<<

Journal of Hydrology (2007) 347, 319– 331

available at www.sciencedirect.com

journal homepage: www.elsevier.com/locate/jhydrol

Evaluation of very high-resolution climate model data for simulating flood hazards in the Upper Basin

Rutger Dankers a,*, Ole Bøssing Christensen b, Luc Feyen a, Milan Kalas a, Ad de Roo a a European Commission DG Joint Research Centre, Institute for Environment and Sustainability, TP261, I-21020, Ispra, Italy b Danish Climate Center/Danish Meteorological Institute, Lyngbyvej 100, DK-2100 Copenhagen Ø, Denmark

Received 8 March 2007; received in revised form 22 June 2007; accepted 11 September 2007

KEYWORDS Summary For the purpose of assessing flood hazard in the Upper Danube Basin in Central Climate change; Europe under current and projected future climate conditions, we evaluated data from a Regional climate recent experiment with the regional climate model HIRHAM at a horizontal resolution of models; approximately 12 km. The climate simulations were used to drive the hydrological model Precipitation; LISFLOOD and the results were compared with observations of precipitation and dis- Floods; charge in the area. To explore the benefits of using these very high-resolution data, we Danube also included the results of two HIRHAM experiments at a lower resolution of 50 km in our comparison. It was found that the 12-km data represent the orographic precipitation patterns and the extreme rainfall events over the Upper Danube Basin better than the low- resolution 50-km data. However, the average precipitation rates are generally higher than observed, while the extreme precipitation levels are mostly underestimated. Using the HIRHAM data as input into the LISFLOOD model resulted in a realistic simulation of the average discharge regime in the Upper Danube. In most the 12-km data also led to a better representation of extreme discharge levels, although the performance was still poor in two relatively small rivers originating in the . At larger spatial scales much of the differences and uncertainties between the high- and low-resolution climate data and the observations are averaged out, resulting in a more or less similar performance of the hydrological model, but at the local and sub-basin scale the 12-km data yield better results. The scenario simulations suggest that future climate changes will have an influ- ence on the discharge regime and may increase the flood hazard in large parts of the Upper Danube Basin. ª 2007 Elsevier B.V. All rights reserved.

* Corresponding author. Tel.: +39 0332 786361; fax: +39 0332 785230. E-mail address: [email protected] (R. Dankers).

0022-1694/$ - see front matter ª 2007 Elsevier B.V. All rights reserved. doi:10.1016/j.jhydrol.2007.09.055 320 R. Dankers et al.

Introduction available in sufficient detail or for a long enough period to be able to derive meaningful statistical relationships. These In the past few decades flood damages in Europe have in- disadvantages can be overcome by dynamical downscaling creased considerably (Munich Re, 2005). In a large part, this using regional climate models (RCMs). upward trend can probably be attributed to human causes In recent years, the horizontal resolution of RCM experi- such as progressing urbanisation in flood . Whether ments has increased considerably and now approaches a le- changes in climate have also played a role is far from cer- vel that allows a realistic simulation of the amount and tain, as huge natural variability and long-term persistence intensity of precipitation at the scale of river basins and make it difficult to discern any trends in extreme weather small catchments. Of the studies mentioned above, only events that are, by definition, rare (Kundzewicz et al., Kay et al. (2006a) used data from an RCM, in their case Had- 2005). In the future the influence of climate change is likely RM3, at a spatial resolution of 25 km, as direct input into a to become more prominent. As a consequence of an in- catchment-based rainfall-runoff model. To take account of crease in greenhouse gas concentrations and the conse- the spatial variability they combined the precipitation fields quent rise in atmospheric temperatures, the hydrological of the RCM with high-resolution, standard average annual cycle will intensify, and extreme precipitation events are rainfall data to calculate the catchment-average rainfall. expected to become more frequent and more intense (Ben- They found that direct use of the RCM data into the hydro- iston et al., 2007; Christensen and Christensen, 2003; logical model resulted in relatively good estimates of flood Semmler and Jacob, 2004). Due to higher winter tempera- frequency (Kay et al., 2006b). Graham et al. (2007) used the tures less precipitation will temporarily be stored as snow, results of seven different RCM simulations, with spatial res- and for a shorter period of time. All this will likely increase olutions ranging from 25 to 50 km, as input into a hydrolog- flood hazard in many areas of Europe, although the risk of ical model of the Lule River Basin in . They snowmelt floods and ice jams in spring may actually be re- compared the delta-change approach of applying only the duced (Kundzewicz et al., 2006). climate change signal on observations with a ‘scaling ap- To date, relatively few studies have appeared that made proach’, i.e. a more direct use of regional climate model a quantitative assessment of the potential impacts of cli- data after correction for systematic bias. They concluded mate change on extreme river flows in Europe. Among these that the largest differences between the two methods oc- studies, there has been a geographical preference for catch- curred in the simulation of extreme runoff events, and that ments located in the UK (e.g. Kay et al., 2006a), the Bene- the results of the scaling approach were more consistent lux countries (e.g. Booij, 2005), (e.g. Shabalova with the changes in extremes in the RCMs themselves (Gra- et al., 2003), and Scandinavia (e.g. Graham et al., 2007). ham et al., 2007). While most studies found an increase in flood frequency In this paper, we present data from a recent experiment and intensity, others found a decreasing trend. Different cli- with the regional climate model HIRHAM at an even higher mate scenarios and hydrological models that have been ap- spatial resolution of approximately 12 km. Our main pur- plied, as well as different ways of linking them make it pose was to evaluate these very high-resolution data for difficult to compare the results and to paint an overall pic- application in impact studies of flood hazard under changing ture at the European scale. So far only Lehner et al. (2006) climatic conditions. To this end, the 12-km HIRHAM model made an integrated, pan-European assessment of the was coupled in an off-line mode to a hydrological model changes in flood frequencies due to global climate change, of the Upper Danube basin in . The perfor- and found northern to north-eastern Europe to be most mance of both the climate model and the hydrological mod- prone to a rise in flood risk. However, their analysis was el was validated against observations of climate and river based on applying the climate change signal of two different discharge, looking specifically at the simulation of hydrolog- general circulation models (GCMs) to an observation-based ical extremes. To evaluate the benefits of using high spatial dataset (i.e. a delta-change approach), and did not take resolution climate information, we compared the results into account a potential increase in climate variability with a second climate model experiment in which HIRHAM (see Lehner et al., 2006). was driven by the same global atmosphere model and green- Due to their coarse horizontal resolution and inability to house gas emission scenario, but run at a lower horizontal resolve important sub-grid scale hydroclimatological pro- resolution of 50 km. The significance of the differences cesses, GCMs are generally unable to capture the extreme in the results due to the increase in resolution was further weather conditions that may cause flooding. For these rea- tested by including a third experiment with the same cli- sons, hydrological impact studies commonly rely on statisti- mate model, also at 50 km but driven by a different scenario cal downscaling techniques (e.g. Wood et al., 2004)or of greenhouse gas emissions. sensitivity studies with hydrological models in which mete- This paper is organised as follows: after introducing orological observations were simply perturbed by some arbi- the Study area and explaining the Methodology the cli- trary amount (e.g. Kwadijk, 1993). However, the implicit matology simulated by HIRHAM in the high- and low-res- assumption that the statistical relationships developed for olution experiments will be validated against the present-day climate also hold under the different forc- observations, and the differences in the climate change ing conditions of possible future climates cannot be veri- signal between the three RCM runs will be discussed. fied. Empirical techniques cannot account for systematic Next, the performance of the hydrological model, when changes in regional forcing conditions or feedback pro- driven by the HIRHAM data, will be evaluated and the cesses. Furthermore, in remote regions or areas with com- differences in climate change impacts on simulated plex topography observational data are mostly not flood risks will be assessed. Evaluation of very high-resolution climate model data for simulating flood hazards in the Upper Danube Basin 321

