazard of insect disturbances due to warm conditions . The relationships between EVIanom residuals and EVIanomyr-1 provide an approximation but do not allow the identification of the underlying drivers.

In DH18, we find a positive effect of spring warming in vegetation growth, leading to weaker departures from long-term vegetation–climate relationships (observed EVIanom more positive or less negative than modelled), but with associated water depletion amplifying the impacts of DH18 in summer in pixels dominated by grasslands and crops (low tree cover in Fig. 6). These results are in line with that showed contrasting seasonal legacy effects of warm springs in cropland- versus forest-dominated regions.

On the contrary, spring and summer Tanomsm in 2019 (or cooling, see Fig. ) are negative correlated with EVIanom residuals in both high and low tree cover pixels. This indicates increasing damage from heat stress, for example due to reductions in evapotranspirative cooling or cascading impacts of compound heat and drought, such as insect attacks .

Including EVIanomyr-1 in the long-term RF regression model improves the predictive skill for 2001–2017 but does not reduce the residuals in DH18 and DH19. The high correlation between EVIanom residuals and EVIanomyr-1 in DH19 can indicate either that pixels strongly impacted by DH18 were associated with amplified impacts by DH19 (negative residuals) or that pixels affected moderately by DH18 (less negative EVIanomDH18) were associated with positive residuals, i.e. stronger recovery. Damage to roots and tissues or depletion of carbon reserves from DH18 leading to higher vulnerability to DH19 could explain the positive correlation in high tree cover pixels in CDecline. Conversely, the moderate DH18 impacts in CGreening may have resulted in increased resistance to DH19. The strong correlation found in low tree cover pixels is surprising though, as European crop species tend to be annual plants, and annual species can also be found in many grasslands. For these pixels, it is more likely that the positive correlation is explained by management practices, e.g. through earlier harvest or active reduction of stand density in DH19 .

CDecline stands out from the other clusters, in that browning is found in spite of drought alleviation in DH19. The strong negative correlation of residuals with SManomsm and AWC in forest-dominated pixels is counter-intuitive and suggests that other environmental effects not considered in our analysis may modulate DH19 impacts. Insect outbreaks are a potential candidate to explain such effects: the stronger correlation of residuals with EVIanomyr-1 in DH19 could reflect increased susceptibility of impaired trees, combined with favourable climatic conditions for insect growth, reflected in stronger negative effects of Tanomsm in DH19 in high tree cover pixels.

Results from field inventories and forest plots support this hypothesis. Increased tree mortality and insect outbreaks in central Europe during 2018 have been reported . A recent assessment by the German Federal Ministry for Food and Agriculture reported crown damage in 36 % of all tree types in summer 2019, a 7 % increase compared to 2018 and predominating in trees over 60 years of age. According to this report, the mortality rate in both needle-leaved and broad-leaved trees almost tripled from 2018 to 2019. Although no large-scale data on insect outbreaks are currently available, local authorities in regions where CDecline is prevalent report an increase in tree mortality from bark beetle infestations: the Environment Ministry of North Rhine-Westphalia in western Germany reported soaring rates of spruce affected by severe bark beetle infestations, from about 1 % in 2018 to over 12 % in 2019 . In the Czech Republic, rates of spruce damaged by bark beetles more than tripled, leading to increased mortality . In Belgium, a “bark beetle task force” was created in September 2018 by the economic office of Wallonia . Increased tree mortality and bark beetle infestations have also been reported in eastern France .

Temperate ecosystems are an important global sink of CO2 and are not usually considered hotspots of drought risk and environmental degradation under climate change . Our results show that the past two extreme summers in central Europe reveal the first signs of large-scale enhanced vulnerability in response to DH events (CHighV, CPRecov) and of potential degradation trajectories induced by consecutive events (CDecline). Even though it is limited to 20 % of the study area, the patterns in CDecline highlight the risks associated with more frequent and intense droughts and heatwaves expected in the coming decades . At the same time, progressive warming conditions can increase the likelihood of compound occurrence of multiple disturbances, such as droughts and insect outbreaks, which are both promoted by warm and dry conditions. Interactions between compounding disturbances can further contribute to forest C losses . To anticipate such impacts, process-based modelling of ecosystem response to such events is needed.

The LSMs perform well in simulating the magnitude and evolution of productivity anomalies in 2018 but not in 2019. The recovery simulated by LSMs in DH19 can be partly explained by a strong recovery of modelled soil moisture (Fig. ) but may also result from limited ability of LSMs to simulate changes in ecosystem vulnerability during the two DH events. The latter is supported by the fact that simulated SManom shows good agreement in the temporal evolution of soil moisture anomalies with both observation-based datasets but not of GPPanom (Table ).

The comparison of the reference and factorial simulations allows showing that the poor performance in 2019 may be related with interannual legacy effects. LSMs estimate legacies from DH18 only in the early growing season (March to May 2019) but do not estimate any legacy effects in summer (Fig.  bottom panel). The poor relationships between EVIanom and simulated GPPanom in response to DH19 indicate that processes controlling legacy effects such as damage from embolism, carbon starvation, and resulting tree mortality or disturbances induced by drought and heat, such as insect outbreaks, currently missing in LSMs, likely explain the amplified impacts of DH19.

LSMs are known to have limited ability to simulate drought-induced stress and tree mortality and lack impacts of biotic disturbances, although rudimentary approaches have been attempted . These model shortcomings add to limitations in simulating soil moisture variability and transitions between energy-limited and water-limited regimes.