Regional climate model and model output statistics method uncertainties and the effect of temperature and precipitation on future river discharges in Scandinavia

Recent climate change studies have given strong evidence that daily weather variability will change along with average climate conditions. Due to these changes, the global hydrological cycle is expected to intensify (Held & Soden 2006). At high latitudes, as well as around alpine regions at lower latitudes, changes in river flows are driven by changes in snow melt/accumulation (López-Moreno & Garcia-Ruíz 2004; Stewart 2009) and evapotranspiration (Nohara et al. 2006), which depend on combined changes in temperature and precipitation. In Scandinavia, changes in snow melt affect river flows in winter and spring (Beldring et al. 2008; Arheimer & Lindström 2015), while changes in evapotranspiration dominate in summer (Olsson et al. 2015). The implications of changes in river flows for socio-economic sectors such as hydro-power production, water resource management and the environment including inflows to the Baltic Sea need a proper assessment.

However, due to uncertainties in future greenhouse gas emissions, inherent limitations in hydrological and climate model formulations and resolutions as well as differences between the statistical tools used to bias-correct climate model simulations (Chen et al. 2011a, 2011b; Hagemann et al. 2011; Dobler et al. 2012; Bosshard et al. 2013; Donnelly et al. 2013), uncertainties in the hydrological climate change impact modeling chain are substantial. Part of the uncertainty is also related to the internal variability of the climate system, which can be as important as the other major sources of uncertainty (Hawkins & Sutton 2011; Lafaysse et al. 2014). Thus, the assessment of potential uncertainties is an integral part of communication for adaptation and mitigation studies. Without them, decision-making might be misled causing potential economic losses.

Model output statistics (MOS) is a technique commonly used when correcting biases in global (GCM) and regional (RCM) climate model simulations before using them as input for hydrological models (Fowler et al. 2007; Graham et al. 2007; Maraun et al. 2010; Teutschbein et al. 2011). MOS consists of several types of methods of differing complexity, including the commonly used quantile mapping methods. The ability of many MOS methods in adjusting climate projections has been documented extensively in the literature (Piani et al. 2010; Chen et al. 2013), and studies have shown that most of the MOS methods perform relatively well in reproducing the identified observed hydrological conditions (Chen et al. 2013). However, several methods should be used in parallel in climate change impact studies, since the identification of a single universally well performing method is practically impossible (Räisänen & Räty 2013; Räty et al. 2014).

As the main drivers of changes in surface hydrology are changes in precipitation and temperature, it is of interest both from the hydrological model development and downscaling perspectives, how a hydrological model responds to changes in these quantities. In a recent study by Bosshard et al. (2014), the separate effects of temperature and precipitation changes on river discharge in the Rhine River catchment were examined. The study showed that in snow-dominated areas, temperature effects tend to have a larger impact on the annual cycle whereas in other areas precipitation effects are also important. Similar conclusions have also been obtained in earlier studies (Barnett et al. 2005 and references therein). The study made by Bosshard et al. (2014), however, covered a relatively small area in central Europe and used a spectral delta change method, which does not take changes in daily variability directly into account. Some processes, such as snow melting and accumulation, may be sensitive even for small changes in daily (co-)variability in areas where temperatures are in the present-day climate close to the melting point (Barnett et al. 2005).

The relative importance of GCM-RCMs and MOS methods as uncertainty sources in hydrological simulations have been studied previously both at global (Hagemann et al. 2011; Chen et al. 2011a) and regional scales (Chen et al. 2011b; Teutschbein et al. 2011; Dobler et al. 2012), but less attention has been given to the spatial sensitivity of these two uncertainty components as well as to their contributions at different parts of the discharge distribution. This study aims to evaluate these aspects in the Scandinavian region. In particular, we aim to:

• 1. evaluate how uncertainties in the forcing data are reflected in hydrological simulations, i.e., how climate model and MOS method differences contribute to the overall spread in the future climate change signal;

• 2. assess separate temperature and precipitation effects on the simulated changes in river discharges and how their contributions depend on a particular GCM-RCM and MOS-method combination.

Analysis of variance (ANOVA) methods are used in a similar manner as depicted in Dobler et al. (2012), Bosshard et al. (2013), and Räty et al. (2014). Similarly, the analysis of separate temperature and precipitation effects on simulated changes in river discharges is based on ANOVA methods (Bosshard et al. 2014). In contrast to the previous study, the daily variability is taken into account in the assessment of these effects, and the effects are evaluated not just for one GCM-RCM but a whole set of GCM-RCMs and MOS methods, thus providing more insight into the generality of the results.