Study area a grid size of 12 km and covering the whole of Europe. The HIRHAM experiment consisted of simulations for two The catchment of the Danube River is the second largest in 30-year time slices: a 30-year control run with a greenhouse Europe, covering 18 countries. The river itself fulfils an gas forcing corresponding to 1961–1990, and a scenario run important function in navigation, water supply, and hydro- corresponding to 2071–2100 according to the A2 scenario of electricity. The upper part of the Danube Basin is character- the Intergovernmental Panel on Climate Change (IPCC) ised by strong topographical gradients, with several rivers (Nakicenovic and Swart, 2000). These runs are hereafter re- originating in the Alps, the Bavarian Forest and in the Mor- ferred to as ‘H12CL’ and ‘H12A2’, respectively. In the con- ava Basin, which makes it suitable for evaluating resolution trol run, the lateral boundaries were derived from the effects. For the present study, we chose the area upstream HadAM3H high-resolution global atmosphere model, which of (), which covers approximately itself had been forced by low-resolution observed sea sur- 130,000 km2 (see Fig. 1). face temperature (SST) and sea-ice extent. The climate change signal in SST and sea-ice extent for future conditions Methodology came from the global coupled atmosphere-ocean model HadCM3 (Gordon et al., 2000; Pope et al., 2000). The models used in this study are the regional climate mod- In this paper we compare the results of the 12-km exper- el HIRHAM (Christensen et al., 1996) and the hydrological iment with three simulations at a lower resolution of about model LISFLOOD (De Roo et al., 2000) which has been devel- 50 km: a 30-year control run (‘H50CL’) and two scenario oped for operational flood forecasting on a European scale. runs with greenhouse gas forcing according to the IPCC sce- narios A2 (‘H50A2’) and B2 (‘H50B2’), respectively. The The regional climate model HIRHAM boundary conditions for these simulations were also coming from the HadAM3H/HadCM3 global models, allowing us to evaluate the influence of choosing a finer horizontal resolu- The HIRHAM regional climate model is based on the limited tion on the simulation results. An overview of the climate area model HIRLAM but uses the physical parameterisation model runs used in this study is given in Table 1. of the general circulation model ECHAM (Roeckner et al., 1996). Here we use data from a recent experiment that has been conducted within the framework of the EC-funded The hydrological model LISFLOOD project PRUDENCE (Prediction of Regional scenarios and Uncertainties for Defining EuropeaN Climate change risks The LISFLOOD model is a partially physically based rainfall- and Effects; Christensen et al., 2002). In this experiment a runoff model that has been developed to simulate the very high horizontal resolution was adopted, approaching hydrological behaviour in large European catchments with

Figure 1 Map of the Upper Danube Basin, showing the location of the gauging stations mentioned in the text. 322 R. Dankers et al.

late potential evapotranspiration were obtained from the Table 1 Overview of the HIRHAM regional climate model Meteorological Archiving and Retrieving System (MARS) runs used in this study Meteorological Database (Rijks et al., 1998) that uses a Run Horizontal Greenhouse gas Corresponding much less dense network of synoptic weather stations. resolution (km) forcing period The optimal parameterisation of LISFLOOD was determined H12CL 12 Control run 1961–1990 with a hybrid optimisation procedure consisting of an adap- H12A2 12 Scenario A2 2071–2100 tive partition-based search and a downhill simplex method H50CL 50 Control run 1961–1990 (Szabo´, 2006). The resulting simulated river discharge of H50A2 50 Scenario A2 2071–2100 the Danube at Bratislava in the validation period October H50B2 50 Scenario B2 2071–2100 1998–September 2002 (Fig. 2) reproduces the observations satisfactorily (Nash and Sutcliffe (1970) coefficient = 0.6).

