Effect of modelling scale on the assessment of climate change impact on river runoff

Abstract

The effect of using two distributed hydrological models with different degrees of spatial aggregation on the assessment of climate change impact on river runoff was investigated. Analyses were conducted in the Narew River basin situated in northeast Poland using a global hydrological model (WaterGAP) and a catchment-scale hydrological model (SWAT). Climate change was represented in both models by projected changes in monthly temperature and precipitation between the period 2040–2069 and the baseline period, resulting from two general circulation models: IPSL-CM4 and MIROC3.2, both coupled with the SRES A2 emissions scenario. The degree of consistency between the global and the catchment model was very high for mean annual runoff, and medium for indicators of high and low runoff. It was observed that SWAT generally suggests changes of larger magnitude than WaterGAP for both climate models, but SWAT and WaterGAP were consistent as regards the direction of change in monthly runoff. The results indicate that a global model can be used in Central and Eastern European lowlands to identify hot-spots where a catchment-scale model should be applied to evaluate, e.g. the effectiveness of management options.

Editor D. Koutsoyiannis; Associate editor F.F. Hattermann

Citation Piniewski, M., Voss, F., Bärlund, I., Okruszko, T., and Kundzewicz. Z.W., 2013. Effect of modelling scale on the assessment of climate change impact on river runoff. Hydrological Sciences Journal, 58 (4), 737–754.

Résumé

Nous avons étudié l'effet de l'utilisation de deux modèles hydrologiques distribués selon différents degrés d'agrégation spatiale sur l’évaluation de l'impact du changement climatique sur l’écoulement fluvial. Les analyses ont été effectuées dans le bassin de la rivière Narew située au Nord-Est de la Pologne, à l'aide d'un modèle hydrologique global (WaterGAP) et d'un modèle hydrologique à l’échelle du bassin versant (SWAT). Le changement climatique a été représenté dans les deux modèles par les changements de températures et de précipitations mensuelles entre la période 2040-2069 et la période de référence, projetés par deux modèles de circulation générale : IPSL-CM4 et MIROC3.2, couplés tous les deux avec le scénario d’émission SRES A2. Le degré de cohérence entre le modèle global et le modèle à l’échelle du bassin versant s'est révélé très élevé pour l’écoulement annuel moyen, et moyen pour les indicateurs de hautes et basses eaux. Nous avons observé que SWAT suggère généralement des changements de plus grande ampleur que WaterGAP pour les deux modèles climatiques, mais que les modèles hydrologiques étaient cohérents sur le signe du changement de l’écoulement mensuel. Les résultats indiquent que le modèle global peut être utilisé sur les zones de plaine d'Europe centrale et orientale afin d'identifier les points chauds où le modèle à l’échelle du bassin versant devrait être appliqué pour évaluer, par exemple, l'efficacité des options de gestion.

Key words:

• global hydrological model
• catchment hydrological model
• WaterGAP
• SWAT
• climate change
• Narew River
• Poland

Mots clefs:

• modèle hydrologique global
• modèle à l’échelle du bassin versant
• WaterGAP
• SWAT
• changement climatique
• Fleuve Narew
• Pologne
• Pologne

1 INTRODUCTION

The most recent report of the Intergovernmental Panel on Climate Change (IPCC 2007) states that “warming of the climate system is unequivocal, as is now evident from observations of increases in global air and ocean temperatures, widespread melting of snow and ice, and rising global sea level”. All available general circulation models (GCMs) agree about the further increase of global mean temperature in future projections. Several local Polish studies suggest that the northeast of Poland has been subject to significant temperature increase. Maksymiuk et al. (2008) detected an increasing, statistically significant trend in winter air temperature and a decreasing trend in snow cover thickness in the Biebrza River basin. Marszelewski and Skowron (2006) analysed various indicators related to lake ice cover and winter air temperature and concluded that the recently observed modifications of the ice cover regime of lakes reflect the increase in winter temperature. Trends and projections in precipitation, another key driver of the hydrological system, are less consistent than for temperature. Kundzewicz et al. (2008) argued that precipitation is not adequately simulated by the present climate models. Poland is one of the countries for which climate models largely disagree about future precipitation projections. Most models, however, predict an increase in winter precipitation (Szwed et al. 2010).

The common method for assessment of climate change impacts on hydrology is to use a pre-processed output from one or several GCMs as climatic input to hydrological models. This step of pre-processing, often referred to as “downscaling”, is essential due to the fact that the GCMs remain coarse in spatial resolution and are unable to resolve several sub-grid scale features (Grotch and MacCracken 1991) such as topography, clouds and land use (Fowler et al. 2007). Anagnostopoulos et al. (2010) showed that GCMs, on their own, cannot accurately reconstruct the past even at sub-continental to continental scales, and perform poorly at regional scales. This is one of the reasons why Kundzewicz and Stakhiv (2010) concluded that climate models are not yet “ready for prime time” in water resources management applications. In contrast, the hydrological models widely used by the water science community have undergone decades of peer review, testing and application in a wide range of conditions all over the globe. They have been applied not only in different geographical settings but also at various spatial scales: from hillslopes (Ambroise et al. 1996) to small catchments (Zehe et al. 2001), large river basins (Barthel et al. 2005) and at the global scale (Hanasaki et al. 2010, Haddeland et al. 2011). When we focus only on runoff prediction models, another differentiating feature is the discretization strategy (strongly related to computational requirements): from fully-distributed, grid-element-based models, such as Système Hydrologique Europèen – SHE (Abbott et al. 1986) and its successors such as SHETRAN (Bathurst et al. 1995) and MIKE SHE (Refsgaard and Storm 1995), to semi-distributed models built on the concept of hydrological similarity, such as TOPMODEL (Beven and Kirkby 1979), SWAT (Arnold et al. 1998, Neitsch et al. 2005) or SWIM (Krysanova et al. 1998).

