The two investigated catchments (Schwechat and Raab) are representative examples for river systems for the eastern region of Austria. Both rivers have their origin in the forested foothills of the limestone Alps, with a pre-alpine character and a low anthropogenic impact. The lower parts of both catchments are characterized by human activities, with primarily urban settlements and agricultural uses in the plains of the Schwechat catchment and dominant industrial activities and agricultural land uses in the valley bottom of the Raab catchment (Fig.  and Tables  and ).

The Schwechat River has its source in the Vienna woods at the northeastern boundary of the Northern Limestone Alps, with a maximum altitude of 893 m a.s.l. After a natural flow section in the narrow and dominantly forested valley of the “Helenental” (70 % of the total catchment area; see Table ), the Schwechat drains into the Vienna basin with flat topography and a predominance of agriculture, viniculture, and settlement areas. The main agricultural crops are winter wheat and summer wheat. Larger areas in the upper part of the catchment are used as pastures (∼10 % of the total area). The largest settlement is the city of Baden with a population of approximately 26 000 inhabitants, while smaller settlements are scattered over the catchment. All municipal wastewater is collected in three wastewater treatment plants (WWTPs; black triangles in Fig. ), where the WWTP Baden is the most relevant one, with a capacity of 45 000 PEs. All WWTPs perform carbon removal, nitrification, denitrification, and enhanced phosphorus removal. Due to the close proximity to the city of Vienna, population growth is a likely prospect for the settlement areas in the lower part of the catchment. The part of the catchment considered in this study has its outlet next to the city of Traiskirchen at an altitude of 185 m a.s.l. and covers an area of approximately 275 km2. The long-term mean annual precipitation in the Vienna basin is around 620 mm yr-1, and the mean annual temperature is 9.9 ∘C.

The Raab River originates at the edge of the southeastern Alps. These are characterized by low mountain ranges with a maximum altitude of 1547 m a.s.l., mostly covered by forests (∼42 % of the total catchment area; see Table ). The Raab flows through the southern part of Austria and crosses the border to Hungary close to the city of Neumarkt an der Raab at an altitude of 232 m a.s.l. The case study encompasses the Austrian part of the Raab with a catchment area of approximately 998 km2. The long, stretched river valley is dominated by agricultural activities (∼25 % of the total area), with urban areas in between. The slopes along the Raab are covered with heterogeneous patterns of forests, pasture areas, and agricultural land use. The main agricultural crops are corn and oil seed pumpkins, but wheat and vegetable production are also common. While the urban areas are of similar small structure to that in the Schwechat catchment, leather industries are present in the catchment that release substantial nutrient inputs into the receiving waters, which has resulted in trans-boundary conflicts . Municipal wastewater in the Raab catchment is collected in 12 relevant WWTPs (black triangles in Fig. ) that all have the same standards for wastewater treatment as in the Schwechat catchment but have almost 3 times the total capacity (approximately 150 000 PEs – population equivalents). Six relevant industrial emitters are located along the main reach of the Raab River (white triangles in Fig. ) that all perform internal wastewater treatment following the respective industry-specific regulations for wastewater treatment e.g.,. The average annual precipitation in the Raab catchment is approximately 800 mm yr-1, and the long-term annual mean temperature is 9.0 ∘C.

The SWAT model is a continuous, process-based semi-distributed eco-hydrological model. In this study we implemented SWAT2012 (revision 622) to simulate daily time series of discharge and NO3--N loads at the catchment outlets. The models' spatial reference to a catchment is given by a subdivision of the basin into subbasins. Areas containing the same land use and soil type and that are lying in the same slope range are lumped together in each subbasin to form hydrologic response units (HRUs). All processes on the land phase of each subbasin are calculated at the HRU scale and are further propagated into the water phase of each subbasin. The processes calculated on the land phase include water balance components such as interception; infiltration; shallow and deep percolation; surface runoff; lateral flow; groundwater flow; plant uptake and evapotranspiration; or the pathways of nutrients such as the input through atmospheric deposition or fertilizer application, the transformation into other forms of a nutrient, and the transport through surface runoff, percolation, lateral flow, and return flow in the groundwater;. In the water phase, the nutrients budgets are calculated. Following the calculation of the water balance and the nutrient budgets, the discharge, the nutrient loads, and other substances are routed through the linked subbasins to the defined catchment outlet . The required input data to set up a model with SWAT are a digital elevation model (DEM), a raster land use map including the model parametrization and the performed management operations for each land use, a raster soil map with soil physical and chemical parameters for all soil layers, and meteorological input data.

A DEM with a 10 m resolution was available for Austria from an airborne laser scan . Based on the DEM we defined three slope classes with slopes of 0 %–3 %, 3 %–8 %, and >8 % in the HRU definition step.