Approach emphasis on predicting floods and droughts (De Roo et al., 2000). The model is a combination of a grid-based water To simulate flood risks in the Upper Danube under changing balance model and a one-dimensional hydrodynamic chan- climate conditions, the LISFLOOD model was coupled to nel flow routing model. Since it is spatially distributed, HIRHAM in an off-line mode. In total, five model runs were the model can take account of the spatial variation in land done with LISFLOOD: two driven by the high-resolution 12- use, soil properties and precipitation. Because of its general km HIRHAM data (H12CL and H12A2) and three driven by nature, LISFLOOD is optimally suited for simulating the dif- the lower resolution 50-km data (H50CL, H50A2 and ferent hydrological regimes across Europe. H50B2). In this way, we could compare the control and sce- In the present study, LISFLOOD was set up to simulate nario periods with each other, as well as the high and low- the discharge of the Upper Danube at a spatial resolution resolution climate data (H12CL vs. H50CL, and H12A2 vs. of 1 km. Wherever possible the input parameters and vari- H50A2). Furthermore, the differences in the climate change ables were derived from European databases. Soil physical signal between the H12A2 and H50A2 runs could be com- properties were derived from the European Soil Geographi- pared to the H50B2 run to assess whether changing the res- cal Database (King et al., 1994), while the HYPRES database olution of the RCM has a significant impact on the results (Wo¨sten et al., 1999) was used to estimate porosity, satu- compared to adopting a different scenario of greenhouse rated hydraulic conductivity and moisture retention proper- gas emissions. ties for each texture class. Vegetation and land use For the control period, the performance of the H12CL information were obtained from the CORINE 2000 land cover and H50CL runs was tested by comparing the output fields dataset of the European Environment Agency (EEA, 2004), with observation-based datasets. For the purpose of cali- and digital elevation data from the Catchment Information brating and validating the LISFLOOD model, we collected a System (Hiederer and De Roo, 2003). In the current version high-resolution meteorological database based on observa- of LISFLOOD, 11 parameters need to be estimated by cali- tions at 2688 stations in the Upper Danube. The entire per- brating the model against historical records of river dis- iod covered by this database (January 1994–October 2002) charge. These include parameters that control infiltration, is, however, too short to obtain a proper statistical descrip- snowmelt, overland and river flow, as well as residence tion of the climatic conditions in the area. For this purpose times in the soil and subsurface reservoirs. For this purpose we also compared the HIRHAM output fields with data from the model was driven by observations of precipitation and the MARS database that covers a longer period (January temperature from the period October 1994–September 1990–December 2005) but contains the data from synoptic 1997 at 2688 meteorological stations in the Upper Danube weather stations only. We also used data from the high-res- Basin and compared with observed daily discharges at 12 olution precipitation climatology database for the Alpine re- locations. The meteorological variables required to calcu- gion (version 4.0) compiled by Frei and Scha¨r (1998) for the

Figure 2 LISFLOOD simulation of the Danube River discharge at Bratislava in the validation period October 1998–September 2002. Evaluation of very high-resolution climate model data for simulating flood hazards in the Upper Danube Basin 323