It is often assumed that spatially explicit and process-based models are best suited to predict the effects of changing environmental conditions (Beven and Binley 1992). However, problems with a priori estimation of model parameters make them difficult to apply and instead, semi-distributed models are often argued to be a more practical alternative (Croke et al. 2004). In particular, Gassman et al. (2007) in their comprehensive review of SWAT model applications reported 22 peer-reviewed papers devoted to climate change impact assessment using SWAT (this number has undoubtedly grown since then; e.g. Kingston and Taylor 2010, Parajuli 2010, Zhang et al. 2011), including related models, such as SWIM (Krysanova et al. 1998). These assessments are typically done for small (49.6 km²; Zhang et al. 2011), medium (2098 km²; Kingston and Taylor 2010) and large catchments (13 000–147 423 km²; Huang et al. 2010). For water resources management, climate change impact assessments made at the medium and large scale are the most desired. These scales conform to the spatial units imposed by the Water Framework Directive of the European Union (EC 2000): water districts and catchments of water bodies (in Poland, the latter are grouped into larger units, such as integrated/consolidated water bodies and water balance units; cf. Pusłowska-Tyszewska et al. 2006, Piniewski and Okruszko 2011). When a broader perspective is needed, an alternative to using catchment models in climate change impact studies is to use large-scale (global or continental) models. The examples include: MacPDM (Arnell 1999, Gosling and Arnell 2011), VIC (Nijssen et al. 1997) and WaterGAP (Alcamo et al. 2003, Döll et al. 2003). Haddeland et al. (2011) recently provided a comprehensive inter-comparison study, in which six land surface models and five global hydrological models were compared.

The future of Europe's waters will be influenced by a combination of many important environmental and socio-economic drivers. In the project “Water Scenarios for Europe and Neighbouring States” (SCENES; Kämäri et al. 2008) a set of qualitative and quantitative scenarios has been developed to describe freshwater futures up to 2050 in a pan-European perspective, covering the area from the Mediterranean rim countries and reaching from the Caucasus to the White Sea in the east. The Narew River basin used in this study was selected as one of the SCENES pilot areas.

The objective of this study is to analyse the effect of modelling scale (using semi-distributed hydrological models with different degrees of spatial aggregation) on the assessment of climate change impact on river runoff. Two models were selected for the comparison: the global hydrological model WaterGAP (Water: A Global Analysis and Prognosis; Alcamo et al. 2003, Döll et al. 2003) and a catchment-scale hydrological model, SWAT (Soil & Water Assessment Tool; Arnold et al. 1998, Neitsch et al. 2005). Consistent climate change signals derived from two GCMs for the time period 2040–2069 drove the hydrological models and generic hydrological indicators were evaluated, such as mean annual runoff, high and low monthly runoff as well as indicators describing the seasonal cycle. An implicit assumption was that, due to more spatially explicit catchment representation, SWAT can be used as a reference to evaluate WaterGAP as a tool to quantify hydrological indicators related to climate change at a large catchment level. In this respect it is worth emphasizing that the aim of this study was not to analyse which model performs better in the Narew basin, as such competition would be highly unfair for WaterGAP due to the different model input data, different set-ups and different calibration strategies. In this study WaterGAP was not set up intentionally for the Narew basin but was applied with its parameters set at the continental scale and calibrated using river flow data from the Global Runoff Data Centre stations across Europe. This included a station on the Narew River, Ostrołęka (GRDC ID 6458810). In contrast, the SWAT set-up was tailored for the study basin and its calibration involved numerous gauging stations and discharge data with finer temporal resolution (Piniewski and Okruszko 2011). Hence, WaterGAP and SWAT were applied and evaluated in this study as a global and a catchment model, respectively.

Comparison studies of this kind (i.e. between global-scale and catchment-scale models) are rare in the literature. For example, the focus in Gosling et al. (2011) was mainly on comparing climate model uncertainty with hydrological model uncertainty and due to the magnitude of their study (six catchment models, six study areas, seven GCMs), rather little was reported on the explanation for the different responses by global and catchment models. In this study, the models were applied in a single study area, the Narew basin in the northeast of Poland, which allowed a focus more on explaining the discovered differences rather than comparing different types of uncertainty.

2 MATERIALS AND METHODS

2.1 Study area

The River Narew situated in northeast Poland (Fig. 1) is a right-bank tributary of the River Vistula; its total drainage area upstream from its mouth is approx. 75 000 km². However, in this study attention is focused only on the part of the Narew basin situated upstream of the Zambski Kościelne (hereafter referred to as “Zambski”) gauging station, which occupies approx. 28 000 km² and is beyond the reach of backwater effects from Lake Zegrzyńskie. Approximately 5% of this area, in its upstream part, lies in western Belarus, i.e. outside the territory of the Republic of Poland.

Fig. 1 Map of the study area.

The Narew is a lowland river and its basin is characterized by a mean altitude of 136 m a.s.l. and flat topography. This region is located in the temperate climatic zone with moderately warm summers (mean July temperature: 17°C) and cool winters (mean January temperature: –3°C), with annual mean precipitation of approx. 600 mm occurring mostly in summer months. Snowmelt occurs usually in early spring causing peak runoff in the rivers. Soils are predominantly loamy sands and sandy loams, with significant areas of organic soils in river valleys. Agriculture is the dominant land use in the basin: 46% of the area is used as arable land, 17% as meadows and pastures, and 33% is occupied by forests. The remaining 4% is occupied by wetlands, lakes and urban areas.

The Narew basin is a good study area for purely hydrological research because it is not greatly impacted on by anthropogenic pressure. The population density is estimated as 59 people per km², which is low relative to the average density of 119 people per km² for the whole of Poland. Only 53.3% of the population live in cities and towns, cf. the 61.3% urban population of all Poland. There is only one city with a population above 100 000 inhabitants (the city of Białystok) whose surface and sub-surface water abstractions as well as the treatment plant discharges influence the hydrograph of the Supraśl River; however, their impact on the Narew is negligible. No heavy industry is present in the study area; agriculture, food and wood production and tourism are the main sources of income for the inhabitants. For further description of the Narew basin see Okruszko and Giełczewski (2004).

2.2 Hydrological models

The catchment-scale model used in this study was the SWAT model, developed at the Grassland, Soil and Water Research Laboratory in Temple, TX, USA. It is a semi-distributed catchment model developed mainly for meso- and large-scale applications, which can be applied to catchments of any size (from very small to large, see e.g. Gassman et al. 2007), provided that it is fed with the necessary catchment-specific input data. The global-scale model used was the WaterGAP model, developed at the Centre for Environmental Systems Research, University of Kassel, Germany. It is a global hydrological model of water availability and water use that comprises two main components: Global Hydrology Model and Global Water Use Model. In this study the latter component was not applied, because, as mentioned previously, water use is not a significant issue in the Narew basin.

