Impact of global warming on streamflow drought in Europe

Transcription

1 JOURNAL OF GEOPHYSICAL RESEARCH, VOL. 114,, doi: /2008jd011438, 2009 Impact of global warming on streamflow drought in Europe Luc Feyen 1 and Rutger Dankers 2 Received 7 November 2008; revised 8 May 2009; accepted 15 May 2009; published 15 September [1] Recent developments in climate modeling suggest that global warming is likely to favor conditions for the development of droughts in many regions of Europe. Studies evaluating possible changes in drought hazard typically have employed indices that are derived solely from climate variables such as temperature and precipitation, whereas many of the impacts of droughts are more related to hydrological variables such as river flow. This study examines the impact of global warming on streamflow drought in Europe by comparing low-flow predictions of a hydrological model driven by high-resolution regional climate simulations for the end of the previous century and for the end of this century based on the Special Report on Emissions Scenarios A2 greenhouse gas emission scenario. For both time slices, low-flow characteristics were derived from the simulated streamflow series using extreme value analysis. More specifically, we employed the methods of block maxima and partial duration series to obtain minimum flows and flow deficits and fitted extreme value distributions by the maximum likelihood method. In order not to mix drought events with different physical causes the analysis was performed separately for the frost and nonfrost season. Results show that in the frost-free season streamflow droughts will become more severe and persistent in most parts of Europe by the end of this century, except in the most northern and northeastern regions. In the frost season, streamflow drought conditions will be of less importance under future climate conditions. Citation: Feyen, L., and R. Dankers (2009), Impact of global warming on streamflow drought in Europe, J. Geophys. Res., 114,, doi: /2008jd Introduction [2] Drought is a natural phenomenon resulting from less than normal precipitation over a large area for an extended period of time. The propagation of a precipitation deficit through the hydrological cycle may eventually lead to extreme low river flows or streamflow droughts. Droughts and low river flows have considerable economical, societal and environmental impacts. They affect several sectors, such as energy production, both in terms of water availability for hydropower and cooling water in electricity generation, river navigation, agriculture, as well as public water supply. With an increasing body of evidence for global warming and its impact on the hydrological cycle [see, e.g., Intergovernmental Panel on Climate Change, 2007] the question naturally arises how streamflow droughts or minimum river flows will develop in a future climate. [3] Over the past 30 years, Europe has been affected by a number of major drought events, most notably in 1976 (Northern and Western Europe), 1989 (most of Europe), 1991 (most of Europe), and more recently, the prolonged event over large parts of Europe associated with the summer 1 Institute for Environment and Sustainability, Joint Research Centre, European Commission, Ispra, Italy. 2 Met Office Hadley Centre, Exeter, UK. Copyright 2009 by the American Geophysical Union /09/2008JD heat wave in The most serious drought in the Iberian Peninsula in 60 years occurred in 2005, reducing overall EU cereal yields by an estimated ten per cent [United Nations Environment Programme, 2006]. Since 1991, the yearly average economic impact of droughts in Europe was C= 5.3 billion, with the economic damage of the 2003 drought in Europe amounting to at least C= 8.7 billion [European Communities, 2007]. It should be noted, however, that these numbers relate not only to streamflow droughts, but also to other types of drought including less than normal precipitation. [4] Despite the already observed changes in climate and the recent droughts experienced in Europe, there is no conclusive proof that climate change has led to more frequent and severe streamflow droughts in Europe [e.g., Hisdal et al., 2001; Hannaford and Marsh, 2006]. A few regional studies have reported a tendency toward reduced low river flows, e.g., in southern and eastern Norway [Hisdal et al., 2006] and in the Pyrenees in France [Renard et al., 2008]. In most cases, however, a possible climate related trend cannot be distinguished from the strong (multi)decadal variability. Also, human interferences in catchments, such as abstraction, irrigation, and reservoirs, partly or completely mask the climate change signal in the hydrological cycle. On the other hand, several stations in Europe have shown trends toward less severe low river flows over the 20th century, consistent with an increasing number of reservoirs becoming operational in the catchments over the period of record [Svensson et al., 2005]. 1of17