25-year period 1971–1995, that has a grid resolution of precipitation along the north side of the Alps and, to a lesser about 25 km. Note that none of these observation-based extent, along the south-west side of the Bavarian Forest. datasets actually has the desired length of 30 years to ob- Note that in the MARS data, based on fewer synoptic weath- tain a correct climatology, and that only the Frei and Scha¨r er stations, this pattern is much less pronounced and the (1998) database overlaps with the HIRHAM control runs that precipitation amounts are lower. The two HIRHAM experi- provide a simulation of the climatological conditions (not ments reproduce this general pattern fairly well. The the historical weather) for a 30-year period corresponding H12CL run is able to simulate the high precipitation amounts to 1961–1990. over the main Alpine range (on the southern boundary of the To drive the LISFLOOD model, the HIRHAM simulations of Danube catchment), while the H50CL run shows a smaller temperature, precipitation, solar and thermal radiation, gradient but extends the orographic influence to the north humidity and wind speed were distributed over the 1-km (Fig. 3). On average, the annual precipitation is higher in grid of the Upper Danube Basin used by the hydrological the two HIRHAM experiments than in the observation-based model. For the purpose of the present study no further datasets (Table 2). In the H12CL run, this overestimation of downscaling was applied to the climate data, i.e. no bias the observed precipitation is most prominent in winter and correction and no correction for altitude. Instead, the HIR- spring over the main Alps and the Basin, and least in HAM output fields were interpolated from the centre points autumn. However, the observation-based datasets have not of the climate model grid to the hydrological model grid been corrected for the systematic bias inherent to rain cells by using an inverse distance interpolation scheme to gauge measurements. This undercatch is particularly large smooth out the patterns. The radiation, humidity and wind in case of snow measurements (Goodison et al., 1998) and speed data were used to calculate reference evapotranspi- may explain at least part of the differences between ob- ration based on the Penman–Monteith model (Monteith, served and simulated precipitation in winter and spring. 1965). LISFLOOD then calculates actual evaporation and Fig. 3 also shows that the annual precipitation is generally transpiration rates based on vegetation characteristics, leaf higher in the H12CL run than in the H50CL data. Averaged area index and soil properties. Since it is a distributed mod- over the catchment area the difference is in the order of el, it produces runoff for every grid cell, which is then rou- 8%, indicating that, at the scale of a river basin, the change ted through the river network using a kinematic wave in resolution of the climate model is not just a mere redis- approach. Here, we compared the simulated river discharge tribution of rainfall but affects the total amount of precip- with observations for the Danube at Bratislava and a number itation over the area. of sub-basins (see Fig. 1). As an indicator of extreme precipitation levels, Fig. 4 shows the annual maximum amount of precipitation falling in five consecutive days (PX5d). The two high-resolution, Comparison of HIRHAM with meteorological observation-based datasets show a pattern similar to that observations of the annual precipitation, which is reasonably well repro- duced by the H12CL data. In the MARS and H50CL data, how- As a first step, we compared the HIRHAM climatology of the ever, the spatial patterns seen in the high-resolution datasets H12CL and H50CL runs with the three observation-based are virtually absent. And while the average precipitation datasets outlined above. Fig. 3 shows that the spatial distri- rates are generally higher in the H12CL data than in the bution of the annual precipitation is fairly well reproduced high-resolution station observations, the maximum 5-day by the 12-km run and, to a lesser extent, also by the 50- rainfall amount is mostly lower. This underestimation of km data. The high-resolution station observations and the the extreme precipitation is strongest in summer (on average Frei and Scha¨r (1998) data show a distinct pattern of higher by about 20%, but with large differences from place to place),

Figure 3 Mean annual precipitation in the Upper Danube Basin according to the different datasets used in this study: (a) high- resolution station observations; (b) MARS; (c) Frei and Scha¨r (1998); (d) HIRHAM, H12CL experiment and (e) HIRHAM, H50CL experiment. Note the different periods to which these averages apply. 324 R. Dankers et al.

Table 2 Statistics of mean annual precipitation in the Upper Danube Basin Dataset Precipitation (mm) Temperature (C) Area average Area maximum Area minimum Standard deviation H12CL 1109 2913 398 371 7.1 H50CL 1065 2003 304 341 7.2 Stations 975 2468 487 347 7.9a MARS 741 1149 408 191 7.8a Frei and Scha¨r 890 1884 463 314 a With correction for elevation.

while in winter the PX5d is generally overestimated by 24%, ments are thus distinctly different from the H50B2 run, but though also in this case the winter precipitation measure- overlap to a large extent with each other. Nevertheless, the ments are likely subject to large uncertainties. In the H12A2 experiment shows a more differentiated pattern with H50CL run the PX5d is mostly lower than in the 12-km data, a larger increase in temperatures over the Alpine region meaning that the underestimation in summer is even more se- than the H50A2 run but a more moderate change in the low- vere (33% on average). In winter, the 50-km PX5d is on aver- land regions. This orographic gradient in the temperature age 12% higher than in the high-resolution station data, but change is most prominent in spring and is, to a lesser extent, even in this season there is a considerable underestimation also visible in autumn, suggesting that it is related to a on the north slope of the Alps. A similar picture was seen in shorter snow season. In all the three scenarios, the warming other indicators of heavy precipitation like the average num- is strongest in summer and autumn (3.7 and 4.3 C in the ber of days on which the daily rainfall exceeds 20 mm, which two A2 scenarios, 2.9 C in H50B2) and least in spring (2.5, has been defined as ‘very heavy precipitation days’ (e.g. Klein 2.7 and 1.3 C, respectively). Tank and Ko¨nnen, 2003). In the H50CL data the frequency of The simulated changes in annual precipitation are pre- heavy precipitation days is seriously underestimated. The sented in Fig. 5. On average, the annual precipitation in- H12CL run reproduces the observed spatial patterns to some creases by 3% in the H12A2 run, 2% in the H50A2 run and extent, but also in this experiment the frequency of heavy 4% in the H50B2 run. Both 50-km experiments show the larg- precipitation days is in places lower than observed. est increases of up to 15% over the Morava Basin in the east. Over the Alps the H50B2 scenario predicts a small increase, while the H50A2 scenario shows hardly any change. The 12- Changes in the HIRHAM scenario runs km data, on the other hand, shows a small decrease in an- nual precipitation over the Alpine region, but an increase The three scenario runs show a significant rise in the mean over the western part of the Upper Danube Basin, which is annual temperature over the Upper Danube Basin. Averaged not present in the 50-km runs (Fig. 5). Although the annual over the study area, it amounts to +2.29 C in the H50B2 changes are relatively small, the differences in seasonal run, +3.73 C in the H50A2 run, and +3.48 C in the H12A2 means are larger. All three scenarios show an increase in run. In spite of the different resolution, the two A2 experi- precipitation in winter and spring (of up to 23% in Decem-