Comparison of SWAT and WaterGAP in terms of their modelling approaches and input data used for the Narew case study shows both similarities and differences between them (Table 1). The SWAT model is a physically-based tool, although it uses many conceptual modelling approaches such as the US SCS curve number method. Instead of using grid cells, SWAT subdivides a river basin into sub-catchments connected by a river network and further delineates hydrological response units (HRUs), obtained through overlay of land use, soil and slope maps in each sub-catchment. It is worth noting that the HRUs are lumped (i.e. non-spatially distributed) units. The current configuration of SWAT for the Narew basin uses 151 sub-catchments and 1131 HRUs.

Table 1 Comparison of SWAT and WaterGAP modelling concepts/approaches and input data used

Aspect	SWAT	WG
Modelling approach
Basic unit	Hydrologic Response Unit	5′ × 5′ grid cell
Potential evapotranspiration (PET)	Penman-Monteith method	Priestley-Taylor method
Actual evapotranspiration (AET)	Evaporation from canopy + sublimation + plant water uptake + soil evaporation	Evaporation from canopy + sublimation + evapotranspiration from vegetated soil
Snowmelt	Degree-day method
Surface runoff	Modified SCS curve number method	HBV method
Redistribution in soil	Storage routing method between up to 10 soil layers	No redistribution, one soil layer
Soil water content	Allowed range of variation from the absolute zero to saturation	Allowed range of variation from the wilting point to field capacity
Groundwater storage	Two groundwater storages (shallow unconfined and deep confined)	One groundwater storage
Baseflow	Recession constant method	Linear storage equation
Flood routing	Variable storage coefficient method	Linear storage equation
Input data
Drainage topology	Based on 30-m resolution DEM and stream network map	Based on the global drainage direction map DDM5
Land-use map	Corine Land Cover 2000
Soil map	Based on ∼3400 benchmark soil profiles in the Narew basin	FAO
Climate	Daily data from 12 precipitation stations and 7 climate stations (temperature) + daily data from MARS-STAT database for other variables	Monthly data from the CRU 10′ resolution global data set

The model features compared in Table 1 are general and do not cover many substantial differences in the parameterizations of hydrological processes. First, the same processes can be modelled using different methods (e.g. potential evapotranspiration – PET) and thus require different parameters; secondly, even if a given process (e.g. snowmelt) is modelled using the same method, the values of the associated parameters might be different.

The current version of WaterGAP works with resolution of 5 arc minutes, which is one of the finest resolutions of state-of-the-art global models. The mean HRU area in SWAT of approx. 24 km² represents a finer resolution than that used in WaterGAP (~51 km²). The SWAT sub-basins and reaches, and the WaterGAP grid mesh are illustrated in Fig. 2.

Fig. 2 Spatial discretization of the Narew basin in SWAT and the WaterGAP grid.

It is to be noted that SWAT, a catchment model, was set up, calibrated and validated intentionally for the Narew basin, whereas WaterGAP was used in its global set-up. In particular, its parameters were not fine-tuned to better represent the study area. Four of the WaterGAP global calibration points were situated in the Vistula basin; one was within the Narew basin on the Lower Narew (Ostrołęka, Fig. 1); the other three were outside the Narew basin (the River San at Radomyśl, the River Vistula at Szczucin and Warszawa). Discharge values for calibration were obtained from the Global Runoff Data Centre. Here we used the WaterGAP 3.0 model version which is an upgrade from version 2.1 as applied by Alcamo et al. (2003) and Döll et al. (2003). One of its main features is the enhanced spatial resolution which was improved from 0.5° to 5′ grid cell size. For SWAT, Piniewski and Okruszko (2011) performed a spatially distributed calibration and validation in the Narew basin for the period 2001–2008 using SWAT2005 with the GIS interface ArcSWAT 2.3, which set the basis for future modelling activities using this tool. In this study we used the same version of SWAT and the model set-up, which was recalibrated and revalidated for the period 1976–2000. As reported in Piniewski and Okruszko (2011), eight SWAT parameters with the highest sensitivities were selected for auto-calibration performed using the ParaSol method (van Griensven and Meixner 2007). The three most sensitive parameters were: ESCO (soil evaporation compensation factor), CN2 (curve number for moisture conditions II) and ALPHA_BF (baseflow alpha factor). The main calibration criterion was a Nash-Sutcliffe efficiency above 0.5 for daily flows; however, other aspects such as maintaining the model bias below 25% and visual inspection of low and high flow modelling were also taken into account. The calibration criteria were met for all 11 calibration gauges. However, spatial validation performed at 12 additional upstream gauges demonstrated that the model performance is significantly lower at smaller spatial scales.

2.3 Climatic input data

The climate data used to drive the hydrological models can be divided into: (a) the observed data from the period 1976–2000 representing the present-day climate, hereafter referred to as the baseline; and (b) the projected climate change data downscaled from two GCMs for the period 2040–2069 representing the future climate, hereafter referred to as the 2050s. The SWAT and WaterGAP models used different data sources for the baseline period and consistent climate change forcing for the 2050s.

2.3.1 Baseline

In WaterGAP, monthly values of the climate variables from the 10-min resolution CRU TS 1.2 data set (Mitchell el al. 2004) were used. The time series of the following variables were used: precipitation, air temperature, cloudiness and wet day frequency. Because WaterGAP simulates river discharges with a daily time step, the climate input data needed to be downscaled from monthly to daily values. Downscaling procedures are implicitly implemented in WaterGAP and were run during the simulations. With this, temperature and cloudiness were downscaled with a cubic-spline-function between the monthly averages, which were assigned to the middle of each month. Precipitation was distributed equally over the number of wet days per month which were distributed within the month using a two-state, first-order Markov Chain, applying the parameterization according to Geng et al. (1986).