The CORINE Land Cover project served as the base land use map to which more detailed agricultural data were added. CORINE does not classify agricultural land uses into crop types. Therefore, tabular data of agricultural land uses at the municipal level derived from the 2010 Austrian Agricultural Census were superimposed onto CORINE data by randomly distributing crops according to the crops' areal share at the municipal level to CORINE pixels containing agricultural and complex cultivation land use. Typical time windows for planting, fertilizer application, tillage, and harvest were derived from field experiment records for the individual crops and written to the HRU management files. The management dates were randomized for all HRUs within the time windows derived for a management operation. Dates with strong rainfall or a high soil moisture potential were not used for scheduling management operations. With 70.0 % and 42.3 % forest land uses were the most dominant land uses in the Schwechat and the Raab catchments, respectively. The SWAT model setups differentiated between deciduous forests, coniferous forests, and mixed forests, derived from the CORINE Land Cover project (see Tables  and ). All HRUs with one of the three forest types as land use were parameterized with an initial biomass and an initial leave area index to simulate intact forests in both catchments.

The SoilGrids database is a consistent global soil information system that provides soil physical and chemical parameters at a 250 m grid resolution and seven soil depths. We utilized the available soil parameters from SoilGrids and estimated further required soil parameters with pedotransfer functions provided by the R package euptf . The seven available soil depths from the SoilGrids data were aggregated to three soil depths (0–30, 30–100, and 100–200 cm), and the gridded data were clustered into soil classes applying k-means clustering , resulting in 14 and 8 “optimum” soil classes for the rivers Schwechat and Raab, respectively.

Meteorological input data were available from the INCA system developed and operated by the Central Institute for Meteorology and Geodynamics of Austria ZAMG;. INCA provides reanalysis data of precipitation and temperature on 1 km grid resolution for Austria, with a temporal resolution of 15 min for precipitation and 60 min for temperature in the period from 2003 to 2015. For all SWAT model setups, daily precipitation sums and daily minimum and maximum temperatures were temporally and spatially aggregated for the model subbasins.

Point source emission data were available from external emission monitoring of municipal WWTPs greater than 2000 PE, according to for both catchments. Municipal WWTPs larger than 2000 PE are responsible for 99.2 % and 86.3 % of municipal point source emissions in the Schwechat and the Raab catchments, respectively. Thus, these data cover a substantial part of the municipal emissions. Additionally, daily and weekly internal monitoring data were available for some large WWTP schemes. In most cases, however, only information on NO3--N emissions was provided. A general budgeting of nitrogen emissions, however, showed that the substantial share of total nitrogen is emitted in the form of NO3--N (87 % in the Schwechat catchment and 89 % in the Raab catchment). For industrial emitters, monthly and annual records from internal and external monitoring agencies were available and only allowed an estimation of industrial emissions with coarse temporal resolution, while covering the annual budgets. Again, mainly data for NO3--N emissions were available. Although nitrogen is emitted in different forms, the available databases only allowed the consideration of NO3--N loads contributed by point sources.

Table  provides an overview of the model input data that were used for the SWAT model setup.

Hourly observations of discharge were available for the period from 2003 to 2015 at two gauges for the Schwechat and the Raab each (Fig. ). NO3--N concentration readings with varying time intervals of 5 to 15 min were available at two stations in both catchments (yellow circles in Fig. ) for selected time periods resulting from monitoring campaigns at the rivers Schwechat and Raab . SWAT simulates output variables with daily time steps. To compare the observations with the modeled SWAT outputs of discharge and NO3--N loads, daily NO3--N loads and daily mean discharge were calculated from the observation data.

Graphical GIS user interfaces such as ArcSWAT or QSWAT facilitate the setup of SWAT models. Yet a model setup requires the modeler to define the number of subbasins as well as the number of HRUs (e.g., by removing HRUs with areas below a certain threshold from the setup and apportion their areas to the remaining HRUs). The size and the number of subbasins in a model setup can affect the process simulations and the resulting model outputs . Removing small HRUs from the model setup and allocating their areas to the remaining HRUs affects the distribution of land use, soil types, and slope classes and thus can impact the model simulations substantially .

We used the ArcSWAT plugin (Version2012.10_1.14) together with ArcGIS 10.1 for the model setup. For both case studies we set up the SWAT model with different numbers of subbasins, whereby we prepared model setups with the full number of HRUs and respective setups with a reduced number of HRUs for each catchment.

In total, we set up four SWAT models, two with 3 and two with 14 subbasins for the Schwechat catchment and six models for the Raab catchments with two each of 4, 29, and 54 subbasins. For the full HRU setups we kept the resulting HRUs unmodified. For the model setups with a reduced number of HRUs we eliminated small HRUs. We determined thresholds for land use, soil, and slope classes to remove HRUs that have an area below these found thresholds. The thresholds were determined using the R package “topHRU” . The topHRU package enables finding thresholds that minimize the number of HRUs of a SWAT model setup while minimizing the aggregation error (sum of changes in the areas of land uses, soils, and slope classes of the reduced set of HRUs compared to the full HRU setup). To maintain a comparability between the reduced HRU setups thresholds were selected that result in an aggregation error of maximum 5 % in all reduced HRU model setups. Table  gives an overview of the final model setups for both case studies.