Figure 4 Average annual maximum precipitation amount falling in 5 consecutive days, (a) high-resolution station observations; (b) MARS; (c) Frei and Scha¨r (1998); (d) HIRHAM, H12CL experiment and (e) HIRHAM, H50CL experiment. Evaluation of very high-resolution climate model data for simulating flood hazards in the Upper Danube Basin 325 ber–January), but a decrease in summer and autumn (up to 13% in June–August). Considerable changes are predicted in the precipitation extremes, for example an increase of up to 60% in the an- nual PX5d (Fig. 6). Again, the largest increases are simu- lated in the eastern part of the catchment, particularly in the 50-km experiments. Both low-resolution runs show lar- gely the same pattern, although in the H50A2 scenario the increase in PX5d extends to a larger portion of the catch- ment. Note also that in some parts of the Alpine region the H50A2 run shows an increase in the annual PX5d where the H12A2 actually predicts a slight decrease. Interestingly, the H50A2 scenario simulates a larger increase in the maxi- mum 5-day amount in summer (June–August, when convec- tive precipitation is more prominent) than the H12A2 scenario (on average +14% and +4%, respectively), and over a larger area. In both A2 scenarios, however, the largest rel- ative increases occur in spring and very little change (+1%) Figure 6 Change in the average annual maximum 5-day in autumn; the H50B2 scenario, on the other hand, shows precipitation amount in the three HIRHAM scenarios relative to a more uniform increase in the maximum 5-day precipita- their corresponding control runs: (a) 12-km, A2; (b) 50-km, A2 tion over all the seasons, ranging from 8% to 14%. In a recent and (c) 50-km, B2. analysis of a large number of RCM scenarios, Beniston et al. (2007) also found an – sometimes strong – increase in heavy precipitation over Central Europe. Averaged over the re- LISFLOOD simulation of discharge gion, they found increases in the winter 5-day maximum precipitation of up to 30%. In summer, the maximum 1- When driven directly by the output of the HIRHAM control day rainfall amount was found to increase up to 20% or re- runs, LISFLOOD reproduces the observed discharge regimes main more or less unaltered, except in scenarios that also in the Danube Basin fairly well. Fig. 7 shows the observed showed a strong reduction in the mean summer precipita- and simulated discharge regime for the locations shown in tion (Beniston et al., 2007). For stations across Europe, Fig. 1 and Table 3. At the basin outlet Bratislava, the sim- Klein Tank and Ko¨nnen (2003) already found an average in- ulated river discharge corresponds well with the observa- crease in the PX5d of 0.6 (0.0–1.2) mm per decade over the tions for most of the year, with the exception of late period 1946–1999. However, the spatial coherence of the spring and early summer when the simulated discharge in trends was low and opposite trends were found at relatively the H12CL and H50CL runs is too high. In Comparison of short distances (Klein Tank and Ko¨nnen, 2003). HIRHAM with meteorological observations, it was already In short, all the three scenarios predict a warmer and noted that the HIRHAM precipitation is on average higher wetter winter and spring and a warmer and drier summer than observed, particularly in winter and spring, which in and autumn, and in many places also an increase in the ex- the mountains may accumulate during the snow season treme precipitation levels over all seasons except autumn in and results in higher simulated runoff following snowmelt. the two A2 scenarios. A similar feature can also be seen for the River at Scha¨rding (Fig. 7d) that drains a large part of the Alpine region. Likewise, the overestimation of river discharge throughout much of autumn, winter and spring in the Mor- ava Basin (Fig. 7b) can be attributed to an overestimation of the precipitation in this area during these seasons. Gen- erally speaking, the differences in runoff between the H12CL and H50CL experiments are small, except for the most upstream part of the Danube River at Berg (Fig. 7f) where the H12CL data result in a significant improvement of the discharge simulation. To analyse how well LISFLOOD is simulating extreme events, we compared the probability of extreme discharge levels in the observed and simulated time series. To this end, a generalised extreme value (GEV) distribution (Coles, 2001; Katz et al., 2002; Gilleland and Katz, 2005) was fit- ted to the annual maximum discharges at the gauging sta- tions shown in Fig. 1. The resulting return level plots of the observations and the H12CL and H50CL model runs are shown in Fig. 8. When looking at the plots it should Figure 5 Change in mean annual precipitation in the three be kept in mind that the 30 years of observations corre- HIRHAM scenarios relative to their corresponding control runs: sponding to 1961–1990 were not available at all stations (a) 12-km, A2; (b) 50-km, A2 and (c) 50-km, B2. (see Table 3). 326 R. Dankers et al.

Figure 7 Simulated and observed discharge regime in the Upper Danube Basin in the control period 1961–1990, averaged over 30 years where possible: (a) Danube at Bratislava; (b) Morava at Angern; (c) Enns at Steyr (period of observation 1971–2000); (d) Inn at Scha¨rding; (e) at (period of observation 1990–2002) and (f) Danube at Berg (period of observation 1990–2002). For locations, see Fig. 1.

Table 3 Observed and simulated river discharge at selected gauging stations in the Upper Danube Basin River/station Area (km2) Observed (m3/s) H12CL (m3/s) H50CL (m3/s) Danube/Bratislavaa 131244 2063 2334 2399 Morava/Angerna 25624 111 154 158 Enns/Steyrb 5915 202 164 169 Inn/Scha¨rdinga 25664 719 887 828 Lech/Augsburgc 3803 117 117 133 Danube/Bergc 4047 41 44 94 For locations, see Fig. 1. a Period of observation 1961–1990. b Period of observation 1971–2000. c Period of observation 1990–2002. Evaluation of very high-resolution climate model data for simulating flood hazards in the Upper Danube Basin 327