In contrast, daily station data from the Polish Institute of Meteorology and Water Management network were used as the climate input for precipitation and temperature in SWAT. Precipitation data came from 12 stations and the temperature data were taken from seven stations. Missing values were filled either by manual interpolation or with values taken from the public domain MARS-STAT database (van der Goot and Orlandi 2003). This data source, which provides daily time series in a 25-km grid for the whole of Europe, was also used to provide daily data for further climate variables required in SWAT: wind speed, relative humidity and solar radiation. As SWAT does not perform any interpolation of climate data, precipitation and temperature were interpolated to the sub-basin level outside ArcSWAT, using the Thiessen polygon method.

It is evident that the daily time scale of the climate data used in SWAT is more adequate than the monthly time series of the original CRU data set used in WaterGAP that was internally downscaled to the daily time scale leading to a loss in daily weather dynamics. However, it is difficult to say which of the models used the more appropriate spatial resolution of climate data. Even though the 10-min resolution of the CRU 1.2 data set is theoretically much higher than the resolution of the climate input used in SWAT, one has to bear in mind that CRU data are based on interpolation from station data and hence the quality of SWAT climate input data should not be worse than the quality of the CRU data set. This assumption was verified by comparing annual basin-averaged mean temperature and precipitation series (Fig. 3) as well as mean monthly values of temperature and precipitation (Fig. 4). Note that SWAT uses daily maximum and minimum temperature as the climatic input, so in order to enable direct comparison of this variable with that from WaterGAP we estimated daily mean temperature as the arithmetic mean of daily maximum and minimum temperature.

Fig. 3 Annual basin-averaged mean (a) temperature and (b) precipitation in the baseline period.

Fig. 4 Mean monthly basin-averaged (a) temperature and (b) precipitation in the baseline period.

The mean annual temperature time series used within SWAT and WaterGAP are very well correlated, with R ² equal to 0.94 (Fig. 3(a)). Long-term mean temperature used within WaterGAP is approx. 0.3°C higher than that used within SWAT. These higher temperature values can be observed especially in spring and summer (Fig. 4(a)). Nevertheless, the differences between SWAT and WaterGAP temperature inputs are rather small, and are partly explained by the indirect method of comparison as well as the different data sources.

Annual precipitation series are also very well correlated (R ² equal to 0.82) and there is hardly any long-term bias between the models (Fig. 3(b)). The greatest difference between WaterGAP and SWAT (71 mm) was observed in 1995. The monthly differences are also rather small (Fig. 4(b)), which suggests that mean areal precipitation derived from the CRU data set is comparable to precipitation derived from station data.

2.3.2 Projections for 2050s

Consistent climate change signals of two types were applied to both hydrological models. The signals were derived from the output of two different GCMs: IPSL-CM4 from the Institute Pierre Simon Laplace, France (Marti et al. 2006) and MIROC 3.2 from the Centre for Climate System Research, University of Tokyo, Japan (Hasumi and Emori 2004), both forced by the SRES-A2 emissions scenario (IPCC 2007). The development of socio-economic scenarios within the SCENES project was a stakeholder-driven process (Kok et al. 2011). Climate scenario development was, however, not part of the project and thus available GCM–emissions scenario combinations were selected. Here the stakeholders played a key role in finally concentrating on the IPCC SRES-A2 scenario emphasizing the trigger role of climate change in all SCENES storylines. The analysis performed at a pan-European scale in the SCENES project revealed that across the range of GCMs driven by the A2 scenario, climate projection by IPSL-CM4 is dry and by MIROC 3.2 is wet, whereas both project an increase in temperature.

Monthly precipitation and temperature derived from GCMs needed to be downscaled to a finer spatial resolution because their original resolution was too coarse relative to that of the catchment processes simulated by the hydrological models. To this end, first, a simple bilinear interpolation approach was applied to downscale GCM data to the resolution of the WaterGAP grid cells.

It is well known that present climate models contain considerable biases in their climatology and do not fit gridded station data well (Kundzewicz and Stakhiv 2010). To reduce the GCM biases, various “bias correction” methods were developed. In this study we applied the delta-change approach. Based on the assumption that GCMs more accurately simulate relative change than absolute values, we assumed a constant bias through time (Fowler et al. 2007). In this method, the delta change factors (DCFs) are calculated at the monthly time scale using the future (here 2040–2069) and present (1976–2000) GCM output. For temperature (additive variable), change factors are defined as the arithmetic difference between the future and present long-term means, whereas for precipitation (multiplicative variable) they are defined as the future to present long-term mean ratios.

Due to the obvious differences between the hydrological models, the final versions of climate input representing the 2050s (the middle decade from the climatic standard normal, 2040–2069) were derived in both models in a slightly different way. In WaterGAP, gridded DCFs were first added to (in the case of temperature) or multiplied by (in the case of precipitation) the monthly time series for the respective grid cells. Next, the number of wet days per month and the cloudiness were taken from the baseline period, in order to downscale monthly climate to daily climate as described above. In SWAT, there is an option of running climate change scenarios by defining monthly change factors at the sub-basin level (parameters RFINC and TMPINC in .sub files) whereby the model automatically creates new daily time series associated with scenarios by scaling the observed climate data for the baseline. In order to use this option, the DCFs calculated beforehand at the WaterGAP grid scale were averaged over the SWAT sub-catchments. On average, there were more than three grid cells per single sub-catchment (cf. Fig. 2 for the map of the modelling units).

Both climate models predict a similar increase in mean annual temperature, but the seasonal variability of this increase is different (Fig. 5(a)). For instance, in April and November the increase in temperature projected by IPSL-CM4 is over 1°C greater than the one projected by MIROC3.2. As regards precipitation, there is hardly any agreement between the two GCMs (Fig. 5(b)). According to IPSL-CM4, relative changes in precipitation do not exceed ±25% for any month and mean annual precipitation is almost the same as in the baseline. According to MIROC3.2, there is an 11% increase in annual precipitation and quite a large variability of within-year changes. There is a very different, hence problematic, behaviour of model projections in two adjacent months: July (15% decrease) and August (44% increase). Two periods can be found where MIROC3.2 projects a substantial increase and IPSL-CM4 little change or even a decrease in precipitation: (a) from March to April; and (b) from August to October.

Fig. 5 Basin-averaged changes in (a) temperature and (b) precipitation from IPSL-CM4 and MIROC3.2.