In a parameter screening, we applied a GSA to the simulations of discharge and NO3--N loads at the catchment outlets of all SWAT model setups to identify influential model parameters. Initially, 42 model parameters were selected that are frequently calibrated in SWAT model setups to simulate discharge and NO3--N loads (see e.g., , and , for a general overview of relevant model parameters; and for parameters controlling the water balance and nutrient cycles; or for a review on the dominant nitrogen parameters). The SWAT model setup initializes the model parameters using values obtained from the SWAT databases (either standard values or user-defined values, e.g., by pedotransfer functions). The selected initial ranges to modify the model parameters and the selected types of parameter changes (e.g., replace parameter values globally or modify a spatially distributed parameter field by a fraction of a parameter) reflect typical procedures often found in SWAT model calibration studies. An overview of the model parameters that were identified as influential and that were further used in the model impact study is provided in Table .

We employed the STAR VARS approach to screen and rank the model parameters. STAR VARS utilizes variograms along each model input's dimension of the input space to infer each model inputs influence on a target variable over different scales (where short lag distances approximate the derivative based method of Morris – , and long distances approximate based on the method of Sobol – ). The calculation of the variograms is based on the tailored STAR sampling design, where “star center” points are randomly sampled in the input space. For each center point, cross sections are sampled along the input factor dimensions with an equally spaced interval. For each sampled input combination the model is evaluated, and variograms along the response surface are calculated. proposed integrated measures of the variograms as measures of sensitivity, where the measures IVARS10, IVARS30, and IVARS50 represent the integrals over 10 %, 30 %, and 50 % of each input dimension, respectively, and therefore provide the sensitivity of a target variable to a model input over different scales. A detailed description of the method is provided in , and the STAR sampling is outlined in . The method proved to be robust and computationally efficient for high-dimensional problems e.g.,.

We drew STAR samples with 50 center points and 10 parameter samples per parameter dimension that resulted in 18 950 parameter combinations per model setup. The Nash–Sutcliffe efficiency criterion NSE;; the Kling–Gupta efficiency criterion (KGE), including its three components ; and a refined version of the index of agreement were used to evaluate the simulated time series of daily mean discharge and daily sums of NO3--N loads. Additionally, we applied the ratio of the root-mean-square error and standard deviation (RSR; ) to evaluate different segments of the FDCs of daily discharge and daily NO3--N load simulations . All calculated criteria were included in the parameter sensitivity analysis as target variables. A model parameter was considered to be sensitive if it showed a relative sensitivity of 10 % compared with the most sensitive parameter with respect to a specific objective criterion for at least one of the employed objective criteria.

The performed GSA for the model parameters of the different model setups of the Schwechat catchment and the Raab catchment, respectively, showed very similar results independent of the number of subbasins and HRUs of the individual model setups (Fig. ). Therefore, for the impact study the same set of model parameters was considered as influential for all model setups of the Schwechat and the Raab, respectively. In total, 19 parameters for the Schwechat and 16 parameters for the Raab were identified as being influential for the analyzed target variables (Table ). The majority of parameters were identified as influential parameters in the Schwechat and the Raab case study. The parameters SNO50COV, CANMX, CDN, and SDNCO were only relevant for the model setups in the Schwechat, and the parameter OV_N was only influential for in the Raab. For the majority of these parameters it is a matter of the selected threshold that defines a parameter to be influential or not. The most dominant parameters were, however, identified as highly relevant in both case studies.

To represent the model parametrization as input in the subsequent sensitivity and uncertainty analysis of the environmental impact study, non-unique parameter sets were identified for the Schwechat and the Raab catchments, respectively. The preceding parameter SA revealed that changes in the model parameter values influenced the simulations similarly independent of the subbasin and HRU configurations in the Schwechat and the Raab catchment, respectively. As a consequence, but also to facilitate the separation of the effects of the model setup and the model parametrization in the analysis, we selected parameter combinations as non-unique ones that result in simulations of daily discharge and NO3--N loads that fulfill certain objective criteria together with all model setups of the Schwechat and the Raab, respectively. For the respective 19 and 16 influential model parameters we randomly sampled 100 000 parameter combinations and simulated daily discharge and NO3--N loads with all model setups of the Schwechat and the Raab catchments. We evaluated the simulations with the following criteria for accepting a parameter set: KGE > 0.5 for daily discharge at the catchment outlets, KGE > 0.4 for daily NO3--N loads at the gauges with longer continuous records (in both case studies the gauging point within the catchment and not at the catchment outlet), percentage bias <50 % for NO3--N loads, and the absolute RSR < 1 for different discharge and NO3--N loads according to. In total, we identified 43 and 52 behavioral parameter combinations for the Schwechat and the Raab catchments, respectively. The ability of the selected parameter sets used with the different model setups to reproduce the observed data is illustrated in Fig. . The initial and final ranges of parameter changes are shown in Table . The 43 and 52 parameter combinations are additionally illustrated in parallel coordinate plots for the Schwechat and the Raab in Fig.  to show any clustering of individual parameters and interactions between parameters. The majority of parameters are scattered randomly and do not show any clustering or interaction with other parameters. The parameters RCN and NPERCO in the Schwechat catchment show a clear inverse relationship. This implies that the parameters compensate each other in the behavioral model setups. This finding seems plausible for the Schwechat catchment, where the NO3--N transport into the receiving waters is strongly groundwater driven and a surplus of NO3--N input is reduced by a decrease in NO3--N percolation. The parameters SLSOIL, SURLAG, and SOL_AWC show a clear bimodal pattern for the Raab catchment. The bimodal patterns of these parameters are strongly related, and a compensation effect between these parameters is visible. Model setups with increased slope values (SLSOIL) and longer lag times of the surface runoff (SURLAG) together with an increased soil available water content (SOL_AWC) resulted in a behavioral model and were able to reproduce historic discharge and NO3--N records similar to the model setups where such clear relationship is not visible.

