Multi-model ensemble projections of European river floods and high flows at 1.5, 2, and 3 degrees global warming

Temperature and precipitation from five GCMs (HadGEM2-ES, IPSL-CM5A-LR, MIROC-ESM-CHEM, GFDL-ESM2M and NorESM1-M) is used to force three HMs at the daily time scale for the period 1950–2099 under three RCPs (2.6, 6.0, and 8.5), which was made available by the ISI-MIP project (Warszawski et al 2014, Hempel et al 2013a). These models have also been widely used for impact studies (Krysanova and Hattermann 2017). This GCM data were downscaled and bias-corrected at a 0.5° global resolution using a trend-preserving approach (Hempel et al 2013b). The 0.5° data is further interpolated within the EDgE project (edge.climate.copernicus.eu) to a 5 km grid over Europe using external drift kriging, which enables HM application at high spatial resolution (Wood et al 2011, Bierkens et al 2015). One variogram for each meteorological variable (i.e. precipitation and temperature) is used for the interpolation, which is derived from daily E-OBS station data (Haylock et al 2008). This approach does not modify long-term trends and is suitable for climate change impact studies.

The period 1971–2000 is selected to represent present-day conditions (1980s in the following) because 1991–2000 is the last decade fully within the historical period of the GCM data. This period is estimated to correspond to a global warming of 0.46 K to pre-industrial conditions in 1881–1910 (Vautard et al 2014). Global warming periods of 30 years for 1.5, 2, and 3 K are estimated with respect to the 1980s employing a time sampling approach (James et al 2017, see supplementary material section S1, available at stacks.iop.org/ERL/13/014003/mmedia). Global warming since preindustrial conditions might be about 0.11 K higher than the 0.46 K assumed here (Hawkins et al 2017), which results in warming periods starting about 2–6 years earlier (not shown). This only has a minor impact on the results because 30 year median values are used in the analysis.

Runoff is simulated by three HMs (mHM, Noah-MP, and PCR-GLOBWB) and routed through the same 5 km river network using the multi-scale routing model that has been developed originally for mHM (Samaniego et al 2010). All models were setup using the same morphologic, land cover, and soil data such that differences among models only originate from different process representations. The mesoscale hydrological model (mHM, www.ufz.de/mhm) is a process-based HM that has been developed for scales from 1 km–50 km (Samaniego et al 2010, Kumar et al 2013). PCR-GLOBWB was developed to represent the terrestrial water cycle, including the human water management, at the global and continental scale with a special emphasis on the groundwater component (van Beek et al 2011, Wanders and Wada 2015). Noah-MP is the land surface component of the Weather Research and Forecast Model representing both the terrestrial water and energy cycle (Niu et al 2011). These three models comprise a wide range of process representations from different conceptualisations of runoff delay using non-linear reservoirs (mHM and PCR-GLOBWB) to numerically computing the Richards equation (Noah-MP). The latter is a physically-based representation for the movement of water in unsaturated soils (Chow et al 1988). Model parameters are calibrated using the E-OBS meteorologic data (Haylock et al 2008) at nine distinct catchments located in Spain, UK, and Norway. Automatic calibration is employed for mHM (following Rakovec et al 2016) and PCR-GLOBWB. Noah-MP is calibrated manually adjusting the parameter for surface evaporation resistance based on the analysis by Cuntz et al (2016).

Using the GCM forcing, all models show a reasonable reproduction of observed indicators at 165 gauging stations (see supplementary section S2). The observed flood indicator is on average underestimated by the multimodel mean by 10% with a standard deviation of 38%. The simulation results for the high flow indicator show a relatively higher bias of 44% for the multimodel mean with a standard deviation of 47%. This is a very rigorous comparison because the high flow and flood indicators obtained from GCM-driven model simulations are directly compared against observations. As a result, higher biases then comparisons using observation-driven HM simulations can be expected. Notably, reducing HM bias through explicit parameter calibration may not reduce the HM uncertainty (e.g. intermodel difference; Mendoza et al 2015).

Two indicators of extreme streamflows are used in this study. These have been co-designed in close collaboration with stakeholders in the water sector across Europe within the EDgE project. The indicators are:

• High flows: the streamflow exceeded 10% of the time (Q₁₀).
• Floods: the annual maximum streamflow (Q_max).

Both indicators are estimated for each year within a given 30 year period. The median of these 30 year values is then considered for the evaluation, which is regarded as a robust estimator for a flow duration curve (Vogel and Fennessey 1994, Blum et al 2017). Furthermore, only river basins with a contributing area larger than 1000 km² are considered in this study to limit the errors due to incorrect delineation of small tributaries and headwater river basins in the routing process.