2.4 Hydrological indicators

Standard goodness-of-fit measures were used to assess the model behaviour in the baseline period. The Nash-Sutcliffe efficiency (NSE) measures the relative magnitude of the residual variance compared to the observed data variance (Nash and Sutcliffe 1970), whilst the coefficient of determination (R ²) describes the degree of co-linearity of the measured and modelled time series (Moriasi et al. 2007). Percent bias is one of the widely used error indices and measures the average tendency of the modelled data to be larger or smaller than the observed data (Gupta et al. 1999).

The response of the hydrological models to the climate change forcing was assessed by relating the modelled runoff from scenario simulations with the runoff from the respective baseline simulations. The impact assessment was done on three levels:

a.	Impact on the mean annual runoff. Here, one indicator was used: the absolute change in mean annual runoff relative to baseline.
b.	Impact on the monthly extreme (high/low) runoff. Here, in the first step, the empirical flow duration curves (EFDCs) were used to make a visual inspection of the extreme parts of the frequency distribution of monthly runoff (Smakhtin 2001). In the second step, two particular indicators (single points from the EFDCs) were reported: the absolute changes in monthly Q10 and Q90 (defined as the monthly runoff exceeded for 10% and 90% of the time, respectively), relative to the baseline period.
c.	Impact on the seasonal cycle of runoff. Here, in the first step, monthly runoff hydrographs simulated by SWAT and WaterGAP for the baseline and under two climate scenarios were analysed in order to interpret the main hydrograph alterations. In the second step, the absolute changes in mean monthly runoff relative to baseline were analysed in order to detect the seasonal pattern in the differences between the future scenarios and baseline conditions and to measure the mean sensitivity of both models to the climate change signals.

All the above-mentioned indicators (apart from the EFDC which was reported for Zambski only) were evaluated at three sites within the catchment: at the basin outlet (Zambski), at the mouth of the Biebrza (Burzyn) and in the upper Narew at Suraż (Fig. 2).

3 RESULTS

Despite the fact that the main objective of our study is not to evaluate model performance during the baseline period, it is an essential step before analysing the climate change impact on hydrological indicators. The analysis of model behaviour in the baseline period can bring an insight into the process of explaining differences between the model behaviours in the future.

3.1 Baseline

WaterGAP tends to underestimate mean monthly runoff in the baseline period at the main catchment outlet (Zambski gauge) and two internal outlets (Fig. 1) by 12% to 24%, whilst SWAT neither underestimates nor overestimates mean monthly runoff by more than 8% (Table 2). As expected, the SWAT-based estimates of Q10 and Q90 are closer to the measured ones than the WaterGAP-based estimates, apart from Q90 at Burzyn. Performance of SWAT at Zambski is apparently better than the performance at Burzyn and Suraż, which is very likely linked to the size of the upstream catchment area (Piniewski and Okruszko 2011). In the case of WaterGAP, this spatial relationship does not exist: the best performance is observed at Burzyn and not at Zambski.

Table 2 SWAT and WaterGAP monthly runoff simulation statistics (mm) and goodness-of-fit measures in the baseline

Gauge	Area (km^2)	Category	Qmean	Q10	Q90	NSE	R ^2	Bias (%)
Zambski	27 500	Measured	13.4	22.6	6.3
		SWAT	13.6	23.5	5.6	0.72	0.73	−2
		WaterGAP	11.7	20.8	4.9	0.35	0.50	12
Burzyn	6800	Measured	14.6	24.9	5.6
		SWAT	14.4	27.6	3.8	0.59	0.61	1
		WaterGAP	11.1	20.6	5.1	0.47	0.58	24
Suraż	3280	Measured	12.6	25.9	4.2
		SWAT	13.6	30.6	2.1	0.61	0.71	−8
		WaterGAP	10.1	21.1	2.0	0.30	0.45	20

The SWAT model captures monthly variability better than WaterGAP in all three locations (Fig. 6). Peak runoff in WaterGAP occurs as often in March as in April, whereas according to the measured data, the peaks occur much more frequently in April. Both models underestimate peak runoff (with one exception of SWAT at Suraż) by 2.8–3.2 mm in the case of SWAT and 2.0–7.1 mm in the case of WaterGAP. As regards the low flow period in the Narew basin, it lasts from July to September. In SWAT this period is shifted one month ahead, whereas in WaterGAP it lasts from September to February, which is arguably the greatest deficiency of the hydrograph simulation by WaterGAP. The largest issue with the SWAT-modelled hydrograph is, in our opinion, that the falling limb is decreasing too gently. It causes overestimation of runoff from May to July, as most clearly seen at Suraż (Fig. 6(c)).

Fig. 6 Mean measured and simulated monthly runoff in the baseline at the three locations analysed.

Correlation of the annual time series of various water balance components simulated by both models (only measured values for runoff could be included) is illustrated in Fig. 7. The R ² of the correlations of the SWAT- and WaterGAP-based estimates of annual runoff with measurements are 0.78 and 0.51, respectively, and the correlation between the two models is good (R ² = 0.75). Other water balance components are either moderately (PET, R ² = 0.52) or weakly correlated (for actual evapotranspiration, AET, and soil water content, R ² equals 0.22 and 0.37, respectively). Note that to derive PET, SWAT uses Penman-Monteith and WaterGAP uses Priestley-Taylor (Table 1). There is a bias in PET time series, especially in the first seven years of the simulation period, when SWAT-based PET estimates are ~100 mm higher than WaterGAP-based estimates. WaterGAP simulates considerably higher AET than SWAT (average difference 44 mm) which partly explains its underestimation of runoff compared to SWAT, by 22 mm on average. Year-to-year soil water storage changes are presented in Fig. 7(c), instead of actual soil water content, because the latter variable is difficult to compare directly between the models. The magnitude of soil water storage changes is comparable between the models and does not exceed 20 mm in terms of absolute values.

Fig. 7 Annual time series of the basin-averaged water balance components in the baseline period, as simulated by WaterGAP and SWAT: (a) potential evapotranspiration; (b) actual evapotranspiration; (c) storage change in soil water (year-to-year); and (d) runoff.

The analysis of the monthly dynamics of these water balance components can help explain the observed differences in runoff simulation (Fig. 8). Estimates of PET by WaterGAP are higher than by SWAT in the hottest months of the year and lower during the rest of the year. WaterGAP simulates significantly (51 mm) higher AET than SWAT in May and June, which is reflected in the decline of soil water content in these months by 72 mm in WaterGAP and only by 17 mm in SWAT. The decrease in soil saturation estimated by WaterGAP lasts until September, which is a potential reason for the underestimation of runoff by WaterGAP observed in autumn and continuing until February.