The study involves future changes of the land use, point source emissions, and the climate. The uncertainties of these variables are expressed as discrete scenarios.

For the land use change scenarios, two scenario storylines were developed for the Schwechat and the Raab catchments. A business-as-usual scenario extrapolates the trends that we determined for the dominant crops in the time period 1970–2010 to the future (2071 to 2100), while a second extensive scenario assumes a more extensive application of agricultural practices and a stronger focus on extensive land uses in both catchments (Table ).

In the Schwechat catchment population growth is the strongest factor for a future change in land use . Hence, a transformation from extensive pastureland (-35 %) to urban land use and an increase in dense urban areas describe the business-as-usual scenario. The extensive scenario assumes no change in population and a shift of half of the wheat-producing area to extensive pastures.

Since 1970, the areas for corn production increased by 220 % in the Raab catchment, mostly for biogas production and at the expense of sugar beets and cereals . For the business-as-usual scenario, an increase in the corn area by a further 100 % until the end of the century was assumed, replacing extensive pastures (-75 %), sugar beets (-80 %), legumes (-70 %), and winter wheat (-30 %).

Groundwater protection measures lead to strict regulations for fertilizer application in the Leibnitzerfeld region adjacent to the Raab catchment . Therefore, the extensive scenario assumes an adoption of similar nitrogen regulations in the Raab catchment. Thus, decreasing areas with intensive fertilizer application, such as corn, by 50 % and transforming these areas into extensive pastureland was carried out in this scenario.

Two municipal point source emission scenarios for both case studies (Table ) and two industrial point source emission scenarios for the Raab catchment (Table ) were developed. The future change in municipal emissions was assumed to be directly related to the change in population. For all provinces in the Schwechat basin future scenarios predict an average population growth of 32 % . The predictions of the population development in the provinces of the Raab are contradicting, with predicted changes between +2.3 % and -20.4 % .

In the Raab catchment 94 % of the industrial point source emissions stem from the leather industry, and almost 70 % of the industrial point source emissions are caused by one leather manufacturing company. Thus, industrial emission scenarios were developed for that particular manufacturer. As boundaries for the production, we defined an upper environmental boundary and a lower economical boundary for the prediction of future industrial emissions. Based on an assessment of effluent dilution , current environmental regulations allow an increase of 30 % in emissions from that leather producer, resulting in a total increase in industrial emissions of 22.6 %. Assuming a relocation of the two manufacturing sites of that leather producer to outside of the catchment would stop their emissions into the Raab, reducing the total industrial point emissions by 75.2 %.

Future climate change was considered with 22 downscaled and bias-corrected climate change scenarios (Table ). Regional climate simulations were obtained from the EU-CORDEX project , providing 11 GCM–RCM simulations for the emission scenarios RCP4.5 and RCP8.5 . In this study we utilized daily precipitation sums and daily minimum and maximum temperatures for the time period 2071 to 2100. The EURO-CORDEX climate simulations are available at a spatial resolution of 12.5 km EUR-11. Statistical downscaling was applied to prepare all climate simulations at a resolution of 1 km. To correct downscaling errors e.g.,, bias correction was applied to the climate simulations employing quantile mapping . Downscaling and bias correction were performed for the historical period 1971 to 2000, involving the reanalysis datasets SPARTACUS for minimum, mean, and maximum temperature and GPARD for daily precipitation sums.