The impact of climate change is quantified for a given indicator by calculating the relative change between a future 30 year median and the median of the 1980s (1971–2000). The non-parametric Wilcoxon rank-sum test is applied to the two 30 year samples to test the null hypothesis of equal medians, which is frequently used in climate impact studies (Gosling et al 2016). Finally, the percentage of ensemble members indicating a significant difference at a 5% significance level is quantified to assess the robustness of ensemble projections. The results are aggregated to the IPCC AR5 European regions (Kovats et al 2011) based on the stratification presented in Metzger et al (2005) (see supplementary section S3).

The signal to noise ratio is used to quantify the uncertainty of projected changes (Hall et al 2014, Giuntoli et al 2015). The signal is estimated as the ensemble median and the noise as the inter-quartile range of the multimodel ensemble, i.e. the difference between the 25th and the 75th percentile of the ensemble members. The sequential sampling approach by Samaniego et al (2017) is used here to quantify the individual contributions by GCMs and HMs (see supplementary section S4).

Zoom In Zoom Out Reset image size

The magnitude of projected changes for high flows (figures 1(a)–(c)) and floods (figures 1(g)–(i)) amplify with enhanced global warming from 1.5 K–3 K. Strong increases in high flows up to 12% are projected in the Northern region under 3 K global warming (table 1). Conversely, large decreases are projected in the Mediterranean. These are of the order of −10% on average for a 1.5 K global warming and further decrease to −30% for 3 K global warming. Particular hotspots of projected changes in the Mediterranean region are the Iberian Peninsula and the Balkans. The Atlantic and Continental regions can be considered as a transition zone between decreases in Southern and increases in Northern Europe. Projected changes in these regions are generally less than 10% in magnitude (table 1).

On average, projected changes in floods are small (i.e. less than 10% in magnitude) with similar spatial patterns to high flows for all warming levels (figures 1(g)–(i)). It is worth noting that a relatively small change of 15% in annual maximum may lead to substantial changes in flood return periods with strong effects on adaptation planning (Hattermann et al 2016). As for high flows, the strongest decreases of floods are seen in the Mediterranean. The magnitude of changes is, however, reduced to −17% under 3 K global warming. A contrasting pattern to high flows is seen in the Northern region, where floods decrease by up to −5% under 3 K global warming and high flows increase up to 12%. This highlights the importance of considering multiple indicators that are relevant for different stakeholders. A decrease of floods in this region has been observed in several studies (Arheimer and Lindstrom 2015, Alfieri et al 2015, Roudier et al 2016). Increased temperature in snow dominated regions will alter snow dynamics, in particular decreases in snow pack are projected (Roudier et al 2016, Donnelly et al 2017) that will lead to less spring melt and consequently reduce spring floods. In the past decades, earlier snowmelt lead to a shift in timing of floods in this region (Bloeschl et al 2017).

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

The percentage of ensemble members projecting a robust change exhibits a substantial spatial variability for both indicators and all warming levels (figures 1(d)–(f) for high flows and figures 1(j)–(l) for floods). The area of likely changes increases with higher warming levels for both indicators. It is generally higher for high flows than for floods covering up to 2.4 and 1.6 × 10⁶ km², respectively (total area of the study domain is 5.4 × 10⁶ km²). Likely changes are found mainly in regions with strong change signals. A large agreement among models is observed in the Baltic states for both high flows and floods, particularly at 3 K warming. Substantial disagreements are found in the Mediterranean under 3 K global warming, with changes in high flows being significant for more than 66% (likely change) of the ensemble members whereas this number reduces to less than 33% (unlikely) for floods.

Recently, Donnelly et al (2017) reported a stronger increase of high flows with increasing global warming. The herein projected magnitude and direction of changes are substantially different, particularly in the Mediterranean. It is worth noting that the indicators are estimated here for routed streamflow, whereas Donnelly et al (2017) used grid-specific runoff in their analysis. The streamflow of a particular river reach is the result of the grid-specific runoff routed over the entire upstream area, which is more relevant for flood conditions. Notably, the projected changes presented here are confirming those reported by Gosling et al (2016) for the Tagus River who used the same GCM forcing dataset. They are also in line with Alfieri et al (2015) and Rojas et al (2012) who estimated decreases in annual maximum streamflow of −30% by the end of the 21st century on the Iberian Peninsula.