2 [5] For the coming century, the latest climate change simulations for Europe project higher temperatures, hence higher evaporative demands, changes in the seasonality of precipitation patterns, with wetter winters and drier summers, as well as an increase in the frequency and intensity of extreme climatic events [e.g., Rowell, 2005; Beniston et al., 2007]. The combination of these patterns of change will likely result in more frequent, severe and persistent streamflow droughts in large parts of Europe, especially in the south. [6] In recent years, many studies have appeared that assess the possible impacts of climate change on the hydrological cycle. Among them, studies with a focus on droughts typically have employed indices that are derived solely from climate variables such as temperature and precipitation [e.g., Beniston et al., 2007; Calanca, 2007; Blenkinsop and Fowler, 2007]. Other studies have incorporated variables from the land surface scheme of climate models, mainly soil moisture, to assess changes in drought frequency and severity [e.g., Wang, 2005; Sheffield and Wood, 2008; Burke and Brown, 2008]. [7] Many impacts of droughts, however, relate more directly to the resulting hydrological conditions such as low surface water levels and reduced groundwater recharge. Current climate models are still poor at providing reliable estimates of such variables. Hydrological models often provide a better representation of the land-surface part of the hydrological cycle. They also allow incorporating spatial details on soil, vegetation and terrain characteristics that impact local storages and flows. By simulating hydrologic variables such as river runoff, the combination of regional climate predictions with hydrological models offers a valuable counterpart to existing climate-based indices [Shukla and Wood, 2008]. We follow this approach by employing high-resolution simulations from a regional climate model (RCM) to drive a spatially distributed hydrological model. Our focus is on streamflow drought, as it reflects the spatially integrated reduction in water resources over river basins, and as such forms a major concern to water managers. [8] Few studies have adopted this approach to study the impact of future global warming on extreme low flows. De Wit et al. [2007] studied the impact of climate change on low flows in the river Meuse by combining RCM simulations from the PRUDENCE project [Christensen et al., 2002] with a local hydrological model. They found that the projected decrease of summer precipitation did not necessarily result in more severe and frequent critical summer low flows, as during prolonged dry periods the discharge in the Meuse is largely derived from water releases from groundwater storage. The aquifers supplying this water are mostly recharged during the winter season. The projected increase of winter precipitation will therefore reduce the occurrence of low flows to a certain extent in this river basin. Lehner et al. [2006] made an integrated European assessment of changes in extreme low river flows due to climate change and changes in water use using the Water- Gap global hydrology and water use model. Their analysis was based on applying the climate change signal of two different General Circulation Models (GCMs) to an observation-based data set (i.e., a delta change approach). The monthly averaged GCM output was disaggregated in space and time to the temporal (daily) and spatial scale (0.5 ) of the WaterGAP model. In the climate signal, only long-term trends and changes in seasonal climate were taken into account, while a potential change in climate variability was neglected [see Lehner et al., 2006].Weiss et al. [2007] also used the WaterGAP model driven by simulations from a GCM to evaluate changes in the frequency of droughts in the Mediterranean area. GCM climate change predictions, however, do not adequately resolve factors that influence regional climates, such as the local effects of mountains, coastlines, lakes, vegetation boundaries, and heterogeneous soils. [9] This study presents a pan-european assessment of the potential impacts of climate change on low river flows by comparing streamflow drought characteristics for current and future climate. We employed high-resolution regional climate data for the end of the current century on the basis of the A2 greenhouse gas emission scenario [Nakicenovic and Swart, 2000] to force a European-wide hydrological model. We derived low-flow characteristics from the simulated streamflow series using extreme value analysis. More specifically, we employed the methods of block maxima to select extreme low flows and fitted a Generalized Extreme Value distribution to estimate return levels. To evaluate the severity of streamflow droughts we furthermore derived partial duration series of flow deficits from the streamflow simulations and fitted a Generalized Pareto distribution. In order not to mix drought events caused by a lack of precipitation and those originating from the temporary storage of water as snow, the analysis was performed separately for the frost-free and the frost season. The accuracy of the simulations was verified by comparing the simulated streamflow drought characteristics with those derived from 30 years of observations at 209 stations throughout Europe. [10] The remainder of the paper is organized as follows. Section 2 (on the methodology) first describes the generation of the streamflow time series at the European scale using regional climate data and the hydrological model. In section 2.2 the low-flow indices and their derivation are explained. In section 3 we present the results with discussion. In section 4 the conclusions of this study are presented. 2. Methodology 2.1. Generation of Streamflow Series for Current and Future Climate [11] In spite of the ever increasing spatial resolution, current state of the art RCMs do not represent the relevant runoff-generating mechanisms in the landscape adequately enough to allow an assessment of the impacts of climate change on streamflow characteristics. Most impact studies therefore rely on hydrological models that are coupled offline with one or more regional climate models. In this study, we used output from the RCM HIRHAM [Christensen et al., 1996] to drive the LISFLOOD hydrological model [van der Knijff et al., 2009].Morespecifically, we employed climate data from an experiment with the HIRHAM model conducted within the framework of the PRUDENCE (Prediction of Regional Scenarios and Uncertainties for Defining European Climate Change Risks 2of17

