Revisiting the recent European droughts from a long-term perspective

Early 21st-century droughts in Europe have been broadly regarded as exceptionally severe, substantially affecting a wide range of socio-economic sectors. These extreme events were linked mainly to increases in temperature and record-breaking heatwaves that have been influencing Europe since 2000, in combination with a lack of precipitation during the summer months. Drought propagated through all respective compartments of the hydrological cycle, involving low runoff and prolonged soil moisture deficits. What if these recent droughts are not as extreme as previously thought? Using reconstructed droughts over the last 250 years, we show that although the 2003 and 2015 droughts may be regarded as the most extreme droughts driven by precipitation deficits during the vegetation period, their spatial extent and severity at a long-term European scale are less uncommon. This conclusion is evident in our concurrent investigation of three major drought types – meteorological (precipitation), agricultural (soil moisture) and hydrological (grid-scale runoff) droughts. Additionally, unprecedented drying trends for soil moisture and corresponding increases in the frequency of agricultural droughts are also observed, reflecting the recurring periods of high temperatures. Since intense and extended meteorological droughts may reemerge in the future, our study highlights concerns regarding the impacts of such extreme events when combined with persistent decrease in European soil moisture.

1 Evaluation of the model performance 19 The mHM setup follows the previous study of Rakovec et al. 1 , in which the model parameterization is constrained against river 20 discharge and the GRACE satellite-based terrestrial water storage (TWS) 1 , and meteorological forcings are taken from the 21 E-OBS datasets. The model exhibits a reasonable ability to capture the observed dynamics of monthly streamflow (Q), TWS 22 and actual evapotranspiration (ET) during the available period of 1950-2011 (see Fig. S1; panels a, b, and c). The median skill 23 in terms of correlations for standardized Q, TWS and ET estimated across 83 European river basins are 0.86, 0.62 and 0.60, 24 respectively ( Fig. S1; see 1 for more details). We would like to highlight that although we used correlation as an skill metric, it 25 was estimated after standardising the variables of interest. This means that the seasonality component is removed and thus the 26 predictive skill of the model variables can be considered free of its seasonal climate behaviour. 27 We also evaluate the model skill for discharge simulations at 29 gauging stations that have long observation records starting 28 before 1900 (median lengths of 32 years before 1900 and 112 years after 1900) to gain more confidence in the modelling results 29 regarding the historical reconstruction of water fluxes and states over the long period . In this respect, this step also 30 provides an independent evaluation of the reconstructed precipitation and temperature fields 2, 3 . The results of this analysis 31 indicate a reasonably good model capability for capturing the observed dynamics of the standardized monthly discharge over  is considered to understand the recession behaviour of runoff hydrograph and TWS anomalies. Model exhibits slightly higher 53 values than the observations, but overall a general tendency of decreasing correlation with increasing lag time is consistent 54 between model and observations. Since we do not have consistent set of soil moisture observations, we use the terrestrial water 55 storage anomaly observations from GRACE satellite to evaluate the skill of model to capture the auto-correlation behaviour.

56
Note that the TWS has considerably lower sample size due to shorter data availability (monthly samples between 2002-2011).

57
In this case also, the modelled estimates are on average higher compared to the observation -but overall the match between 58 them are reasonably good. We also find a relatively larger spread in both modelled and observed auto-correlation estimates for 59 the TWS anomalies as compared to those obtained for the streamflow values -which do reflect the basin to basin diversity in 60 terms of TWS values. Example time-series of standardized observed and reconstructed routed runoff for the Salzach river at 61 Burghausen (Germany) are given in Fig. S3a. Overall, the mHM simulation agrees well with observations. Fig. S3b compares   62 the distribution of SDI derived from the mHM simulation to those from the observed runoff for the same 29 stations as in Fig.   63 S1. The average and the spread of the simulated SDI values correspond well with the observed SDI values. 64 We show here also the validation results for the groundwater module of the mHM model (Fig. S4). Using the example from 65 South Germany, we demonstrate that even the low complexity groundwater model is able to capture spatial mean of observed 66 groundwater heads reasonably well. Note that the same holds also for capturing the patterns of regional variability (not shown).  (Fig. S7). In all cases, the differences are small.   3 Trend significance 78 We test the significance of trends in various characteristics of precipitation, soil moisture and runoff. The magnitude of the 79 trend is determined by the Theil-Sen slope estimator, i.e., the median of the slope estimates obtained from all distinct pairs of 80 points in the time series 9 . The significance of the trend is evaluated using a modification accounting for auto-correlation 10 . If 81 not stated otherwise, the reported significance refers to the 0.1 significance level.

82
The significance of the trends in the average, 20th percentile (q 20 ) and SDI for standardized precipitation, grid cell runoff 83 and soil moisture is tested at each grid cell. Table S1 reports the fraction of grid cells with significant increasing or decreasing 84 trends over the CEU and MED regions. Fig. S8 shows the localization of the grid cells with significant trends together with 85 their magnitudes. century. This trend is statistically significant for 95% of the MED area and 70% of the CEU area (Table S1).

93
Regardless of the increase in mean precipitation and grid cell runoff in CEU, the lower quantiles of the distribution are 94 decreasing for all variables and regions. The strongest decreasing trend of the 20th percentile (q 20 ) is noticed in soil moisture, 95 followed by precipitation, and the least in grid cell runoff, across both the CEU and MED regions. The trends are significant 96 over more than 90% of the MED area for all variables and more than 30%, 60% and 75% of the CEU area for grid cell runoff, 97 precipitation and soil moisture, respectively (Table S1).

98
In general, the fraction of grid cells with significant trends is larger for MED than CEU, and the trend magnitudes are larger 99 in MED, especially for the case of a decrease in q 20 and an increase in SDI.