Table  summarizes the land use change, point source emissions, and climate change and the model setups and model parametrizations that were used for the analysis of simulated discharge and NO3--N loads in the Schwechat and the Raab catchments. In total, 7000 combinations of land use, point source emissions, climate, model setups and model parametrizations were drawn for both case studies, applying a quasi-random sampling . The number of combinations results from previous experiments that apply the SA method of Sobol (results not shown) using the sampling strategy proposed by . A base sample size of Nb=1000 was used to meet the suggestions shown in . Thus, the total sample size of 7000 is defined as N=Nb(k+2), where k is the number of model inputs (k=5). Although report publications that required substantially larger base sample sizes (e.g., Nb=12000 in , or Nb=8192 in ) for convergence of the ranking of influential continuous model parameters, a sample size of 7000 includes 46 % and 12 % of all possible model input combinations in the Schwechat and the Raab case studies, respectively. All sampled combinations were assembled to executable SWAT models. Daily discharge and daily NO3--N loads at the outlets of the Schwechat and the Raab catchments were simulated for the period from 2071 to 2100.

The analysis of discharge and NO3--N loads follows two main goals: (i) to identify the dominant controls on the simulation of discharge and NO3--N loads in the two case studies and (ii) to assess how the considered inputs control the simulation of discharge and NO3--N loads.

Figure  summarizes the influence of the implemented land use, point source emission, climate scenarios, the different model setups, and the model parametrizations for the simulation of future discharge and NO3--N loads in the Schwechat (panel a) and the Raab (panel b) catchments. Each plot panel shows the calculated PAWN indices for the analyzed target variables for one model input in a catchment. Related target variables are grouped by colors to support the interpretability (e.g., to identify changes in sensitivity from high to low discharge). In its entity each panel provides a general overview of the importance of input for the simulation of discharge and NO3--N loads. Individual PAWN indices (a single bar in a plot panel) highlight the importance of input for the simulation of specific characteristics of the time series of discharge and NO3--N loads.

The white boxes on top of each bar show the bootstrap means and the 95 % CIs of each PAWN index and therefore provide an indicator for the adequacy of the sample size that was used to perform the analysis and the impact of differing n+umbers of discrete values of the analyzed input variables. In general the bootstrapping resulted in narrow confidence intervals (maximum +0.05 and -0.08) for all analyzed model inputs and all signature measures, providing high confidence in the resulting sensitivities. Although the numbers of discrete realizations of the analyzed model inputs (e.g., only two land use scenarios, but 43 and 52 model parametrizations) differ strongly and therefore result in different subset sizes to calculate the PAWN indices, no substantial differences in the confidence intervals is visible.

The land use scenarios applied to SWAT demonstrated a rather negligible impact on all signature measures, with mean PAWN indices below 0.05 and 0.07 and confidence intervals in the same range for the Schwechat and Raab, respectively (first row Fig. ). The point source scenarios, in contrast, showed a considerable influence on the signature measures of NO3--N loads and concentrations in the Raab case study, while the impacts of the point sources in the Schwechat case study were negligibly low (second row Fig. ). Thus, based on the implemented point source emission scenarios, industrial emitters in the Raab catchment are relevant for the development of in-stream NO3--N loads and concentrations, particularly for low discharges and low NO3--N loads. The importance of the industrial point sources in SWAT increases when higher NO3--N load quantiles (low NO3--N loads, from dark yellow to light yellow in Fig. ) and NO3--N concentrations for low discharges (from dark red to light red in Fig. ) are simulated, which is evident from an increase in the mean PAWN index from 0.11 to 0.49 and 0.22 to 0.43, respectively. The climate scenarios and the model parametrizations show respective decreases in their importance for the simulation of low NO3--N loads and NO3--N concentrations for low discharges (with decreases in the mean PAWN index from 0.71 to 0.28 for the climate scenarios' influence on NO3--N loads and from 0.79 to 0.36 for model parametrization's influence on NO3--N concentrations).

The implemented climate scenarios showed large impacts on all calculated signature measures of discharge and NO3--N loads (third row Fig. ). The mean PAWN indices range between 0.25 to 0.90 and 0.25 to 0.96 for the Schwechat and the Raab, respectively. The climate scenarios were the most relevant inputs for the simulation of seasonal mean discharges and seasonal sums of NO3--N loads. For the simulation of low discharge quantiles (large daily discharges) climate scenarios showed the highest relevance. For the simulation of low discharges however, the importance of the climate scenarios decreases, while the model parametrization becomes more relevant (from dark green to light green in Fig. ). The mean PAWN indices of climate scenarios drop from 0.74 to 0.47 in the Schwechat catchment and from 0.82 to 0.51 for the simulation of lower discharges, while the mean PAWN indices for the model parametrization show respective increases from 0.43 to 0.87 and 0.44 to 0.80.

In general, the model parametrization was highly influential for all calculated signature measures and is comparable to that of the climate scenarios, with mean PAWN indices ranging between 0.43 to 0.90 in the Schwechat and 0.36 to 0.80 in the Raab (fifth row Fig. ). Particularly, for the simulation of NO3--N concentrations the model parametrization was the most dominant control of the variable simulated. In contrast to the large impact of the model parametrization, the relevance of the model setup was much lower for the simulation of discharge and NO3--N loads and concentrations. Overall, values of the PAWN index for the choice of the model setup did not exceed 0.37 and were much smaller (2 to 5 times) compared to the model parametrization. The model setups yielded insignificantly low PAWN indices for the majority of signature measures with values below 0.1 in the Raab case study (2.5 % CI almost 0 for many signature measures), indicating that the model setup had a low influence on most of the analyzed processes. Only for high discharges and large NO3--N loads is a mean value for the PAWN index above 0.1 visible.