3 Figure 1. Schematic overview of the LISFLOOD model. P, precipitation; Int, interception; EW int, evaporation of intercepted water; D int, leaf drainage; ES act, evaporation from soil surface; T act, transpiration (water uptake by plant roots); INF act, infiltration; Q rs, surface runoff; D us,ls, drainage from upper to lower soil zone; D ls,ugw,drainage from lower soil zone to upper groundwater zone; D pref,gw, preferential flow to upper groundwater zone; D ugw,lgw, drainage from upper to lower groundwater zone; Q ugw, outflow from upper groundwater zone; Q lgw, outflow from lower groundwater zone; Q loss, loss from lower groundwater zone. Note that snowmelt is not included, even though it is simulated by the model. and Effects) project [Christensen et al., 2002]. In this experiment the HIRHAM model was run with a very high horizontal resolution of 12 km. The climate simulations consist of two 30-year time slices: a control run with a greenhouse gas forcing corresponding to and a scenario run corresponding to according to the A2 scenario of the Intergovernmental Panel on Climate Change (IPCC) [Nakicenovic and Swart, 2000].In the control run, the lateral boundaries were derived from the HadAM3H high-resolution global atmosphere model, which itself had been forced by low-resolution observed sea surface temperature (SST) and sea-ice extent. The climate change signal in SST and sea-ice extent for future conditions came from the global coupled atmosphereocean model HadCM3 [Gordon et al., 2000; Pope et al., 2000]. The HIRHAM simulations that were used in the present study and the changes in climatology in the scenario period have been described by Christensen and Christensen [2007], Dankers and Feyen [2008], and others. [12] The LISFLOOD model is a combination of a gridbased water balance model and a 1-D hydrodynamic channel flow routing model that has been developed to simulate the hydrological behavior in European catchments [van der Knijff et al., 2009]. Even though the model was originally developed for flood forecasting purposes, it is capable of simulating the whole spectrum of river flows, including low flows, as it calculates a complete water balance for every grid cell and for every time step (1 day in the present study). A schematic overview of LISFLOOD is presented in Figure 1. Driven by meteorological input the model calculates actual evaporation and transpiration rates on the basis of vegetation characteristics, leaf area index and soil properties. Processes simulated for each grid cell include snowmelt, soil freezing, surface runoff, infiltration into the soil, preferential flow, redistribution of soil moisture within the soil profile, drainage of water to the groundwater system, groundwater storage and groundwater base flow. Runoff produced for every grid cell is routed through the river network using a kinematic wave approach. It should be noted that the model does have an option to simulate lakes and reservoirs (as described by van der Knijff et al. [2009]), which can be relevant for low-flow analysis as they tend to increase low flows. Similar as in other large-scale hydrological modeling studies [e.g., Arnell, 1999; Lehner et al., 2006], the main reason for not including them in the present study is a deplorable lack of relevant data at European scale. As a consequence, results are likely to underestimate the human influence on low flows. [13] The current European-wide model setup with a 5-km grid resolution uses spatially variable parameters on soil, vegetation and land use derived from European data sets. A set of 8 parameters that control infiltration, snowmelt, overland and river flow, as well as residence times in the soil and subsurface reservoirs, have been estimated in 231 catchments by calibrating the model against historical records of river discharge. The calibration period varied between the different catchments depending on the availability of discharge measurements, but all spanned at least 4 years between 1995 and The meteorological variables used to force the model in the calibration exercise were obtained from the Meteorological Archiving and Retrieving System (MARS) database [Rijks et al., 1998]. The upstream area of the calibrated catchments varied in size between 1,000 and 178,475 km 2. The locations of the stations used in the calibration are presented by the gray dots in Figure 2. This overview shows that the coverage is sufficient in most parts of western and central Europe. Discharge information was also available for several stations in northern Europe, which allowed tuning snow related parameters of the model in this area. For the Balkan area, southern Italy and the Iberian peninsula (except for the Ebro river basin) no discharge series were available at the time of the model calibration. For catchments where discharge measurements were not available simple regionalization techniques (regional averages) were applied to obtain the parameters. However, it should be noted that the transfer of parameters to ungauged catchment may introduce additional bias, as the geomorphological characteristics in these basins may not fully correspond to those of 3of17

4 Figure 2. Location of the discharge gauging stations used in the calibration of the hydrological model (indicated by the small gray circles) and for validation of the climate driven simulations (indicated by the larger white circles). the calibrated catchments. Hence, the assessment of changes in low-flow indices in these areas is more prone to uncertainty. A more detailed description of the different model processes and governing equations, as well as of the European-wide model setup and calibration exercise are given by van der Knijff et al. [2009] and Feyen et al. [2007, 2008]. [14] To drive the LISFLOOD model, the HIRHAM daily simulations of temperature, precipitation, solar and thermal radiation, humidity and wind speed were regridded to the 5-km grid of the hydrological model using a nearest neighbor approach on the basis of the center points of the 12-km grid cells of the HIRHAM model. Given the high horizontal resolution of the climate data and the lack of a high-quality, high-resolution meteorological data set at European scale no further downscaling or bias correction was applied to the climate data. This means that any bias in especially the precipitation fields will directly influence the LISFLOOD simulations. The radiation, humidity and wind speed output fields were used to calculate reference evapotranspiration using the Penman-Monteith model [Monteith, 1965]. [15] On the basis of the 30-year climate simulations for the control and scenario period the LISFLOOD model then calculated the water balance for every grid cell and routed the produced runoff through the river network. As such, 30 years of daily discharges were produced at each river pixel for both the control and scenario period. From the simulated discharge time series a flow duration curve (FDC) was derived for each river pixel for the control and scenario period. The FDC describes the relationship between the magnitude and frequency of streamflow discharges and as such provides information about the complete range of river discharges from low flows to flood events. It also provided 4of17