With the exception of two gauging stations (three for the fit, while a 30-year event at Angern has an estimated return H50CL simulation), the simulated return levels in Fig. 8 period of 33 years in the H12CL run but of only 11 years in show a reasonable agreement with the observation-based the H50CL run. estimates. This is particularly the case at the basin outlet, Fig. 8 also shows a significant underestimation of the ex- Bratislava. Contrary to the discharge regime shown in treme discharge levels at Steyr and Augsburg, while at Berg Fig. 7a the simulated extreme discharge levels are mostly the H12CL simulation performs very well but the H50CL run slightly lower than the observations, and especially in the overestimates the peak flows considerably (this in line with H12CL run this results in lower estimates of events with a the general overestimation of discharge shown in Fig. 7f). return period in the range of 50–100 years. Interestingly, The reason for the underestimation of peak flows at Steyr at Bratislava the H12CL and H50CL simulations are quite and Augsburg is not exactly clear. Both the Enns and the close to each other, while for the Morava the H50CL run re- Lech are, however, relatively small (<6000 km2) river basins sults in higher, and for the Inn River in lower estimates of originating in the Alpine region (also the Inn River originates extreme discharge levels. In the latter basin, the H12CL in the Alps but has a much larger catchment area, see Table experiment seems to overestimate the more frequent peak 3). The underestimation could therefore be due to the rep- flows, but results in a better estimate of the most extreme resentation of small, mountainous catchments with short events (Fig. 8d). Note that at Scha¨rding the 50-year return response times in LISFLOOD, or due to the simulation of level in the H50CL simulation is only a 6-year event in the snowmelt-generated floods. Also, both rivers are heavily H12CL run and an 11-year event in the observation-based regulated, which may have affected the model calibration.

Figure 8 Return level plots of simulated and observed extreme discharge levels in the Upper Danube Basin, based on a generalised extreme value (GEV) distribution fit to the annual maxima. Note the logarithmical scale on the horizontal axis that extends to 100 years. The symbols indicate individual annual maximum values. For period of observation, see Fig. 7. 328 R. Dankers et al.

From Fig. 4 it can be deduced that the H50CL and, to a les- and can be explained by less precipitation in summer, more ser extent, the H12CL experiments underestimate the ex- precipitation in winter and spring, and higher temperatures treme precipitation levels on the north face of the Alps, resulting in an earlier snowmelt (see Changes in the HIRHAM where both the Enns and the Lech originate. In summary, scenario runs). Regimes that are dominated by snowmelt, particularly the H12CL simulation shows a relatively good like those of the Enns, Inn and Lech rivers, also show a shift representation of extreme discharge levels, except in the and reduction in the runoff peak in late spring, which is Enns and Lech river basins. However, deviations from the least obvious in the H50B2 scenario that is characterised observations and differences between the two model runs by less warming than the two A2 scenarios. become larger with increasing return periods and especially Fig. 10 shows the return level plots after fitting a GEV beyond the 50-to-100-year return interval. distribution (Coles, 2001; Gilleland and Katz, 2005) to the time series of annual maximum discharges in the three sce- Changes in flood hazard nario runs. Compared with the two control runs, a tendency towards higher extreme discharges in the scenarios can be The changes in simulated river discharge under the three seen at most stations, with the exception of Scha¨rding and climate scenarios (H12A2, H50A2 and H50B2) are summa- Augsburg, where the H12A2 return levels are lower than in rised in Figs. 9 and 10. A general pattern of change in the the H12CL control run. The reason for this is the slight de- discharge regime can be seen towards higher runoff in win- crease in extreme precipitation levels over the Alpine re- ter and spring, and lower runoff in summer. This pattern can gion that can be seen in Fig. 6a. The two 50-km scenarios, be seen at all stations and in each of the three scenarios, on the other hand, show a slight increase in extreme precip-

Figure 9 Change in discharge regime in the Upper Danube in the three HIRHAM scenarios (2071–2100; dashed lines) together with their corresponding control runs (solid lines). Evaluation of very high-resolution climate model data for simulating flood hazards in the Upper Danube Basin 329

