This study makes use of the latest ensemble of high-resolution climate models from CORDEX (Jacob et al. 2014) which uses an ensemble of GCMs from CMIP5 (Taylor et al. 2012). A subset of 11 projections was chosen to represent the spread from the full CORDEX ensemble using cluster analysis (Moss et al. 2010). They are based on five unique GCM/RCM combinations (four GCMs and four RCMs) and three RCPs (RCP2.6, 4.5 and 8.5), and these were combined in different ensembles for the different warming levels. To drive the hydrological models, the climate projections of precipitation and temperature were bias-corrected to the E-OBS (Haylock et al. 2008) gridded observation database using quantile mapping (Wilcke et al. 2013). Other forcing variables were bias-corrected using WFDEI (Weedon et al. 2014).
The method to define warming thresholds in climate models follows that of Vautard et al. (2014) where scenarios that pass the target warming level are used as snapshots in time representing these levels of warming. This is necessary because sufficient ensembles of climate models stabilising at each of these warming thresholds are not available (unless GCM patterns are scaled so as to reach the same warming levels at the same time, as in Heinke et al. 2013). To assess the impacts of climate change at 1.5, 2 and 3 °C GMT rise, change is quantified for the 30-year period centred at the year when each GCM reaches the defined increase in GMT relative to preindustrial levels (1881–1910).
The RCP8.5 runs reach the +1.5 °C threshold very early in the twenty-first century when the uncertainty from the initial state of the climate models is still very high, so the ensemble for +1.5 °C is made up of the lower emission RCPs: RCP2.6 and RCP4.5. For +3 °C, it is mostly only the RCP8.5 runs that reach +3 °C by the end of the century so the ensemble for +3 °C is made up of the RCP8.5 (high-emission) runs (Table 1). For +2 °C, the impacts were calculated twice: (a) for an ensemble of the low-emission RCPs and (b) for an ensemble of the high-emission RCP. This is done so that the sensitivity of the impacts to the choice of RCP forcing the ensemble can be studied.
To account for uncertainties in simulation of hydrological processes, five HMs were systematically used to simulate the impacts of climate change. The chosen HMs differ in complexity of process description, resolution, input data and level of calibration (see S1 for further information on the HMs and how they are parameterised, also Greuell et al. 2015) but were all forced with the same data and produced the same output variables. They include model concepts varying from land-surface schemes (VIC, Liang et al. 1994), to varying levels of process-based hydrological modelling (LISFLOOD, Burek et al. 2013; WBM, Vörösmarty et al. 2000; and E-HYPE, Donnelly et al. 2016) and a coupled water and carbon cycle model with vegetation dynamics (LPJmL, Schaphoff et al. 2013). Methods to simulate evapotranspiration and snow processes vary with each model, e.g. both energy balance and degree day methods are used. Despite these differences, each of the models has demonstrated ability to simulate hydrologic conditions at large scale across Europe (Nijssen et al. 2001; Burek et al. 2013, Donnelly et al. 2016, Biemans et al. 2009) and Greuell et al. (2015) showed that the models showed an ability to respond to differences in wet and dry and warm and cold years (interannual variability). This was postulated as a measure of how HMs might respond to climate change as it was shown that the magnitude of the interannual variability was of a similar order to the expected mean climatic changes (Greuell et al. 2015).
The E-HYPE, Lisflood, WBM and LPJmL models were run on a daily time-step, while VIC was run on a three hourly time-step with results aggregated to daily time-step for the model ensemble. The original HM resolutions vary from 5 km to 0.5° (ca.56 × 56 km at the equator, 28 × 56 km at 60°N) grid resolution, and in the case of E-HYPE a high-resolution subbasin rather than grid scale, so all model results were post-processed to a common 0.5° grid before calculation of ensemble results.
Here, the changes to a number of simple indicators, indicative of the climatic development of aspects of the water cycle relevant for user sectors, were quantified. From this, potential impacts that the different warming levels might have on water-related sectors can be inferred. The following hydrological indicators were calculated for all of Europe:
Evapotranspiration: Mean annual evapotranspiration
Runoff: Mean annual runoff (indicative of available water resources, e.g. for agriculture, water supply, navigation, etc.).
High runoff: Mean annual maximum runoff (indicative of recurring high flows)
Low runoff: Mean annual low runoff (mean of annual 10th percentile runoff, indicative of dry conditions)
Snowpack: Mean annual snow water equivalent (SWE) maximum (indicative of snow storage for hydropower production and tourism)
Runoff, rather than discharge, is analysed due to the different HM resolutions. In total, 55 model runs are analysed, i.e. 11 climate projections and 5 HMs. Changes to the indicators are calculated by taking the mean indicator value over the 30-year periods representing each warming level in each GCM and comparing this to the mean indicator value over the reference period, 1971–2000. So while the warming levels of +1.5, 2 and 3 °C are defined relative to preindustrial levels, the impacts of the change are analysed relative to a more recent historical period. As well as considering changes to runoff and snow across Europe, changes to discharge from six major European rivers were analysed. To do this, the outlet point of each river was identified separately for each HM. These rivers were chosen to demonstrate the needs for adaptation for navigability (e.g. Rhine, Danube, Wisla), irrigation (Ebro), water supply (Ebro, Wisla), ecology (e.g. Ebro delta), hydropower (Glomma, Lule) and cooling water supply (Wisla).
The differences in impacts between warming levels were compared by plotting the ensemble mean changes at one warming level (e.g. 1.5 °C) versus those at another (e.g. 2 °C) for each grid point. These scatter plots were summarised into box-plots by dividing the x-axis into 30 equal size bins and showing the interquartile range in the y-direction for each bin. This was done to help with visual interpretation of the plots (original scatter plots available in online supplementary material). A line-of-best-fit, forced through the origin, was fitted to the median values of each bin; a line steeper than the 1:1 line indicates increasing impacts with increasing warming level. A correlation coefficient, R, is used to indicate the uniformity of the changes (scalability) across the grid; a value of 1 indicates perfect scalability of the changes across the grid.
Assessing reliability of results
Projected changes are assumed robust where the absolute value of the ensemble mean change exceeds the standard deviation (SD) of the changes of the individual members of the ensemble. Robust changes are stippled in the maps. This is consistent with the method used in the Fourth IPCC Assessment Report by Working Group I (Meehl et al. 2007).
As well as the boxed scatter plots to compare warming levels, similar plots were also made for the two different 2 °C ensembles to investigate the sensitivity of the method to the choice of RCP forcing in the climate model ensemble. Finally, the slope of the scaling relationship from the scatter plots is compared for each HM to show the variations in HM sensitivity.