The paper is organized as follows. Immediately below, the hydrological model setup and the input data are described, while the following section introduces the MOS methods and the ANOVA frameworks used in the study. The next section presents the results, particularly the different sources of uncertainty and the effects of temperature and precipitation on different components of the water balance. Also the main aspects of simulated changes in river discharges as well as in daily temperature and precipitation are illustrated. Finally, conclusions are drawn.

The ERA-Interim re-analysis (Dee et al. 2011) is used as the baseline data. To reduce differences between the re-analysis and the observed precipitation climatology, monthly precipitation means have been adjusted to the monthly means from the Global Precipitation Climatology Centre (GPCC) gridded data set (Schneider et al. 2011). These adjusted ERA-interim data (hereafter referred to as ERA-I-adj) are used as the reference data in the bias-correction. For future projections of climate, six GCM-RCMs (Table 1) were selected from the ENSEMBLES database (van der Linden & Mitchell 2009). To maximize independence between the GCM-RCM formulations, all the selected models have a different GCM and RCM component. The model simulations have been driven with SRES A1B emission scenario forcing and their spatial resolution is 0.25 ° (∼25 km) in both latitude and longitude. Both ERA-I-adj and the RCM simulations were interpolated to the corresponding sub-basins using the nearest-neighbor interpolation method. One should note that the selected GCM-RCMs might not be optimal in terms of their performance in the present-day climate, but act as a useful test for MOS method ability to adjust GCM-RCM biases.

The modeling experiments are kept as simple as possible, i.e., only daily mean temperature and daily precipitation are used as input variables. The study concentrates on analyzing simulated changes in the seasonality of river discharges in the mid-21st century (2041–2070) as compared to current conditions (1980–2009), but also changes to total runoff and evapotranspiration are discussed.

Previous studies with the E-HYPE v2.1 have shown that in Scandinavia the model tends to underestimate river discharges (Donnelly et al. 2016). This is at least partly related to the underestimation of precipitation in the baseline data even after the adjustment with GPCC, as the data set has not been corrected for undercatch for snow falling as precipitation or the under-representation of stations at higher latitudes. To take the underestimation of discharge values in the present-day climate into account, a regional correction of precipitation and an adjustment of evapotranspiration parameters (soil type and land-use dependent) was made to the original model. These parameters were optimized to the Nash–Sutcliffe efficiency coefficient (NSE) and the relative volume error (RE) for 213 stations around the Scandinavian region in the years 1991–2000 and then validated over the whole baseline period (1980–2009).

To study the separate effects that changes to temperature and precipitation have to changes in river discharges, the analysis follows the approach described in Bosshard et al. (2014). Assuming that as a first approximation non-linear effects are negligible, the relative contributions of temperature and precipitation changes are derived using four sets of HYPE simulations: one for present-day conditions, one including both temperature and precipitation changes, one for precipitation changes, and one for temperature changes only. From these simulations, using a simple ANOVA analysis one can assess the magnitudes of temperature and precipitation effects to the simulated changes together with the interaction term. The separation of temperature and precipitation effects is done individually for each GCM-RCM simulation and the two delta change methods, which allows us to compare how the relative importance of these two effects vary from one GCM-RCM and MOS-method combination to another. Due to methodological constraints (see Bosshard et al. (2014) for more details), only the two delta-change methods are included in the assessment of temperature and precipitation effects on the changes in the annual cycle of simulated surface hydrology.

We briefly demonstrate the performance of the HYPE model in the present-day climatic conditions. Validation against the observed discharges showed that the median daily NSE was close to 0.5 and ranged between 0.87 and −13.51. Locations with poor performance could often be related to strong regulation of the river for hydro power. Also the limited resolution of the baseline forcing and the limited capability of river routing network data sets to delineate small catchment areas (Donnelly et al. 2012) hamper the use of the model in catchments with size <5,000 km². The median annual mean relative volume error for river discharge was small (−1%) after the simple calibration and ranged between ±20% for 74% of the stations.

Further issues related to the model formulation were also identified: first, the snow melting scheme used in this model version was found to be sensitive to changes in the parameter values, which causes uncertainties to the simulated snow accumulation and melting; second, the potential evapotranspiration is calculated using a relatively simple temperature-based scheme, which may not be sufficient in regions where wind or other processes strongly affect evapotranspiration. These limitations, together with the fact that the model parameters and structure are uncertain for future climatic conditions, need to be kept in mind when assessing the model results, even though HYPE is able to simulate seasonal aspects as well as inter-annual variability realistically in the present-day climate (Donnelly et al. 2016). Qualitatively, the accuracy of such a model is somewhere between a land surface scheme and an individual catchment model.

During the summer months, the average projected increases for temperature are smaller than in winter. Also the spread between the model–method combinations follows the spatial pattern of the mean changes and varies less between different areas than in the winter. Compared to the winter season, disagreement between the projections is smaller during the summer around Norwegian coastal areas, but slightly larger around Finland and the southern parts of Sweden.