Figure 10 Return level plots of extreme discharges in the three HIRHAM scenarios (dashed lines) together with their corresponding control runs (solid lines); the symbols indicate individual annual maxima. Note that in some plots the vertical scale is different from Fig. 8. itation levels over the same area (Fig. 6b and c). This is re- most other stations the control and scenario runs are very flected in the extreme discharges that are also higher in the close to each other and only start to diverge for return peri- H50A2 and H50B2 runs than in the H50CL experiment, ods greater than 5 or 10 years. Particularly strong increases although the increase is relatively small at Scha¨rding (and in the most extreme discharge levels (return period of even negative for short return periods; Fig. 10d). 30 years or more) of up to 50% or more can be observed at At all other stations the extreme discharge levels are Angern and Berg, but it should be noted that here the GEV higher in the three scenario runs than in the corresponding fit is heavily influenced by just one or two events (Fig. 10b control experiments, especially for events with a longer re- and f). In the three river basins that originate in the Alps, turn period. A relatively strong increase can be seen at the the Enns, Inn and Lech, the relative increase in the extreme catchment outlet, Bratislava, where in the H12A2 scenario discharge levels is higher in the two 50-km scenarios than in a 50-year event has become a 9-year event or, put differ- the H12A2 experiment (that, as noted before, even shows a ently, is 35 % higher than in the control run. In the H50A2 decrease at Scha¨rding and Augsburg), while at the other and H50B2 scenarios the increases for a 50-year event are three stations the 12-km scenario shows a stronger response. less dramatic (18% and 6%, respectively), but more signifi- cant at shorter return periods: the return level of a 10-year Discussion and conclusions event, for example, is 22% higher in the H50A2 experiment and 14% higher in the H50B2 scenario, compared with +20% In their ‘Guidelines for Use of Climate Scenarios Developed in the H12A2 run. The Danube at Bratislava also shows an in- from Regional Climate Model Experiments’ Mearns et al. crease in discharge level at short return periods, while at (2003) recommend to use observed data to drive impact 330 R. Dankers et al. models whenever possible. However, high-resolution, Acknowledgement observation-based meteorological datasets for a long en- ough period of time (30 years or more) and with sufficient Access to the high-resolution precipitation climatology quality to build a proper statistical description of the cli- database for the Alpine region (version 4.0) of Frei and mate are hard to obtain, even for densely monitored regions Scha¨r (1998) was kindly provided by the Data Centre of such as the Alps. In the present study we found that very the Mesoscale Alpine Programme (MAP) in Zu¨rich, high-resolution RCM simulations may provide a valuable, . though not perfect, alternative. Compared with the observations, the 12-km HIRHAM data References that were used here represented the orographic precipita- tion patterns and the extreme rainfall events over the Beniston, M., Stephenson, D.B., Christensen, O.B., Ferro, C.A.T., Upper Danube Basin better than the low-resolution, 50-km Frei, C., Goyette, S., Halsnaes, K., Holt, T., Jylha¨, K., Koffi, B., data from the same model setup. Using the same high-reso- Palutikof, J., Scho¨ll, R., Semmler, T., Woth, K., 2007. Future lution dataset to drive the LISFLOOD hydrological model re- extreme events in european climate: an exploration of regional sulted (at most gauging stations) in a realistic simulation of climate model projections. Climatic Change 81, 71–95. the average discharge regime in the Upper Danube. How- Booij, M.J., 2005. Impact of climate change on river flooding ever, with the exception of one station (the Danube River assessed with different spatial model resolutions. Journal of Hydrology 303, 176–198. at Berg) the results of the H50CL model run are, generally Christensen, O.B., Christensen, J.H., 2003. Severe summertime speaking, not much better or worse than those of the flooding in Europe. Nature 421, 805–806. H12CL run. In some rivers (namely the Morava, Inn, and Christensen, J.H., Christensen, O.B., Lopez, P., Van Meijgaard, E., the Danube upstream of Berg) the 12-km data lead to a Botzet, M., 1996. The HIRHAM4 Regional Atmospheric Climate (much) better representation of extreme discharge levels. Model. DMI Scientific Report 96-4, Danish Meteorological Insti- In the Lech and the Enns, both relatively small rivers origi- tute, Copenhagen, Denmark. nating on the northern side of the Alpine mountain range, Christensen, J.H., Carter, T., Giorgi, F., 2002. PRUDENCE the simulation of extreme events is, however, still poor. employs new methods to assess european climate change. For catchments of this size that are characterised by a large EOS 83, 147. orographic variability, the models are still not of sufficient Coles, S., 2001. An Introduction to Statistical Modeling of Extreme Values. Springer-Verlag, London. quality. At larger spatial scales it seems that much of the De Roo, A.P.J., Wesseling, C.G., Van Deurzen, W.P.A., 2000. differences and uncertainties between the high- and low- Physically based river basin modelling within a GIS: the resolution climate data and the observations are averaged LISFLOOD model. Hydrological Processes 14, 1981–1992. out, resulting in a more or less similar performance of the European Environment Agency (EEA), 2004. Corine Land Cover 2000 hydrological model that is increasingly matching better with (CLC2000). . the observations as the catchment size increases. However, Frei, C., Scha¨r, C., 1998. A precipitation climatology of the Alps it is often not just the response at the outlet that is impor- from high-resolution rain-gauge observations. International tant but also the simulation of extreme events within the Journal of Climatology 18, 873–900. catchment, and it is exactly at this local and sub-basin scale Gilleland, E., Katz, R.W., 2005. Extremes Toolkit (extRemes): that the high-resolution RCM data give better results. Weather and Climate Applications of Extreme Value Statistics. . The present study also found that the changes in dis- Goodison, B.E., Louie, P.Y.T., Yang, D., 1998. WMO Solid Precip- charge regime were to a large extent consistent between itation Measurement Intercomparison: Final Report. WMO/TD – the three scenarios considered – the high-resolution A2 No. 872, World Meteorological Organization, Geneva, and the low-resolution A2 and B2 experiments. This should Switzerland. not be surprising as the changes in the seasonal precipita- Gordon, C., Cooper, C., Senior, C.A., Banks, H.T., Gregory, J.M., tion were also consistent between the three scenarios. Johns, T.C., Mitchell, J.F.B., Wood, R.A., 2000. The simulation The only significant difference was found in rivers that are of SST, sea ice extents and ocean heat transports in a version of dominated by snowmelt, where the higher temperatures the Hadley Centre coupled model without flux adjustments. in the two A2 scenarios resulted in a larger reduction and Climate Dynamics 16, 147–168. shift of the spring runoff peak. Also the changes in extreme Graham, L.P., Andre´asson, J., Carlsson, B., 2007. Assessing climate change impacts on hydrology from an ensemble of regional flood levels were more or less consistent between the three climate models, model scales and linking methods – a case study scenarios, with the exception of the Lech and Inn rivers in on the Lule River Basin. Climatic Change 81, 293–307. the H12A2 scenario. The two 50-km scenarios resulted in a Hiederer, R., De Roo, A.P.J., 2003. A European Flow Network and larger increase over the Alpine region, while the 12-km sce- Catchment Data Set. Technical Report EUR 20703/EN, European nario showed a stronger response over the lowland areas. Commission Joint Research Centre, Ispra, Italy. These patterns can be explained by looking at the changes Katz, R.W., Parlange, M.B., Naveau, P., 2002. Statistics of extremes in extreme precipitation in the three scenarios. in hydrology. Advances in Water Resources 25, 1287–1304. The present simulations using state-of-the-art models Kay, A.L., Jones, R.G., Reynard, N.S., 2006a. RCM rainfall for UK suggest that future climate changes will have an influence flood frequency estimation. II. Climate change results. Journal on the discharge regime and flood hazard in the Upper Dan- of Hydrology 318, 163–172. Kay, A.L., Jones, R.G., Reynard, N.S., 2006b. RCM rainfall for UK ube. Not everywhere this will result in an increase in flood flood frequency estimation. I. Method and validation. Journal of hazard – in fact we found a decrease in extreme discharges Hydrology 318, 151–162. at two stations in the H12A2 scenario – but an increase in King, D., Daroussin, J., Tavernier, R., 1994. Development of a soil the extreme river discharge levels seems likely in most parts geographical database from the soil map of the European of the Upper Danube Basin. Communities. Catena 21, 37–56. Evaluation of very high-resolution climate model data for simulating flood hazards in the Upper Danube Basin 331