To a large extent, the projected changes in the high flow indicator are related to the changes in annual total catchment precipitation (figures 2(a)–(c)), as highlighted by a coefficient of determination (r² ) larger than 0.5 for all warming levels. Higher r² values are observed for higher warming levels mostly because of the increased spread. There is, however, a considerable difference between positive and negative precipitation changes. Areas exhibiting a decrease in annual total precipitation are located mostly in the Mediterranean, Continental, and Atlantic regions and show higher r² than areas with positive precipitation changes. Annual precipitation is projected to increase mostly in the Northern region (see van Vliet et al 2015, for the same GCMs). An increase in this region has also been reported in other studies (Arheimer and Lindstrom 2015, Alfieri et al 2015, Donnelly et al 2017). In northern regions, high temperatures have a strong impact on snow processes with altered snow accumulation and melting. Snow acts as a buffer that diminishes the relationship between precipitation and high flows resulting in lower r² values compared to areas without a significant snow cover. In contrast, the flood indicator shows a weak correspondence to the changes in annual precipitation (figures 2(d)–(f)) regardless of the region and warming level. Other precipitation statistics, such as maximum five-day precipitation (rx5day) often used as an explanatory variable for changes in floods (Rojas et al 2012), do not exhibit a strong relationship either (figures 2(g)–(i)). One reason might be that GCMs are underestimating extreme precipitation amounts, in particular convective precipitation (Jacob et al 2014). However, recent research using an observation-based datasets indicates that a link between extreme precipitation and extreme streamflow is only found in few cases (Wasko and Sharma 2017). This highlights the complexity of flood events that not only depend on the total amount of precipitation but also on its spatial distribution, time evolution, and antecedent soil moisture conditions (Bloeschl et al 2013). It is worth noting that changes in rx5day are mostly positive regardless of the warming level, which has also been reported in the literature for end of the century projections (Rojas et al 2012, Alfieri et al 2015). To some extent, the Atlantic and Mediterranean regions show higher sensitivity to changes in rx5day, which might be related to the minor importance of snow processes in these regions (table S4: r² values for individual regions). While our analysis indicates that precipitation-based extreme indicators (rx5day) may not be suitable for analysing floods, future research should examine this relationship in greater detail.

A particular focus of this study is to evaluate changes in high flow and flood indicators between 1.5, 2 and 3 K global warming. As expected, larger differences in global temperatures lead to larger differences in high flow and flood indicators across Europe and for all stratified regions (figure 3). The differences in the magnitude of relative change is less than 10% for both indicators between a 1.5 K and 2 K warmer world with robust signals found in at most 8% of the ensemble members because of the large uncertainties of the relative changes at both the 1.5 K and 2 K global warming and the relatively small change in global mean temperatures. Across all regions, higher and more robust changes are observed between all other warming levels.

The impact of climate change does not only depend on the incremental level of warming, but also on the absolute value of global mean temperatures. This can be seen by comparing distributions of relative change when moving from present-day climate (1980s) to a 1.5 K warming and from a 2 K to a 3 K warming. In both cases, the increase of global temperature is in a similar order (about 1 K), but the distribution of differences varies across regions. This is most evident for the Alpine region (figure 3(b)). Projected high flow changes exhibit a bimodal distribution moving from 2 K to 3 K global warming, but a unimodal one from the 1980s to 1.5 K warming. The bimodal distribution is due to a different response in the Alps and the Alpine regions in Norway (see supplementary section S6). Higher temperatures lead to a shorter snow season in both of these mountainous regions, but precipitation is projected only to increase in Norway.

Zoom In Zoom Out Reset image size

Across all regions, the robustness of projected changes is higher between the 1980s and a 1.5 K warming than between a 2 K and 3 K warming (figure 3). In the former, 35%–37% of the ensemble realisations indicate significant changes whereas these are only 14%–25% in the latter (figure 3(a)). This implies that the robustness depends on the absolute global warming level. Moreover, it is not distributed evenly over the continent. For high flows, regions exhibiting changes that are as likely as not (percentage from 33%–66%) are located in the Mediterranean and large parts of Eastern Europe (see figure S5). The differences in the relative changes for both indicators between 1.5 and 3 K global warming and their significance follow the same distributions as those between 2 and 3 K, but are more strongly pronounced because of the larger difference in global mean temperature. The significance is, however, in general smaller than between the 1980s and 1.5 K warming. One reason may be that the inter-annual variability within the 30 year samples is smaller for the 1980s than for the future periods. The variability is pathway dependent in the latter case. For example, a GCM with a global warming of 1.5 K might give a different response for a 30 year period with a strong temperature increase (e.g. under RCP 8.5) compared to a period with a nearly constant warming of 1.5 K (e.g. under RCP 2.6; see also figure S1). This is a limitation of the time sampling approach used here (James et al 2017).