5 the basis for the calculation of the low-flow indices described below Streamflow Drought Indices [16] The low-flow regime of a river can be evaluated in a variety of ways, consequently a large number of low-flow measures exist. An extensive overview of the variety of low-flow indices and their derivation are given by Smakhtin [2001] or Tallaksen and van Lanen [2004]. In the present study we used two approaches. One is to analyze low flows through the magnitude of the river discharge. Such low-flow characteristics, however, do not consider the development in time of a drought event. In the second approach, droughts are therefore defined as periods in which discharge remains below a threshold flow. From a series of drought events obtained from the discharge time series deficit characteristics expressing the duration and severity can be derived. The two methods are described in more detail below Low-Flow Characteristics [17] To evaluate how the magnitude of minimum flows may be affected by climate change we considered the 7-day minimum flows at several recurrence intervals n (7Qn). The 7-day period eliminates the day-to-day variations in river flow and was obtained by moving an averaging window with an interval of 7 days over the discharge time series. For each year the minimum was extracted from the smoothed discharge time series, resulting in a sample of 30 yearly minima. According to the Fisher-Tippett extreme value theorem [Fisher and Tippett, 1928] the block maxima (or minima) of a sequence of independent and identically (iid) distributed random variables in the limit follow a Generalized Extreme Value (GEV) distribution [Coles, 2001; Katz et al., 2002]. The GEV is a three-parameter distribution defined by a location (m), scale (s), and shape (g) parameter. To estimate the probability of extreme low river flows a GEV distribution was fitted to the annual minimum values in every river cell using Maximum Likelihood Estimation (MLE), following Gilleland and Katz [2005]. It has been argued that the MLE is known to be unstable for small sample sizes and can give unrealistic estimates of g [e.g., Martins and Stedinger, 2000]. Alternative methods, such as the probability-weighted moments, may provide more robust estimates but lack a theoretical justification [Katz et al., 2002]. For the purpose of the present study we have therefore not explored any alternative methods of parameter estimation (see Frei et al. [2006] for an example based on a Bayesian prior distribution of the shape parameter g) Deficit Characteristics [18] The derivation of deficit characteristics is based on the threshold or truncation concept, which originates from the theory of runs [Yevjevich, 1967]. In this approach a drought is defined as a period in which the river flow is below a certain threshold or truncation level. The main deficit characteristics normally considered are the run duration (length of event) and the severity (cumulative deficit or the negative run sum) [Smakhtin, 2001]. Typically, drought durations and deficit volumes are highly correlated [Woo and Tarhule, 1994], therefore only deficit volumes are considered here. The threshold level is defined here as an exceedance frequency of the FDC rather than as a percentage of the mean flow. As such, series throughout Europe experience the same number of days with flow below the threshold level but deficit volumes may vary. Meaningful thresholds typically are in the range of discharges with 70 90% exceedance frequency for perennial streams [Tallaksen et al., 1997; Engeland et al., 2004; Fleig et al., 2006]. Several thresholds within this range were evaluated, but results will be presented here using the 80% exceedance frequency of the FDC (i.e., flow that is exceeded 80% of the time, indicated here as Q 80 ). Thresholds imposed to select drought events in the scenario period were derived from the FDC of the control climate. [19] Pickands [1975] showed that excesses over a high threshold, asymptotically, follow a Generalized Pareto (GP) distribution. In the case of river flow droughts this applies to shortfalls below a low threshold. It was assumed that the drought deficit volumes corresponding to the shortfalls are independent and identically distributed. With the location parameter (or distribution origin) set to the lower bound (or threshold) of the partial duration series, drought deficits for different recurrence intervals were then derived by inferring the scale (s) andshape(g) parameters of the GP distribution in each river pixel using MLE [Gilleland and Katz, 2005]. [20] When selecting drought events with the threshold method the daily resolution of the time series introduces two problems: dependency among droughts and the occurrence of minor droughts [TallaksenandvanLanen, 2004; Engeland et al., 2004; Fleig et al., 2006]. Mutual dependence among droughts occurs when the flow shortly exceeds the threshold level in periods of prolonged low flow, thereby dividing the large event into smaller dependent events. Minor droughts are events of very small duration and deficit volume that may distort the extreme value analysis. A number of methods to reduce such events was assessed by Tallaksen et al. [1997]. On the basis of their results, and similar as for the low-flow indices described above, a moving average (MA) procedure with a 7-day averaging window was applied to the discharge time series before selecting the events, which pools mutually dependent droughts and filters out minor droughts. It was observed, however, that the MA procedure did not exclude all small events. Therefore, following Zelenhasic and Salvai [1987], additionally all events with a deficit smaller than 0.5% of the maximum deficit of the time series were excluded from the further analysis Characteristics of Streamflow Regimes [21] Climatologic and hydrologic conditions vary strongly across Europe, which complicates the application of a general procedure. In highly seasonal climates streamflow droughts may be generated by different physical processes and two distinct low-flow seasons may result [e.g., Fleig et al., 2006]. In cold or mountainous regions, low winter flows are due to temporary storage of precipitation as snow, whereas summer low flows are caused by reduced precipitation and high evapotranspiration losses. As rivers in these regions typically have their lowest flows in winter, an annual analysis masks any changes that would occur in summer droughts. Combining drought events generated by different mechanisms also violates the basic requirement of frequency analysis, namely that the sample should come from an independent and identically (iid) distributed random variable. Droughts caused by temporary storage of precipitation as snow can be separated from those caused 5of17