Klein Tank, A.M.G., Ko¨nnen, G.P., 2003. Trends in indices of daily Nash, J.E., Sutcliffe, J.V., 1970. River flow forecasting through temperature and precipitation extremes in Europe, 1946-99. conceptual models part I – a discussion of principles. Journal of Journal of Climate 16, 3665–3680. Hydrology 10 (3), 282–290. Kundzewicz, Z.W., Graczyk, D., Maurer, T., Pin´skwar, I., Radzie- Pope, V.D., Gallani, M.L., Rowntree, P.R., Stratton, R.A., 2000. The jewski, M., Svensson, C., Szwed, M., 2005. Trend detection in impact of new physical parametrizations in the Hadley Centre river flow series: 1. Annual maximum flow. Hydrological Sciences climate model - HadAM3. Climate Dynamics 16, 123–146. Journal 50 (5), 797–810. Rijks, D., Terres, J.M., Vossen, P., (Eds.), 1998. Agrometeorological Kundzewicz, Z.W., Radziejewski, M., Pin´skwar, I., 2006. Precipita- Applications for Regional Crop Monitoring and Production tion extremes in the changing climate of Europe. Climate Assessment. Technical Report EUR 17735 EN, European Commis- Research 31, 51–58. sion Joint Research Centre, Ispra, Italy, pp. 31–55. Kwadijk, J.C.J., 1993. The Impact of Climate Change on the Roeckner, E., Arpe, K., Bengtsson, L., Christoph, M., Claussen, M., Discharge of the River Rhine. Ph.D. Thesis, Department of Du¨menil, L., Esch, M., Giorgetta, M., Schlese, U., Schulzweida, U., Physical Geography, Utrecht University, The Netherlands. 1996. The Atmospheric General Circulation Model ECHAM-4: Model Lehner, B., Do¨ll, P., Alcamo, J., Henrichs, T., Kaspar, F., 2006. Description and Simulation of Present-day Climate. MPI Rep. 218, Estimating the impact of global change on flood and drought Max-Planck-Institute for Meteorology, Hamburg, Germany. risks in Europe: a continental integrated analysis. Climatic Semmler, T., Jacob, D., 2004. Modeling extreme precipitation Change 75, 273–299. events – A climate change simulation for Europe. Global and Mearns, L.O., Giorgi, F., Whetton, P., Pabon, D., Hulme, M., Lal, Planetary Change 44, 119–127. M., 2003. Guidelines for Use of Climate Scenarios Developed Shabalova, M., Van Deursen, W., Buishand, T., 2003. Assessing from Regional Climate Model Experiments. Available from the future discharge of the river Rhine using RCM integrations and a Intergovernmental Panel on Climate Change, Task Group on hydrological model. Climate Research 23, 233–246. Scenarios for Climate Impact Assessment, Data Distribution Szabo´, J.A., 2006. An efficient hybrid optimization procedure of Centre. . adaptive partition-based search and downhill simplex methods Monteith, J.L., 1965. Evaporation and environment. In: Fogg, G.E., for calibrating water resources models. In: Proceedings of the Kohn, P.G. (Eds.), The state and movement of water in living XXIII Conference of the Danubian Countries on the Hydrological organisms. Proceedings of the 19th Symposium of the Society for Forecasting and Hydrological Basis of the Water Management, Experimental Biology. Cambridge University Press, London, pp. 2006, . 205–234. Wood, A.W., Leung, L.R., Sridhar, V., Lettenmaier, D.P., 2004. Munich Re, 2005. Annual Review: Natural Catastrophes 2004. Hydrologic implications of dynamical and statistical approaches Knowledge Series Topics Geo. Mu¨nchener Ru¨ckversicherungs- to downscaling climate model outputs. Climatic Change 62 (1–3), Gesellschaft, Munich, Germany. 189–216. Nakicenovic, N., Swart, R. (Eds.), 2000. IPCC Special Report on Wo¨sten, J.H.M., Lilly, A., Nemes, A., Le Bas, C., 1999. Development Emission Scenarios. Cambridge University Press, Cambridge, and use of a database of hydraulic properties of European soils. United Kingdom. Geoderma 90 (3–4), 169–185.