Using all 7000 combinations of land use, point source emissions, climate, model setups, and model parametrizations, the simulated discharges and NO3--N loads deviated by up to 350 % (grey bands in Fig. ) from the simulations of discharge and NO3--N loads in the reference period 2003 to 2015 (dashed line in Fig. ). In the Schwechat (left column in Fig. ) wider uncertainty bands are visible for the spring and early summer months. The results for the Raab catchment (right column) show that wider uncertainty bands emerged for summer as well as for winter and early spring. A notable difference between the two case studies is how the simulations of long-term monthly discharges and NO3--N loads in the reference period compare with the ranges of future simulations. While the majority of model combinations for the Schwechat simulated larger discharges and NO3--N loads for all months in the future, for the Raab catchment the simulations of discharge and especially NO3--N loads are lower in comparison to the reference period.

The analyses of the uncertainty bands with respect to the implemented land use scenarios and the point source scenarios fully confirm the results from the SA (Fig. ). The attributed uncertainty bands for the two land use scenarios almost entirely overlap and show only minor deviations. A similar result is illustrated for the two point source scenarios in the Schwechat case study. The scenarios in the Raab catchment involved industrial point source emissions. The grouped uncertainty bands that include scenarios with an increase in industrial production (red) and the uncertainty bands that include a decrease in industrial production (blue) show similar patterns. Yet the blue and red uncertainty bands show a clear shift to each other. On average the scenarios with an increase in industrial production show long-term monthly sums of NO3--N loads that are 15 t higher compared with the scenarios with a decrease in industrial production. The same scenarios show larger amplitudes for medium and low NO3--N loads, while large NO3--N loads remain uninfluenced by the two scenarios for the development of the leather industry.

With the GSA we identified that the climate scenarios have a great influence on all signature measures of the simulated variables. Attributing the uncertainty bands to the individual GCM–RCM combinations unveils diverse outcomes for the future flow regime, the distribution and amplitude of monthly NO3--N loads, and the appearance of high and low discharges and NO3--N loads (Fig. ). A visual analysis of the separated uncertainty bands identifies that the deviations of the mean annual precipitation of the GCM–RCM combinations have a strong impact on the simulation of discharge and NO3--N loads. In comparison to the reference period (dashed line), wetter future climate scenarios (blue) simulated larger discharge and NO3--N loads, while drier future conditions lead to a drastic reduction in discharge and NO3--N loads. These findings further imply that NO3--N applied in fertilizers will remain in the upper soil layers and be transformed (mineralized or immobilized or denitrified) instead of being transported to the receiving waters. A comparison of the NO3--N budgets of simulations with dry and wet climate scenarios for the Raab shows a difference of up to +27 % of NO3--N accumulated in the soil, as well as a decrease of 43 % and 38 % in NO3--N yield in the fast and slow runoff, respectively.

Half of the 22 implemented GCM–RCM combinations simulated an increase of more than 75 mm (dark blue), and for two GCM–RCM combinations, an increase of more than 25 mm (light blue) in precipitation for the Schwechat catchment was simulated. In contrast, for the Raab, nine and four GCM–RCM combinations simulated a decrease in precipitation of more than 75 mm (dark red) and 25 mm (light red), respectively. Consequently, a decrease in discharge and NO3--N loads due to a decrease in precipitation is pronounced in the Raab catchment, while the majority of simulations of the Schwechat catchment show an increase in discharge and NO3--N loads.

While a grouping of the individual climate scenarios with respect to their temperature deviations shows a more indefinite picture, all climate scenarios simulated an increase in temperature. Nevertheless, the expectation that an increase in annual mean temperature increases evapotranspiration and thus reduces discharge and NO3--N loads is not met in Fig. . A clear separation of warmer and cooler climate scenarios as observable for precipitation is not the case with temperature. Consequently, the differences in precipitation predominantly account for the influence of the climate scenarios rather than the differences in temperature.

Although the influence of the model setups was much lower compared to the influence of the climate scenarios or the model parametrization, the analysis of the uncertainty bands for the different model setups provides interesting insights (Fig. ). The uncertainty bands do overlap to a great extent, which confirms a low impact of the use of different model setups in the simulation of discharge and NO3--N loads. It is noteworthy that model setups that use the full set of HRUs agree more strongly in their simulations compared to the model setups where the number of HRUs was reduced. The difference between the full HRU and the reduced HRU model setups is distinct in the Schwechat case study. The uncertainty bands of the two full HRU model setups almost completely overlap, although their numbers of subbasins are different (4 and 14 subbasins). The two model setups with a reduced number of HRUs (but also with 4 and 14 subbasins) show differences of up to 15 mm in the simulated monthly specific discharge and up to 7 t in the monthly NO3--N loads (∼20 % of the uncertainty bandwidth).