Fig. 8 Basin-averaged monthly dynamics of the water balance components in the baseline period, as simulated by WaterGAP and SWAT: (a) potential evapotranspiration; (b) actual evapotranspiration; (c) storage change in soil water (month-to-month); (d) runoff.

3.2 Hydrological model responses to climate change forcing

3.2.1 Mean annual runoff

Comparing the change in mean annual runoff of the simulations to baseline, there is a large difference between the results driven by IPSL-CM4 and MIROC3.2, but a negligible difference between the results obtained for SWAT and WaterGAP driven by the same climate model at all selected locations (Fig. 9). The largest difference between SWAT- and WaterGAP-based estimates of change in runoff is for IPSL-CM4 at Suraż where the runoff decrease according to SWAT would be 41.2 mm and according to WaterGAP, 27.8 mm. However, the sign of the projected change is the same in each case. It is noteworthy that, for all sites, the differences between the results of a hydrological model driven by two climate models are greater than the differences between the results of two hydrological models driven by one climate model. Hence, the climate scenarios largely contribute to the uncertainty of findings.

Fig. 9 Absolute changes in mean annual runoff relative to baseline under two GCMs as simulated by SWAT and WaterGAP at Zambski, Burzyn and Suraż.

3.2.2 High and low monthly runoff

The EFDC (Fig. 10) indicates a decrease in both high and low runoff under IPSL-CM4 for both SWAT and WaterGAP at any exceedence level. The magnitude of this decrease is variable; however, at the exceedence levels of 5–10% the consistency between SWAT and WaterGAP is higher than at exceedence levels below 5% (for the low runoff part there is no clear relation in this regard). In the case of MIROC3.2, SWAT suggests an increase in high runoff at any exceedence level, whereas WaterGAP suggests a negligible change in runoff at exceedence levels in the range 7–10% and a decrease below 7%. The low runoff part of the EFDC shows that under MIROC3.2 the WaterGAP model suggests an increase in runoff at any exceedence level, whereas SWAT suggests a small increase at exceedence levels between 90 and 91% and a negligible change above 91%. Overall, the analysis of the EFDCs shows that the consistency between SWAT and WaterGAP is higher for runoff corresponding to less extreme exceedence levels. Hence, hereafter we will focus on Q10 as the high runoff indicator and Q90 as the low runoff indicator.

Fig. 10 Empirical flow duration curves (EFDCs) of the monthly runoff simulated by SWAT and WaterGAP for the baseline and two climate scenarios at Zambski: (a) high runoff; and (b) low runoff.

The diversity in the change of Q10 and Q90 due to the selected GCMs with regard to the baseline is larger than for the annual runoff (Fig. 11; note that this figure shows monthly and not annual runoff, cf. Fig. 9). For Q10 at Zambski and Burzyn, IPSL-CM4 forcing causes greater decrease in the WaterGAP model than in the SWAT model, whilst at Suraż the decrease rate is greater in SWAT. The MIROC3.2 forcing causes an increase in SWAT and a negligible change in WaterGAP. In the case of Q90, for IPSL-CM4 forcing, SWAT suggests a larger decrease than WaterGAP, whereas for MIROC3.2 the results are not spatially consistent: at Zambski both models suggest an increase in runoff, whereas at Burzyn and Suraż WaterGAP also shows an increase, whilst SWAT shows a decrease. It is worth noting that most of projected changes in runoff are considerable relative to the measured Q90 (6.3, 5.6 and 4.2 mm for Zambski, Burzyn and Suraż, respectively).

Fig. 11 Absolute changes in monthly (a)-(c) Q10 and (d)-(f) Q90 relative to baseline under two GCMs as simulated by SWAT and WaterGAP at Zambski, Burzyn and Suraż.

The differences in low and high runoff are greater between climate scenarios than between hydrological models (Figs 10 and 11), as in the mean annual runoff case.

3.2.3 The seasonal cycle

The projected seasonal cycle of runoff simulated by the hydrological models, illustrated in Fig. 12 (baseline runoff is plotted for comparison), gives a general impression of the hydrograph alteration caused by the climate change forcing. There is a consistency between the hydrological models under both climate scenarios that peak monthly runoff will shift from April to March in all cases, except one, the SWAT-MIROC3.2-Burzyn combination. In the latter case, January is the month with peak runoff; however, the difference between January and March is only 0.3 mm. It is equally worth noting that under the IPSL-CM4 climate scenario, not only a shift in timing is observed but also a substantial decrease in peak runoff at all analysed sites and for both models. Under the MIROC3.2 climate scenario, SWAT shows a moderate decrease in peak runoff, and WaterGAP shows negligible change.

Fig. 12 Monthly runoff hydrographs simulated by SWAT and WaterGAP for the baseline and under two climate scenarios.

The IPSL-CM4 climate model forcing is also likely to significantly alter the hydrographs in their low runoff part (Fig. 12). Under this scenario, according to simulations by SWAT, in the period June–November, runoff will be lower than the minimum SWAT-modelled baseline monthly runoff at all sites (at Suraż between July and November). According to simulations with WaterGAP, runoff will be lower than the minimum WaterGAP-modelled baseline monthly runoff for the period between August (or September in the case of Suraż) and November. However, it should be recalled that simulation of the low runoff period in the baseline was less accurate in WaterGAP than in SWAT (cf. Fig. 6).

Figure 13 gives further insight to the seasonal aspects of runoff as it presents the absolute deviations from baseline for each hydrological model, each climate model (GCM), and each site. Two observations are noteworthy:

Fig. 13 Absolute changes in mean monthly runoff relative to baseline under two GCMs as simulated by SWAT and WaterGAP at Zambski, Burzyn and Suraż.

a.

With a few exceptions, the models are generally consistent in showing the direction of change in mean monthly runoff. Lack of consistency in the sign of change occurred in only four out of 72 cases (neglecting very small changes, <0.2 mm).

b.