Domain mean changes in daily discharges from 1980–2009 to 2041–2070 are shown in Table 2. The results show that several of the GCM-RCM and MOS-method combinations project decreases in the annual mean discharge. These values are somewhat lower than the results obtained in recent literature (Donnelly et al. 2013; van Vliet et al. 2015). However, there are large regional variations. In Norway, discharges seem to mainly increase, while in southern Sweden and Denmark discharges are projected to decrease (not shown). Variations between the GCM-RCMs are also substantial (as also seen by van Vliet et al. (2015)). Some models project increases also at annual level, which underlines the uncertainty related to the choice of the GCM-RCM component.

The changes in the seasonal characteristics of river discharge from 1980–2009 to 2041–2070 are shown in Figure 1. During the summer, the mean river discharge seems to decrease nearly everywhere, especially in the Scandinavian mountains and in the northern parts of the model domain. During the winter months, discharges are projected to increase by over 40% in most parts of the domain reflecting the shift of the snowmelt from spring season to winter with the exception of southern Sweden and Denmark where the discharges seem to decrease instead due to projected decreases in precipitation and increases in evapotranspiration. Thus, even if mean annual precipitation is expected to increase in almost all of Scandinavia, the future discharge is projected to increase only in some parts of the region. Although the annual multi-model-mean changes of daily discharge are uncertain, seasonal and spatial variations are relatively well in line with previous results (e.g., Donnelly et al. 2013).

The left panel of Figure 3 shows the ANOVA decomposition for individual percentiles of domain-mean daily discharge distribution changes for the period 2041–2070. It is seen that the GCM-RCM effect (which includes the internal variability) is larger than the other two components for most percentiles and usually exceeds 40% of the total variance. However, for low and high flows (i.e., the tails of the distribution), the contribution of the MOS-method differences increases, which is important information when using these methods for assessing changes to extremes such as floods and droughts. This is likely caused by the differing treatment of extremes in the MOS methods, which varies from constant scaling to adjustment to an empirical frequency distribution. For high flows the interaction term is also relatively large. These results resemble those obtained in Räisänen & Räty (2013) and Räty et al. (2014), who showed that most of the uncertainty in daily mean temperature and precipitation projections comes from the GCM-RCM differences, rather than differences between the MOS methods.

Similar decompositions are also shown for daily mean temperature and precipitation. For temperature, the model component dominates over the other two components, as in Räisänen & Räty (2013). The same also holds for precipitation, although the interaction component is slightly larger. To cover the most important water balance components, results are also shown for evapotranspiration. In winter the spatial pattern resembles those for the total runoff and river discharge, with somewhat lower (higher) contribution from the model (method) differences, indicating greater dependence on the selected method. Yet, it has to be kept in mind that evapotranspiration values are small during the winter also in the future climate, which might add to the interaction term.

The results are slightly different for the summer season (Figure 5), when climate model differences generally explain a larger part of the overall uncertainty in river discharge and total runoff. Note the absence of the relatively larger values for the MOS-method component in the northern parts of the domain seen during the winter. This further indicates that it is snow melting and accumulation which are affected by the choice of the MOS method. Also for evapotranspiration, the RCM component is noticeably larger in summer.

Previous results lead to the question, to what extent are future changes in river discharge and other hydrological aspects regulated by changes in either precipitation or temperature? Figure 6 illustrates how temperature and precipitation changes modify separately the annual cycle of river discharge for the years 2041–2070 in comparison to the baseline period. Comparing the simulations that include either temperature or precipitation changes only, it is seen that changes in temperature have a positive effect on river discharges in winter and early spring whereas in other seasons the temperature effect tends to be negative. On average, the temperature effect is much stronger than the precipitation effect and explains most of the changes in discharge seasonality. In particular, decreases in river discharges during the spring and summer seem to be mostly temperature regulated. Based on the previous results, this is probably due to the smaller snow reservoir in winter and earlier timing of snow melt in spring in a large part of the model domain. Also the sensitivity of evapotranspiration to temperature changes might explain the larger temperature dependence in summer. Evapotranspiration is mostly energy-, rather than moisture-limited in Scandinavia, although moisture dependence becomes more important towards the southern parts of the domain.

Differences between the curves showing the sum of the separate temperature and precipitation change simulations and simulations including changes for both variables illustrate how non-linear processes contribute to the simulated changes in the annual cycle of river discharge. The importance of non-linear processes varies from one GCM-RCM to another, but is generally largest during the time of the spring maximum in river discharge values. A possible explanation for this is that some threshold processes related to snow melting become important when temperature and/or precipitation changes grow large enough. This is indicated by the fact that the difference between the sum of the separate temperature and precipitation effect simulations (long-short dashed lines in Figure 6) and the full non-linear simulation (solid lines in Figure 6) is largest for those GCM-RCM combinations that also have the largest non-linear response in the simulated discharge changes.