The model parametrizations were relevant for all signature measures of discharge, and NO3--N loads and were most dominant for medium and low flows. The most dominant model parameters in both case studies were the parameters CNOP_till and SOL_AWC. Both parameters control the water retention and thus the immanent contribution of rainfall to the river discharge. Large values of CNOP_till and small values of SOL_AWC reduce the water retention capacity and increase the amplitude of medium and low discharges (third row in Fig. ). A similar but inverse behavior is visible with medium NO3--N loads (last row in Fig. ), where a higher water retention results in an increase in NO3--N loads. For the long-term monthly mean discharges and sums of NO3--N loads two effects are observable in Fig. . First, smaller values of CNOP_till and larger values of SOL_AWC decrease the upper boundary of the uncertainty bands. Second, selected model parametrizations with large values of CNOP_till and small values of SOL_AWC cause considerably larger discharges in spring and a strongly reduced runoff in the autumn months in the Schwechat case study.

The illustrated case studies emphasized the necessity for characterizing, identifying, and explicitly communicating the uncertainties in a modeling chain, particularly for future simulations of environmental variables where large uncertainties are inherent in several modeling inputs. While the sensitivity analysis of signature measures related to discharge, NO3--N loads and NO3--N concentrations provided a comprehensive overview of the dominant influencing inputs on specific modeled variables, and the analysis of the uncertainty bands for the simulation of the modeled variables provided insights into which properties of the model inputs (e.g., mean annual precipitation or mean air temperature of a climate scenario) control the uncertainties and how these control the simulation. The analyses allow for drawing conclusions that are beneficial to consecutive steps of an impact study, for instance in refining the impact study setup and focusing on the most influential components and ultimately reducing the uncertainties in the modeling simulation chain.

The land use scenarios showed an almost negligible impact on the simulation of discharge and NO3--N loads. The discharge and the NO3--N loads at the catchment are, however, integrated signals for the entire catchment, and changes in land use may have a greater importance for particular points in a catchment. Many case studies have applied the SWAT model to assess the impact of land use change on different variables of the water cycle , water quality , or sediment yield . found very low increases induced by land use change in discharge for a catchment in China. Only an assumed strong intensification of the agriculture led to a 4 % increase in discharge. At the same time however, a strong increase in sediment yield of up to 450 % for the summer months was simulated due to the intensification of agriculture. also found only small changes in simulated discharge caused by future land use change in a German lowland catchment. In absolute numbers the simulated future NO3--N loads showed small differences between the baseline scenario and the two applied methods of land use change presented by . Yet the temporal patterns in NO3--N loads caused by the different approaches of changing the land use were the major observable difference. however found that including agricultural land use change into the impact assessment of a southern German watershed strongly increased the NO3--N and total phosphorus loads. support the findings of and also found that corn-intensive scenarios lead to an increase in discharge and significant water quality problems, while an extensive scenario where mainly switchgrass is planted leads to water quality improvements under future climate change. Consequently, the low impact of land use change found in the present study seems reasonable with respect to other literature, particularly as no extreme scenarios were implemented. This does however not generally imply a low importance of land use change in environmental impact assessments. Land use change or changes in the management can be the most relevant input, particularly when strong future changes such as possible bans of the emission of substances are considered .

Industrial emitters were the main cause for the impact of point sources on medium to low NO3--N loads. The future scenarios of the development of industrial emitters were however highly uncertain. The developed scenarios are based on expert knowledge. Yet there is no reliable basis available on status of the industrial emitters by the end of the century. Therefore, the developed scenarios should be noted as feasible futures, rather than, for example, politically realizable futures . Setting a feasible range as boundaries for the future development of industrial emitters can lead to an overestimation of their impact in comparison with other influencing variables. Nevertheless, the visualization of the NO3--N FDC of the Raab case study highlights the effect of the industrial emissions for medium and small NO3--N loads. Large NO3--N loads, however, are hardly affected by the implemented scenarios, indicating that large NO3--N emissions are mainly driven by agricultural activities.

The selection of climate scenarios had a strong influence on the simulation of discharge and NO3--N loads in both case studies. The analysis of the uncertainty bands identified the differences in precipitation between the GCM–RCM combinations as being the main control, while the differences in air temperature had a low impact on the simulation outcome. This finding stands in contrast to other studies. and for example, identified empirical approaches for the calculation of evapotranspiration as the main source for overestimation of the climate's influence on hydrological processes, particularly when evapotranspiration is a function of air temperature . In the climate scenarios used in this study, the impact of large differences in mean annual precipitation on the simulated outputs exceeded the impact of the differences in air temperature.