6 by a lack of precipitation by combining the streamflow data with information about snow cover or temperature, for example by defining the frost season as the period when the monthly average temperature drops below 0 C [Hisdal et al., 2001]. In this study, we performed an analysis for the nonfrost and frost season, whereby for both current and future climate the frost season in a certain river pixel is defined as the period when the monthly average temperature in the upstream area falls below 0 C. Using monthly average temperatures over the upstream area rather than at the grid cell itself avoids that at a given downstream location a period is labeled as nonfrost while the majority of the upstream area may still be covered by snow. This is, for example, the case for downstream reaches of large rivers draining from the Alps, such as the Rhine, Rhone, or Po. [22] In transition regions between cold and warm climates not all years may have a frost season, which means that the series of annual minima may contain less than 30 values. In these regions a conditional probability procedure is recommended to correct for the years without frost season. Following Stedinger et al. [1993], an extra parameter was introduced that describes the probability p 0 that there is no data in a given year (i.e., that no frost season exists) and the following model was considered F m;s;g ðþ¼p x 0 þ ð1 p 0 ÞG m;s;g ðþ; x where the parameters of G m,s,g (x) were obtained by applying the GEV fitting procedure to the minima of the frost years. Stedinger et al. [1993] recommend not to apply this method when more than 25% of the record has zero values. Therefore, river pixels were included in the frost season analysis when at least 23 out of 30 years have a frost season. [23] The application of a general procedure is furthermore complicated by the presence of intermittent and ephemeral streams that are characterized by periods of no flow. This results in data series of annual minima containing several zero values, which requires the use of discontinuous probability distribution functions. However, evaluation of the LISFLOOD simulations in the control and scenario periods showed that for catchments with an upstream area larger than 1000 km 2 the time series of annual minima in both the nonfrost and frost season did not contain any zero flow values. Hence, larger catchments in Europe can be considered to be perennial even under future climate conditions. The analysis was therefore limited to catchments with an upstream area larger than 1000 km Results and Discussion 3.1. Validation of Simulations Using Control Climate [24] Simulated low flows for the control period were validated against observations at 209 gauging stations across Europe for which long enough daily discharge time series (30 years covering or ) were available. For nearly all validation stations the discharge series were retrieved from the Global Runoff Data Centre database ( The upstream area of the validation stations varies between 1,500 and 807,000 km 2. As shown in Figure 2, the distribution of the validation stations is very uneven, with the majority of stations located ð1þ in western parts of Europe and hardly any in the Mediterranean countries. From the 209 validation stations, 76 were also used in the calibration of the hydrological model (stations in Figure 2 with overlapping symbols). Note, though, that the calibration period consisted of only a limited number of years between 1995 and 2002, hence not the full 30 years, and used observed meteorological information rather than simulated climatology by the RCM. At 158 stations the number of years with a frost season in the control climate was at least 23 out of 30; hence they were used for validation in both the nonfrost and frost season analysis. Minimum flows range between 2 and 3200 m 3 /s and between 3 and 4300 m 3 /s in the nonfrost (Figure 3a) and frost season (Figure 3b), respectively. For the threshold adopted (Q 80 ), deficit volumes range between 1.4e +6 and 2e +9 m 3 and between 4.4e +5 and 1.3e +9 m 3 for the nonfrost (Figure 3c) and frost season (Figure 3d), respectively. This shows the variability in flow magnitude and volumes that is captured by the 209 validation stations. [25] Control climate simulations of the HIRHAM model do not reproduce the historical weather of the period, but only the average climate conditions. Hence, a direct day-to-day or even year-to-year comparison is not feasible. Therefore, the accuracy of the simulations was assessed by comparing observed and simulated low-flow indices visually as well as through the statistical measures coefficient of determination (r 2 ) and model efficiency (EF), defined as 8 9 X n >< Y obs;t Y obs Ysim;t Y sim >= r 2 t¼1 ¼ " X n # " 0:5 2 Y obs;t Y X # n 0:5 2 >: obs Y sim;t Y sim >; t¼1 ", # EF ¼ 1 Xn 2 X n 2 Y obs;t Y sim;t Y obs;t Y obs ; ð3þ t¼1 where Y obs,t and Y sim,t are the observed and simulated lowflow variable at (yearly) time step t, respectively, and the horizontal bar denotes the average over all years. The coefficient of determination r 2 describes the proportion of the total variance in the observed data that can be explained by the model. It ranges from 0.0 to 1.0, with higher values indicating better agreement. The coefficient of efficiency EF is the ratio of the mean squared error to the variance in the observed data, subtracted from unity. It ranges from minus infinity to 1.0, with higher values indicating a better agreement. [26] Figure 3 compares for the nonfrost and frost season the observed versus simulated average annual minima (Figures 3a and 3b), average deficits based on Q 80 (Figures 3c and 3d), minimum flows with return period of 20 years (Figures 3e and 3f), and maximum deficits with return periods of 20 years (Figures 3g and 3h). The gray points in these plots represent the stations that were also used in the calibration of LISFLOOD. The results show that the hydrological model driven by the regional climate simulations reproduces the observed low-flow sta- t¼1 t¼1 2 ð2þ 6of17

7 Figure 3. Observed versus simulated (a and b) average annual minimum, (c and d) average deficit based on Q 80, (e and f) minimum flow with return period of 20 years based on GEV fit, and (g and h) maximum deficits with return period of 20 years based on GPD fit. Figures 3a, 3c, 3e, and 3g represent the nonfrost season; Figures 3b, 3d, 3f, and 3h represent the frost season. Stations used in the calibration of the hydrological model are indicated by the gray circles. tistics reasonably well, with a general tendency of better performance with increasing catchment size. Values of r 2 range from 0.87 to 0.95 and the EF ranges from 0.78 to Also at stations not used in the calibration of the hydrological model the performance is acceptable. Uncertainty arising from fitting the extreme value distributions through the data results in slightly inferior agreements between observed and simulated fitted minima and deficit volumes compared to the averages (compare Figures 3e 3h with Figures 3a 3d). [27] Notwithstanding the overall good agreement between the observed and simulated low-flow statistics, large discrepancies do occur at a small number of stations, where the relative error can be 1 or 2 orders of magnitude (note the 7of17