Overall, this analysis highlights that the effort to limit global warming to 1.5 K (UNFCC 2015) would yield a limited benefit for high flow and flood indicators across Europe in comparison to 2 K global warming. The underlying uncertainty in the ensemble members precludes a robust distinction between changes under 1.5 K and 2 K. However, under a 1.5 K global warming, the projected changes are robust for at least 30% of the ensemble members in all stratified European regions (figure 3). These results should not be confused with the fact that a 2 K warming from the 1980s in comparison to 1.5 K from the 1980s shows more substantial changes. The temperature difference in both cases is larger than 0.5 K and the variability of the dataset is too large to identify significant differences given a 0.5 K warming. Adaptation is required for both 1.5 and 2 K warming, but adaptation actions could be similar for both warming levels. These include water resources management, flood protection, and the design of infrastructures.

The signal to noise ratio (SNR), expressed as the ensemble median divided by the ensemble inter-quartile range (Giuntoli et al 2015), is small for changes in floods and high flows under all warming levels in Europe. SNR is less than one in about 70% of the rivers under investigation (figure 4). This implies that there is a substantial uncertainty in the multi-model median presented in figure 1 and it is challenging to derive quantitatively-based adaptation from these results. SNR is on average higher for high flows than for floods (compare figures 4(a)–(c) and figures 4(g)–(i)). This indicates that floods exhibit a higher ensemble variability than high flows. Lower signal to noise ratios are generally found for 1.5 K warming level in comparison to 2 K and 3 K warming because the magnitude of relative changes also increases with increased warming (figure 1) and thus strengthens the signal. The spatial patterns of SNR follow those of the relative changes with hotspots of strong decreasing signals in the Mediterranean for high flows and the Baltic countries for increasing floods. Overall, these results highlight that there is substantial uncertainty associated with the results presented in this study. There is a mixed pattern regarding the uncertainty contribution of GCMs and HMs on changes in high flows (figures 4(d)–(f)) and floods (figures 4(j)–(l)). Overall, in 75% of Europe, uncertainty is dominated by GCMs (relative GCM/HM uncertainty contribution > 1.25) for the high flow indicator irrespective of the warming level. For floods, the relative GCM contribution is significantly lower, dominating around 50% of the area under 1.5 K and 2 K global warming and about 40% for 3 K global warming. This implies that the choice of HM substantially contributes to the uncertainty of the results in 25% up to 60% of the area. A higher spread induced by HMs has also been found in Donnelly et al (2017), although no rigorous uncertainty contribution was conducted in that study. The results of this study also confirm the findings of Dankers et al (2014) who stated that the use of a single hydrologic model underestimates the uncertainty in projected floods in their global scale analysis. It can be expected that GCMs are the main contributor to the high flow uncertainty because this indicator has shown a large dependency on precipitation changes (figure 2). Such a relationship could not be found for the flood indicator, which depends, among others, more strongly on antecedent wetness condition (Bloeschl et al 2013). Hotspots of HM uncertainty are areas with a high contribution of snowmelt to runoff or in semi-arid regions (e.g. Scandinavia, the Iberian Peninsula, and the Balkans). Under dry conditions, the representation of soil moisture dynamics and runoff processes has a large impact on the simulation results. For instance, the fraction of fast surface runoff is relatively higher for Noah-MP than for mHM and PCR-GLOBWB because it does not use linear reservoirs to delay the runoff signal. In regions with a considerable snow cover, different representations of snow processes enhance the variability caused by HMs. Noah-MP solves the full energy balance of snow, whereas mHM and PCR-GLOBWB use a degree-day factor approach. High uncertainty contribution in snow-dominated regions was also reported by Giuntoli et al (2015).

The 45 ensemble members used in this study consider only two major sources of uncertainty (i.e. choice of GCM and HM). This underestimates the true uncertainty of projected changes because of two reasons. First, the used GCMs and HMs only cover a fraction of the possible projections by all GCM/HM combinations (McSweeney and Jones 2016). Second, other factors influencing the uncertainty do exist. These include the choice of downscaling technique, bias-correction method, choice of reference period, internal climate variability, hydrologic model parameters and model resolution (Wilby 2005, Deser et al 2014, Kundzewicz et al 2017). For example, other downscaling techniques such as dynamical and statistical downscaling schemes will provide different representation of extreme precipitation and altered climate change signals (Jacob et al 2014, Donnelly et al 2017, Themeßl et al 2012). These should be recognised in future studies.