The effect of the model setup, with different watershed subdivisions, on the simulation of discharge or water quality variables has been investigated in various studies e.g.,. emphasize the greater impact of changes during the HRU definition over the defined number of subbasins, as a consequent change in the distribution of land use, soil, or topography strongly affects runoff and the nutrient budget in a catchment. The analysis of the uncertainty bands with respect to the different model setups clearly confirmed the study by , especially in the case of the Schwechat. Nevertheless, the impact of the model setup was lower than the effect of the model parametrization by a factor of up to 5 in the Schwechat study and up to 8 in the Raab case study. Yet the model setup strongly affects the computation time. In the present case, where aggregated discharge and NO3--N loads at the catchment outlets were the variables of interest, a strong focus on the model parametrization is of higher priority than the spatial distribution of the model setup. Therefore, to maintain short computation times (and at the same time to maintain the distributions of land use, soil, or topography) a model setup with a low number of subbasins without any reduction in the number of HRUs is beneficial.

The impact of parameter non-uniqueness on the simulation of hydrological and water quality variables has been demonstrated previously e.g.,. The importance of the model parametrization for the simulation of discharge and NO3--N loads was confirmed in the present study as well. Large sensitivities of all signature measures of discharge and NO3--N loads to the different model parametrizations were identified. Although all selected parameter sets represented historical observations of discharge and NO3--N loads with a certain goodness of fit (based on defined objective criteria), the colored grouping of the uncertainty bands illustrated that the selected model parameter sets control the simulation of future discharge and NO3--N loads in different ways. Thus, the large impact of the model parametrization and the distinctive patterns identified in the uncertainty bands suggest a great potential to further refine the model parametrization and consequently reduce simulation uncertainties with a more intensive model calibration. Additional information on the time series of observations can help to constrain the model parameters and adequately describe the relevant processes e.g.,.

Scenarios always reflect subjective assumptions made by the modeler. Assumptions that are made in the scenario development, however, can strongly influence a simulation and thus affect a comparison of different model inputs and their impacts on the simulation. All steps in a scenario development involve subjective assumptions and can lack plausibility ; regardless of whether the process involves expert knowledge, the input of stakeholders in an participatory process, or an exploratory approach that extrapolates trends, these practices potentially introduce uncertainties in the definition of scenarios. Technical aspects such as how the scenario is represented in the model are also strongly biased by the modeler's decision and represent an additional source of uncertainty . The communication of the potential uncertainties inherent in the developed scenarios and the boundaries of the explanatory power of an scenario ensemble is essential for the integrity of any impact study .

In the present study, several assumptions were made in the development of scenarios that are highly subjective, such as the extrapolated gradient of future land use changes, the drastic changes in future industrial emissions, and also the selection of objective criteria that define a behavioral SWAT model setup. Scenarios must cover a broad range of possible futures and have to be adequately represented in the model setup. An explicit delineation of the implemented scenarios and their limitations is essential in clearly illustrating the limitations of an impact study's conclusions. An immanent risk in any impact study is that the model representation of a future change or the uncertainties in a model input fail to reproduce the response of a simulated variable that would have taken place in the real environmental system. Hence, a detailed analysis of the simulation uncertainties perfectly compliments an SA in identifying possible shortcomings in the study setup. Attributing the uncertainty bands resulting from the simulation of an environmental variable to individual model inputs proves to be a useful visual analysis tool that gives the power to illustrate the uncertainties in a transparent way. Furthermore, the colored differentiation provides a visual guidance to judge the impacts of different implemented scenarios.

The presented approach implements large samples combining scenarios for different model inputs and different model setups and parametrizations in a GSA to identify the dominant contributors of uncertainties in the simulated outputs. The utilization of SA with large sample sizes, however, raises the following issues. (i) Compared to a standard approach for performing an impact assessment, where a few different future scenarios are implemented into a model, the computational demand of a GSA requiring hundreds or thousands of model executions is larger by several orders of magnitude. Thus, a practical implementation of the presented procedure in impact studies is questionable, and a strong cooperation between research and the practitioners is essential. (ii) Scenarios of different model inputs are often interrelated . A change in one model input therefore expects the change of another model input into one direction and makes a change into another direction unlikely. While the implementation of input dependencies, although challenging, is feasible for continuous model inputs; for instance by a transformation of the input space e.g., or the determination of input distribution functions , the dependencies of composite model inputs are usually difficult to express mathematically. To identify the dependencies between composite model inputs, expert knowledge is required to properly constrain the model input combinations, and this therefore complicates the implementation in approaches such as the presented one.

therefore suggest identifying consistent hydrologic storylines that result in least severe, most likely, and most severe responses of the modeled system. Such an approach would tremendously reduce the number of necessary model evaluations but also establish consistency between the considered influencing variables. Nevertheless, the feasible combinations of influencing variables that lead to extreme or likely responses of the modeled system are hardly known a priori. Consequently, a sensitivity analysis with a constrained sampling space to avoid infeasible combinations of influencing variables might be a pragmatic compromise.