8 Figure 4. Return level plots of minimum flows for the nonfrost season at selected stations in Europe (for locations see Figure 1). The black crosses and solid lines represent the observations and corresponding GEV fits, the dark blue circles and dashed lines represent the control climate simulations and corresponding GEV fits, and the red triangles and dashed lines represent the scenario climate simulations and corresponding GEV fits. Information is provided on the upstream area of the station (UpA) and whether (Y) or not (N) the station was used in the calibration of the hydrological model (cal). logarithmic scale). Deviations from the observation-based statistics can be attributed to errors in the climate simulations, as well as to errors in the hydrological model, its static input and in the calibration and regionalization of its parameters. Several studies [e.g., Wilby, 2005], however, showed that uncertainty in the hydrological model is generally much lower than the uncertainty of the climate input. Some of the disagreements can also be linked with man-made modifications of low-flow regimes in many catchments in Europe [see Dynesius and Nilsson, 1994] that are not accounted for in the hydrological model. Lower observed streamflow levels can, for example, be due to an increased outtake of water for irrigation purposes, whereas minimum flow requirements and river regulation may result in artificially higher flows than natural during low-flow periods. The relative impact of water extraction and river flow regulation can be considered to be higher in the low region of the flow spectrum, which renders low-flow regime analysis more susceptible to large errors introduced by unaccounted alterations than average or peak flow analysis. [28] Figure 3 also shows that simulations in the frost season are inferior to those in the nonfrost season, with a tendency to underestimate observed low flows and deficits. This suggests that the hydrological model driven by the HIRHAM simulations does not adequately reproduce the generation of runoff and base flow in the frost season. This 8of17

9 Figure 5. Return level plots of minimum flows for the frost season at selected stations in Europe (for locations see Figure 1). The black crosses and solid lines represent the observations and corresponding GEV fits, the dark blue circles and dashed lines represent the control climate simulations and corresponding GEV fits, and the red triangles and dashed lines represent the scenario climate simulations and corresponding GEV fits. Information is provided on the upstream area of the station (UpA) and whether (Y) or not (N) the station was used in the calibration of the hydrological model (cal). may be due to (1) an underestimation of winter temperatures by the regional climate model, especially during intermittent melt events, hence more water remains stored as snow; (2) an underestimation of the autumn and early winter precipitation by the regional climate model in Northern Europe, as suggested by May [2007], resulting in a soil and groundwater storage that is too low at the start of the winter; or (3) errors in the conceptualization and parameterization of the snow and frost module in the hydrological model, particularly with respect to drainage in the cold season. Uncertainties in the observed winter precipitation [see Goodison et al., 1998; Yang et al., 2001] that was used in the calibration of the hydrological model may have further affected the parameterization of the groundwater reservoir. The relative contribution of these factors is difficult to quantify because observed events cannot be compared on an individual basis with simulations for the control climate, as it does not reproduce the historical weather. In part, higher observed minimum flows could also be caused by the release of storage from reservoirs to ensure minimum flow requirements for hydropower purposes. Flow regulation may also result in artificially higher thresholds (Q 80 ) compared to the thresholds derived from the simulated time series, which under severe drought conditions may result in large deficit volumes. As a consequence, observed deficit volumes tend to be larger compared to those simulated at the majority of stations. [29] Figures 4 7 show the results of the extreme value fitting for a selection of 9 validation stations. These stations are a representative sample from the set of 209 stations in terms of geographical spread (see Figure 2), catchment size (ranging from 4390 to 709,100 km 2 ), the presence of a frost season (three stations with frost season in both control and scenario climate), and whether or not they were used in the calibration of the hydrological model (three stations, see Figure 2). [30] Figure 4 (nonfrost season) and Figure 5 (frost season) show that the GEV distribution generally fits the annual minima reasonably well at most stations, also at higher return periods. The correspondence between the observations (black crosses and solid line) and control climate simulations (dark blue circles and dashed line) differs strongly between the stations. At station Pulo do Lobo on the Guadiana (Figure 4a), for example, control climate simulations largely overestimate the observed minima. In this region of Spain and Portugal, however, the natural flow regime is strongly affected by water extraction for irrigation. At the station of Neuhausen on the Rhine (Figures 4g and 5a), on the other hand, the observation based minima and fit are higher than those based on the control climate simulations. This is likely due to the regulation of outflow from the Lake Constance upstream of this station, which results in artificially higher minimum flows. This is also the case for the regulated river Gloma (Figures 4h and 5b) in Norway [Borgestrøm and Løkensgard, 1984]. Although the Kemijoki in Northern Finland is also regulated for hydropower, the location and shape of the GEV fits compare very well in the nonfrost season (Figure 4i). Flow regulation affects low flows also in the frost season (Figures 5a 5c), but here the underestimation of the minima is more pronounced than during the nonfrost season (Figures 4g 4i). This may be explained by inadequacies in the simulation of runoff in winter, as discussed above. [31] Fits of the GP distribution to the observed and simulated deficit volumes at the 9 stations are presented in Figure 6 (nonfrost season) and Figure 7 (frost season). These fits show that, even after removing the smallest events (see discussion on the derivation of deficit series in section 2.2.1), an imbalance between many smaller events and a few big events may affect the goodness of fit of the GP distribution to the data. In the case of a negatively fitted value for the shape parameter g the distribution has no upper limit, and, especially for return periods higher than 20 years, the extrapolation uncertainty of the GP distribution can be considerably high (e.g., Tweed, Figure 6d). For the nonfrost season, there does not seem to be a consistent over- or underestimation of observed flow deficits (see also Figure 3g). Note that at station Pulo do Lobo on the Guadiana, where observed minimum flows were largely overestimated (Figure 4a), the simulated deficits correspond closely to the observed deficits (Figure 6a). This suggests that water extraction in this region has lowered the absolute values of the low flows, including the threshold 9of17