The differences between changes simulated by SWAT and WaterGAP for a given GCM are generally smaller than the differences between changes simulated by a given model forced by IPSL-CM4 or MIROC3.2. The largest difference between the departures from baseline simulated by SWAT and WaterGAP under a given climate scenario equals 5.7 mm. For the absolute changes, in four out of six cases the largest differences occur in March.

Analysis of the results from Fig. 13 in relation to the climate forcing data illustrated in Fig. 5 results in the following points:

a.	A uniform reaction of both models and both climate scenarios can be observed in April at all sites. This particular consistency between the models can be explained by the fact that, regardless of the different projections of precipitation change, the high temperature increase projected in winter by both models accelerates the occurrence of peaks. Hence in April, which was the peak runoff month in the baseline, the hydrograph is already decreasing.
b.	MIROC3.2 suggests an increase in temperature between May and June by 3–3.5°C and a relatively small change in precipitation. This drives SWAT, presumably due to increased evapotranspiration, to decrease the total runoff at Zambski in this period by 5.7 mm relative to the baseline, whilst the change in runoff in WaterGAP is negligible. Figure 8 suggests that this might be due to significant overestimation of AET by WaterGAP in the baseline in May and June.
c.	For the August–November period the total increase in precipitation according to MIROC3.2 is 53 mm and the temperature increase stays in the range 2.5–3.5°C. This drives SWAT to increase the total runoff in this period by 8.4 mm compared to the baseline, whilst the increase in WaterGAP equals 3 mm only.

The above observations indicate that SWAT is more sensitive to various seasonal climate change signals than WaterGAP. Results reported in Table 3 confirm this hypothesis. It is interesting to note that: (a) this measure of sensitivity is higher for the MIROC3.2 model than for the IPSL-CM4 model; and (b) in the case of SWAT it is much higher for the sub-catchments than for the whole basin, while this is not the case for WaterGAP. This is the reason why the hydrological models’ inconsistency in assessing the effect of climate change on monthly runoff is larger at Burzyn and Suraż than at Zambski. Indeed, the number of months for which the differences between the absolute changes simulated by SWAT and WaterGAP for any GCM do not exceed 1 mm (in terms of the absolute values) are equal to 9, 2 and 3 for Zambski, Burzyn and Suraż, respectively. The number of months for which the same characteristics exceed 2 mm are 5, 15 and 11, respectively.

Table 3 The averages of the absolute changes in monthly runoff (mm) for all combinations of GCMs, hydrological models and sites

Location	IPSL-CM4		MIROC3.2
	SWAT	WaterGAP	SWAT	WaterGAP
Zambski	3.3	2.9	3.3	2.1
Burzyn	4.7	2.8	4.5	2.0
Suraż	4.9	3.3	4.6	2.2

4 DISCUSSION

The results of our analysis of the global and catchment-scale model responses to the same climate change signal indicate that:

a.	SWAT and WaterGAP were very consistent in showing the direction and quantifying the magnitude of future change in mean annual runoff due to climate change.
b.	The consistency in identifying the high (Q10) and low (Q90) monthly runoff change was not as good as for the mean annual runoff. It was quite often observed that when one model showed a negligible change in these indicators, the other showed at least medium change. For more extreme indicators (e.g. Q5 and Q95) the difference between SWAT- and WaterGAP-based estimates was even larger (Fig. 10).
c.	Some patterns of change in the seasonal cycle of runoff were comparable in both models (e.g. earlier occurrence of peak runoff, large decrease in April runoff) while others were not (e.g. different responses to the August–November precipitation increase from MIROC3.2). The magnitudes of projected seasonal changes varied significantly, SWAT showing overall more sensitivity to climate change than WaterGAP.

Our interpretation of these results is that the modelling scale does not have much influence on the assessment of simple indicators and general descriptive patterns, but when it comes to more detailed indicators and in particular their magnitudes, the impact of the modelling scale is visible. This partly corresponds to the observation pointed out by several authors (Gosling et al. 2011, Hughes et al. 2011, Nóbrega et al. 2011) that the mean annual runoff can mask considerably greater seasonal variations which are of high importance to water management.

As regards the potential reasons for the differences between simulations by SWAT and WaterGAP in climate change impact assessment, it is not straightforward to discriminate between the different model behaviour in the baseline and the different model reactions to the climate change forcing. As catchment-specific calibration was not performed for the global model, it was not surprising to observe generally better behaviour of the catchment model in the baseline. At present (and very likely for the near future), global models such as WaterGAP are not specifically calibrated for catchments of the size of the Narew. Hence, an important question emerges: which process descriptions/parameterizations in WaterGAP should be rethought in order to reduce the uncertainty in climate change impact assessments? The same question should apply to SWAT; however, in this study we tacitly assume, since SWAT performed better in the baseline, that its results are more reliable and can be used as a benchmark for WaterGAP.

The comparison of the annual time series (Fig. 7) and the seasonal dynamics (Fig. 8) of various water balance components revealed a large difference between SWAT- and WaterGAP-based estimates of actual evapotranspiration (AET) and soil water content. We suppose that WaterGAP actually overestimates AET in May and June. This is consistent with a large decrease in soil water content in these months compared to SWAT. We expect that this results in too little soil moisture content in summer months and in consequence, as total runoff simulated in WaterGAP is a nonlinear function of soil moisture (Bergström 1995, Döll 2003), in underestimation of runoff starting from September and lasting until the soils are completely rewetted (i.e. until February).

The above considerations suggest that either the main parameters controlling the vertical soil water balance in WaterGAP should be reconsidered, or the process description itself should be rethought. Because the methods used for estimation of soil water balance components in WaterGAP are well established and used in many other models, such as HBV (Bergström 1995), one should rather focus on the parameters. In particular, three parameters may be critical, namely: soil depth, set to 1 m in WaterGAP which may be too small; total available water capacity within the effective root zone (S_smax); and runoff coefficient (γ), which is a WaterGAP calibration parameter (Döll 2003). This statement is not restricted only to the Narew basin, but should apply also to other lowland river basins lying in the same climatic zone.