Maps for the summer season show noticeable differences in comparison to the winter months. The temperature effect is systematically larger in the southern parts of the domain, where precipitation is projected to decrease. As in winter, DQ shows a greater contribution from precipitation changes than DC, probably due to differences in the evapotranspiration response. These results suggest that the simple delta change scaling likely overestimates the effect of temperature on changes in river discharges through other hydrological processes, which illustrates the importance of the selection of proper statistical tools for downscaling climate model data in hydrological climate change impact studies.

The proper communication of uncertainties in the response of surface hydrology to future climate changes is of primary importance for stakeholders in several socio-economic sectors, such as hydro power production and water management. In this study, the relative importance of uncertainties arising from GCM-RCM and MOS-method differences when simulating changes in future discharges in Scandinavia are studied using the hydrological model HYPE. To gain more insights into the processes which might explain these uncertainties, effects of temperature and precipitation changes on simulated changes in the annual cycle of river discharge were analyzed to assess their relative importance to the simulated hydrological changes and to also reveal how their relative contributions vary between the GCM-RCM and MOS-method combinations. The main results are two-fold:

1. Most of the spread in the changes of the annual cycle of river discharges is explained by GCM-RCM differences caused by the internal variability and model-specific response to external forcing. This result is in line with previous studies, which have shown that GCM-RCM differences generally explain a major part of the spread in daily mean temperature and precipitation projections (Räisänen & Räty 2013; Räty et al. 2014) as well as in simulated future river discharges (Chen et al. 2011a, 2011b). However, the importance of method differences increases towards the tails of the distribution of daily discharge values (Dobler et al. 2012). Also, the magnitude and location of the spring flood peak is sensitive to the selected MOS method. It was found that this behavior is related to the snow accumulation and melting processes in northern parts of the domain in winter and that this depends on whether daily variability is taken into account in the bias correction step or not. Overall, the results imply that the regional-scale spatial variations in the relative contributions of GCM-RCM and MOS-method effects are non-negligible and should be covered when assessing uncertainties in future hydrological changes.

2. The effect of temperature changes is generally larger than the precipitation effect and explains a major part of the changes in the intra-annual variations of river discharges. A similar finding for Alpine tributaries was made by Bosshard et al. (2014). The smaller snow reservoir and earlier timing of snow melt season in winter and in spring together with the increased evapotranspiration in summer probably explain the relatively large importance of temperature changes. The results also show that non-linear processes are important especially in the spring, which has implications on the bias correction step; due to the threshold effects the use of bias correction methods (including the studied ones), which assume stationary biases, can lead to biased results in the future climate. Also, differences in temperature and precipitation effects were generally larger between different GCM-RCMs than between the two delta change methods. This illustrates that the response of HYPE is more sensitive to the selected GCM-RCM and that the inclusion of daily variability to the MOS-adjustment step has a secondary (albeit important) role in modulating the projected changes in the annual cycle of river discharges.

We stress that the results are specific for the selected GCM-RCMs and hydrological model. Furthermore, the overall set of different methods should have optimally been larger, although the selected ones should cover a relatively broad range of uncertainties related to the MOS methods, since a large part of the uncertainty arises from differences between the delta change and bias correction approaches (Räty et al. 2014). Emission scenario uncertainties, which are likely to become increasingly important in relation to other sources of uncertainty at longer time scales, should also be assessed.

Due to the structural limitations of HYPE and uncertainties in the model parameters, care should be taken in the interpretation of the results. The parametric uncertainty, for example, might be in some cases important but is not expected to be the dominant source here (Dobler et al. 2012). To get a better idea of the overall uncertainty in the future changes of surface hydrology, several hydrological models should be used in parallel, as hydrological model differences in simulated future changes can be one of the largest sources of uncertainty at some locations (e.g., Hagemann et al. 2011).

An obvious extension to the present study is to compare the results between single-variable and multi-variable bias correction methods (Vrac & Friederichs 2015), which take co-variability into account. The comparison of joint and univariate bias correction methods in the hydrological modeling step will be targeted in future research.

We thank COST action VALUE (ES1102) for providing funding for the STSM during which the analysis was made. Olle Räty is funded by Vilho, Yrjö and Kalle Väisälä Foundation. The climate model simulations used in this work were funded by the EU FP6 Integrated Project ENSEMBLES (contract number 505539) and the HYPE model setup by the EU FP7 project SWITCHON (grant number 603587). This study has also been supported by the Academy of Finland (project RECAST, decision 140801) and the Academy of Finland Center of Excellence program (project no. 272041).