10 Figure 6. Return level plots of deficit volumes for the nonfrost season at selected stations in Europe (for locations see Figure 1). The black crosses and solid lines represent the observations and corresponding GEV fits, the dark blue circles and dashed lines represent the control climate simulations and corresponding GEV fits, and the red triangles and dashed lines represent the scenario climate simulations and corresponding GEV fits. Information is provided on the upstream area of the station (UpA) and whether (Y) or not (N) the station was used in the calibration of the hydrological model (cal). Q 80 for the derivation of the deficits, but that the variability in flow, expressed by shortfalls below the threshold, has remained fairly constant. Comparison of Figures 5a and 5b with Figures 7a and 7b, respectively, shows that flow regulation in the Rhine and Gloma basins affects the magnitude of the low flows in the frost season, whereas deficits below the threshold are less affected as well. At station Isohaara on the Kemijoki (Figure 7c), on the other hand, simulated deficits deviate strongly from those observed, indicating that here flow regulation and inadequacies in the simulations affect not only the reproduction of the minimum discharges (Figure 5c) but also of deficit volumes Future Changes in Streamflow Drought Characteristics Changes in Meteorological Forcing [32] For a better understanding of the results, we briefly summarize the changes in temperature and precipitation between the control and scenario climate simulations of the HIRHAM model. Changes in the average temperature and precipitation are presented in Figures 8 and 9. Note that in these plots the changes in every grid cell represent the average change over the entire upstream area of that cell rather than at the location itself, as streamflow at a given 10 of 17

11 Figure 7. Return level plots of deficit volumes for the frost season at selected stations in Europe (for locations see Figure 1). The black crosses and solid lines represent the observations and corresponding GEV fits, the dark blue circles and dashed lines represent the control climate simulations and corresponding GEV fits, and the red triangles and dashed lines represent the scenario climate simulations and corresponding GEV fits. Information is provided on the upstream area of the station (UpA) and whether (Y) or not (N) the station was used in the calibration of the hydrological model (cal). location reflects the spatially integrated hydroclimatological conditions over the upstream river basin. [33] Temperatures are projected to increase all over Europe, with increases that are most pronounced in southern parts of Europe in the nonfrost season. The gray shaded areas in Figure 8b indicate regions that were characterized by frost in the control climate, i.e., where at least 23 out of 30 years had a frost season, but not in the future climate. This implies that because of global warming many catchments in central and eastern Europe will no longer experience a permanent frost season in at least 3 out of every 4 years. In regions with a frost season also in the future climate, temperature increases during the frost season are projected to be higher than in the rest of the year. This is consistent with other studies using different RCMs [e.g., Christensen et al., 2007], which for northern parts of Europe found the strongest increase in temperature during winter. [34] Changes in average precipitation in the frost-free season (Figure 9a) are less uniform across Europe. Reductions in precipitation are projected mainly in the south, but also in some central and eastern parts of Europe, as well as locally in northern regions. During the frost season the climate model projects strong increases in average precipitation over most parts of northern Europe, which is in line with other studies using different RCMs and GCMs [see Christensen and Christensen, 2007]. [35] It should be noted that in this study only changes in climatology are considered and not changes in land use or Figure 8. Change in average temperature in the HIRHAM A2 scenario run compared to the control run for (a) nonfrost and (b) frost seasons. Changes represent the average change over the entire upstream area rather than at the location itself. The gray shaded area indicates regions characterized by a frost season in the control climate and without a frost season in the scenario climate. 11 of 17

12 Figure 9. Change in average precipitation in the HIRHAM A2 scenario run compared to the control run for (a) nonfrost and (b) frost seasons. Changes represent the average change over the entire upstream area rather than at the location itself. vegetation characteristics (e.g., Leaf Area Index). Such changes may affect evepotranspiration as well as soil moisture redistribution and groundwater recharge, and consequently the development of droughts Extreme Value Fitting for Future Climate [36] GEV fits to the minima for the scenario climate at the selection of validation stations are shown in Figures 4 and 5 for the nonfrost and frost season, respectively (red triangles and dashed lines). At all 9 stations but the most northerly one (Isohaara on the Kemijoki, Figure 4i), the GEV fits for the scenario climate indicate that minimum flows in the nonfrost season will be more severe by the end of this century. The magnitude of change differs between the stations and depends on the return period, with some stations having stronger reductions in low flows with smaller return periods (e.g., Tweed, Figure 4d), whereas in other stations the reductions are more pronounced at larger return periods (e.g., Gloma, Figure 4h). At the station Beaucaire on the Rhone (Figure 4f), the direction of the signal changes for return periods of 100 years or higher, but this may be an artifact of extrapolation based on a GEV distribution fitted to only 30 years of data. In the frost season, low flows are projected to increase, mainly due to warmer and wetter winters, with a stronger increase for lower return periods. [37] GP fits for the scenario period (Figure 6) show an increase in deficit volumes by the end of this century in the nonfrost season, except in the Kemijoki. In the Gloma (Figure 6h) the GP distribution fits the data well at small return periods (<3 years), but does not reproduce the tail of the distribution. This leads to an underestimation of the simulated deficits by the GP distribution at return periods between 3 and 10 years, and a strong overestimation at larger return periods. In the frost season, deficits are projected to decrease strongly at the 3 stations shown in Figure 7. Note that the thresholds for establishing flow deficits are based on the control climate simulations. No drought events occurred in the scenario period in the Kemijoki (Figure 7c), and hence no GP distribution could be fitted. Even though the exact magnitude in change is not quantifiable, this implies, of course, a strong reduction in future streamflow deficits at this location. We also note that the change of the curvature of the GP fit at several stations between the control and scenario climate relates to fitted shape parameters of the GP distribution with opposite sign. [38] The extreme value analysis for the control and scenario period show that large bias may be introduced when drought indices, especially volume deficits, are extrapolated to large return periods on the basis of a limited time series of 30 years. Since relatively short return periods are often sufficient in low-flow design [Tallaksen and van Lanen, 2004], we therefore focus on changes in flow indices with return periods of 2 and 20 years in the remainder of this paper Future Changes in Streamflow Minima [39] The evaluation of changes in minimum flow across Europe is based on the GEV distributions that were fitted in each river pixel to the simulated yearly minima in both the control and scenario period. Figure 10 shows, for river pixels with an upstream area larger than 1000 km 2,thechangesin 7-day minimum flows with recurrence intervals of 2 and 20 years in both the nonfrost and frost season. A decrease in minimum flows (indicated in red) means that low flows are getting lower, while an increase (indicated in blue) implies that streamflow droughts are becoming less severe. In the nonfrost season (Figures 10a and 10b) minimum flows are projected to decrease in most parts of Europe, except in the mostnorthernandinnortheasternregions.inmanyregions, for example the Iberian Peninsula, southern France and the Alpine region, reductions of 20 up to 40% are projected. We note again, however, that the simulations in regions with few calibration and validation stations, mainly southern parts of Europe, are more prone to uncertainty. The decrease in minimum flows is caused by reduced precipitation and higher evaporative demands due the rise in temperature. The latter, however, does not necessarily result in higher absolute actual evapotranspiration rates, which may be limited by the reduced availability of water in the soil 12 of 17