Differences in snowmelt estimation might be another reason for differences between SWAT- and WaterGAP-based estimates, especially those related to winter and spring runoff generation. It was observed that peak runoff in the baseline period occurred quicker in WaterGAP than in SWAT and in the observation records (Fig. 6), which was likely caused by the fact that snow cover was thawing quicker in WaterGAP. Both models use a degree-day approach to estimate snowmelt. However, although the snowmelt base temperature was set to 0°C in both models, two other important parameters controlling snowmelt were set to different values: Snowfall temperature was set to 1°C in SWAT and 0°C in WaterGAP. Degree-day factor (DDF) in WaterGAP was set to range from 1.5 to 7 mm d⁻¹ °C ⁻¹, depending on land cover type, whereas in SWAT this parameter ranged between 0.5 (21 December) and 1.5 (21 June), as a unique value for the whole basin, like all snow-related parameters in SWAT. Higher DDFs in WaterGAP induced quicker snowmelt, and since there was less accumulated snow (due to lower snowfall temperature), peak runoff occurred up to 1 month in advance. Verzano and Menzel (2009) compared hydrographs modelled in WaterGAP with measured ones in two large basins situated in the Alps and the Scandinavian mountains and also found that WaterGAP underestimated winter runoff, but the magnitude of underestimation was smaller. It requires further studies to examine if improvement of the estimation of peak runoff occurrence in WaterGAP could be reached by manipulating the snow-related parameters. Another possible reason for too rapid snowmelt in WaterGAP could be that the global hydrological model internally generates daily climate input time series from the monthly CRU data set, which in the case of temperature, and especially temperatures around snowmelt events, may affect simulated runoff stronger than in any other season of the year.

Although differences between SWAT- and WaterGAP-based estimates in assessing the effect of climate change on runoff are undeniable, it is worth noting that the inter-GCM differences are even larger and this is where the uncertainty is dominant. In particular, the largest difference between estimates of the mean annual runoff using IPSL-CM4 and MIROC3.2 is equal to 56 mm, whereas differences between the SWAT- and WaterGAP-based estimates do not exceed 13 mm (Fig. 9). It is also relevant to note that regardless of whether it was a decrease or an increase in the monthly runoff due to the climate change forcing, the reaction of SWAT was in 63 out of 72 cases (2 models · 3 sites · 12 months) more pronounced than in WaterGAP (Fig. 13 and Table 2). The SWAT model is equally sensitive to climate change forcing from IPSL-CM4 and MIROC3.2, whereas the WaterGAP model shows significantly lower sensitivity to the latter model. Since the difference between the climate models is mainly in future precipitation changes, we suggest that there is a mechanism in WaterGAP that triggers a more pronounced reaction to a climate model with a large temperature increase and a small change in precipitation than to a model with similar temperature increase and a considerable increase in precipitation.

It was noted that the differences between SWAT and WaterGAP are smaller for the whole catchment (Zambski) than for its two sub-catchments (Burzyn and Suraż, occupying 24 and 12% of the whole catchment area, respectively). This can be explained by the fact that various model inputs have higher uncertainty for smaller areas, whilst for larger areas these differences are likely to cancel out (Qi and Grunwald 2005). Piniewski and Okruszko (2011), who performed spatial calibration and validation of SWAT in the Narew basin, noted also that the goodness-of-fit measures were connected to catchment area, i.e. the smaller the catchment the lower the NSE value.

5 CONCLUSIONS AND OUTLOOK

The results of our study show that the global model is able to capture some of the major responses to the climate change forcing. Given that the set-up, calibration and validation of a SWAT-type catchment model requires much time, human and financial resources, whilst the results of the global model are readily available (the SCENES webservice (http://www.cesr.de/SCENES_WebService/ [last accessed 11.04.2012]), we recommend using the latter for climate change impact assessments at a general level, for instance for indicators such as mean annual runoff, direction of change in monthly runoff, or shift in timing of peak runoff. We are not in position to extend this recommendation for the pan-European scale, but we believe that for the river basins situated in the same climatic zone (such as the Central and Eastern European lowlands) this statement should hold true. However, for more sophisticated assessments, taking into account, for example, the magnitudes of changes in mean and extreme monthly runoff, the local model has advantages over the global one. In practice, for instance in the Polish case, WaterGAP could be used for the country-wide general assessment and a SWAT-type model could be applied in selected hot spots, of special interest to water managers or decision makers.

As regards the reasons for the inconsistencies identified in the model results, we have found some evidence that if there is any part of WaterGAP that could be improved in the future, it is the modelling of vertical soil water balance and in particular the soil parameterization. We found that soil over-drying in summer and autumn is a likely reason for the underestimation of runoff in autumn and winter. Underestimation and too early occurrence of peak runoff in winter in WaterGAP could be explained either by improper parametrization of snow-related parameters or by the process of generating daily temperature time series from monthly values.

In order to gain more insight into the cross-scale issues related to climate change impact assessments, it would be beneficial to apply the approach undertaken in this paper to several more “case study” river basins situated in different parts of the European continent. This should be straightforward provided that the local models (not necessarily SWAT) are already set-up and calibrated for a baseline period similar to that used in WaterGAP. Given that there is considerable uncertainty across different global models in hydrological projections (Haddeland et al. 2011), such a study could also be a valuable complement to the work of Gosling et al. (2011) who found that it is equally feasible to apply the global hydrological model Mac-PDM.09 (Gosling and Arnell 2011) as it is to apply a catchment model to explore catchment-scale changes in runoff due to global warming from an ensemble of GCMs.

Further impacts of our findings on water management in the Narew basin should be analysed with respect to water use (domestic, industrial and agricultural) and environmental flows. In the first case, there is no evidence that relative changes even in the low flow period may alter the water-use possibilities, whether assuming the current use level or projected future water use (Giełczewski et al. 2011) in this region of low population density. In contrast, environmental flows should be of concern to the nature conservation authorities. The high ecological values of riparian wetlands located in the basins of the rivers Biebrza and Narew depend strongly on the availability of a flood pulse in spring (Okruszko et al. 2005). A shift of the inundation period may significantly change the habitat condition both for spawning of phytophilous fish species, such as pike and wels catfish (Piniewski et al. 2011), and for the waterfowl bird community. The buffering capacity of particular ecosystems and/or adaptation strategies should be considered in further study.

Acknowledgements

The authors gratefully acknowledge financial support for the project Water Scenarios for Europe and Neighbouring States (SCENES) from the European Commission (FP6 contract 036822). The authors appreciate constructive comments made by two anonymous referees that helped us clarify our presentation and, generally, improve the paper.