13 Figure 10. Change in the estimated minimum flow in the scenario run relative to the control run for recurrence intervals of (a and c) 2 and (b and d) 20 years for nonfrost (Figures 10a and 10b) and frost seasons (Figures 10c and 10d). The return levels are estimated on the basis of a GEV distribution. Thin yellow lines indicate regions characterized by a frost season in the control climate and without a frost season in the scenario climate. and subsurface storages. The projected strong reductions in minimum flow in southern parts of Europe compares well with the results of Lehner et al. [2006], even though their analysis is based on another greenhouse gas emission scenario and used (much courser) climatology simulations from other GCMs. [40] In many regions of Europe the reductions in minimum flows are projected to be relatively less severe at larger recurrence intervals than for those with shorter return periods. This can be seen, for example, in the British Isles, the Benelux, Germany, France, northern Italy and eastern parts of Europe. Comparison of the GEV distributions for the control and scenario climate at stations Tweed and Beaucaire (Figures 4d and 4f) shows in more detail how the distributions converge at larger return periods. [41] The fact that in these regions minimum flows with lower recurrence intervals are more affected by climate change than those with higher recurrence intervals can be explained by the projected changes in precipitation. In these regions, precipitation is projected to reduce strongly in summer and to a lesser extent also in autumn, while increasing strongly in winter and slightly in spring [see Dankers and Feyen, 2008]. Minimum flows in these regions generally occur in summer or autumn and at lower recurrence intervals they reflect the strong reduction in summer and autumn precipitation. More severe streamflow droughts, i.e., with a higher recurrence interval, typically result from precipitation deficits over longer periods (not only summer and autumn season). Over such a long period of time, the strong reduction in summer and autumn precipitation may be partly offset by higher subsurface storages at the beginning of the summer season due to increased winter and spring precipitation. As such, the change in minimum flows for higher recurrence intervals more closely follows the average change in precipitation over the nonfrost season (Figure 9a), 13 of 17

14 Figure 11. Change in the estimated flow deficit volumes in the scenario run relative to the control run for recurrence intervals of (a and c) 2 and (b and d) 20 years for nonfrost (Figures 10a and 10b) and frost seasons (Figures 10c and 10d). The return levels are estimated on the basis of a GP distribution. Thin yellow lines indicate river sections characterized by a frost season in the control climate and without a frost season in the scenario climate. Thin gray lines indicate river sections where the number of events in the scenario period was less than 10 and no GP distribution was fitted. which, especially in areas without or with a short frost season, reflects the change in average annual precipitation. [42] In the most northern and in northeastern regions of Europe minimum flows are projected to become less severe in the nonfrost season, with the increase being more pronounced for larger return periods. Higher minimum flows result from the increase in precipitation, which offsets increased evapotranspiration due to higher temperatures. The stronger relative increase of minimum flows that occur less frequently reflects the pronounced increase in precipitation during the frost season, which results in larger subsurface storages at the start of the summer season. In some high-latitude areas, however, the projected increase in precipitation may be offset by a reduced and earlier snowmelt, resulting in lower minimum flows later in the frost-free season (e.g., parts of Sweden). [43] For the frost season, changes can only be evaluated when a sufficient number of years with frost occur in both the control and scenario period. The yellow lines in Figures 10c and 10d indicate rivers that are characterized by a frost season in the control period (i.e., at least 23 out of 30 years have a frost season) but not in the scenario period. In all rivers still having a frost season in the future climate, minimum flows are projected to be considerably less severe. In warmer and wetter winters, a smaller fraction of precipitation will fall as snow, resulting in increased flows in the cold season. In some regions, for example in southern parts of Sweden and the most eastern parts of the model domain (Romania, Belarus and western parts of Ukraine), the relative increase of the minimum flows is less pronounced at higher return periods. Again, this can be explained by the fact that the more extreme low flows are affected to a greater extent than average minimum flows by the hydroclimatologic conditions prior to the start of the frost season. In these regions summer and to a lesser extent also autumn precipitation is projected to decrease [see Dankers and Feyen, 2008] (see also Figure 9a 14 of 17