Improving catchment discharge predictions by inferring flow route contributions from a nested-scale monitoring and model setup

Identifying effective measures to reduce nutrient loads of headwaters in lowland catchments requires a thorough understanding of flow routes of water and nutrients. In this paper we assess the value of nested-scale discharge and groundwater level measurements for the estimation of flow route volumes and for predictions of catchment discharge. In order to relate field-site measurements to the catchmentscale an upscaling approach is introduced that assumes that scale differences in flow route fluxes originate from differences in the relationship between groundwater storage and the spatial structure of the groundwater table. This relationship is characterized by the Groundwater Depth Distribution (GDD) curve that relates spatial variation in groundwater depths to the average groundwater depth. The GDDcurve was measured for a single field site (0.009 km 2) and simple process descriptions were applied to relate groundwater levels to flow route discharges. This parsimonious model could accurately describe observed storage, tube drain discharge, overland flow and groundwater flow simultaneously with Nash-Sutcliff coefficients exceeding 0.8. A probabilistic Monte Carlo approach was applied to upscale field-site measurements to catchment scales by inferring scale-specific GDD-curves from the hydrographs of two nested catchments (0.4 and 6.5 km2). The estimated contribution of tube drain effluent (a dominant source for nitrates) decreased with increasing scale from 76–79% at the field-site to 34–61% and Correspondence to: Y. van der Velde (ype.vandervelde@wur.nl) 25–50% for both catchment scales. These results were validated by demonstrating that a model conditioned on nestedscale measurements improves simulations of nitrate loads and predictions of extreme discharges during validation periods compared to a model that was conditioned on catchment discharge only.


Introduction
Intensive agriculture in lowland catchments often leads to high nutrient losses and eutrophication of downstream waters (Oenema et al., 2007;Van der Molen, 1998;Vitousek et al., 2009).To identify effective measures to reduce these nutrient loads, the flow routes of water that enter a stream and their nutrient concentrations need to be quantified (Tiemeyer et al., 2010).In densely drained lowland catchments, surface water discharge is fed by groundwater flow toward streams and ditches, tube drain effluent, and overland flow.Many field-scale studies identified tube drain effluent as the major source of nitrate (Tiemeyer et al., 2006;Nangia et al., 2010;Rozemeijer et al., 2010c).However, the field scale at which these contributions can be directly measured (De Vos et al., 2000;Van der Velde et al., 2010a) often is not the scale of interest to water management authorities.Extrapolation of fields site results to entire catchments can easily lead to wrong conclusions as field sites can prove nonrepresentative of the patterns and processes that emerge at larger scales (Sivapalan, 2003;Soulsby et al., 2006;Didszun and Uhlenbrook, 2008).Therefore, our challenge is Published by Copernicus Publications on behalf of the European Geosciences Union.
Y. van der Velde et al.: Nested scale monitoring and modeling to effectively integrate information from field-scale measurements into the prediction of catchment-scale flow route contributions.
In Van der Velde et al. (2010a), we presented the results of a field-scale measurement setup that separated tile drain flow from overland flow and groundwater flow.We also measured discharges at two larger nested scales and showed that, rather than the actual measured volumes at the field site, the characteristic response of individual flow routes can be used to upscale the field-site flow routes to the catchment scale.This elementary upscaling approach was purely based on measured data.A model framework was needed to upscale the measured field-scale fluxes to the catchment-scale for periods without complete sets of measurements.
To develop such a model upscaling approach, Sivapalan ( 2003) advocated the search for concepts that "easily connect scales, and that can also be easily scaled".This should lead to "a watershed-scale representation that is clearly tied to process descriptions at a lower level of scale, and which is not overly complex".In sloped terrain, scaling research has focused on the way in which hillslopes connect to headwaters (Uchida et al., 2005, Jensco et al., 2009;Tetzlaff et al., 2008;Clark et al. (2009) and headwaters to entire basins (Shaman et al., 2004).Rodgers et al. (2005), Tetzlaff et al. (2007), and Didszun and Uhlenbrook (2008) studied the scaling behavior of both discharge and tracers across nested-scale catchments and found that scaling effects in discharge and solutes could largely be attributed to scalerelated morphologic, topographic and land-use features.In contrast to sloped catchments, lowland catchments generally have little morphological heterogeneity and the main flow routes occur at all scales.Therefore, the scale effects in discharge of lowland catchments are primarily driven by scaledifferences in drainage density of ditches and tube drains, micro-topography (Appels et al., 2011), and soil type.
Tracer based hydrograph separation is an often applied technique to identify scale effects in contributions of different flow routes to catchment discharge (Didszun and Uhlenbrook, 2008;Tiemeyer et al., 2008;Van der Velde et al., 2010).Tiemeyer et al. (2008) and Rozemeijer et al. (2010c) demonstrated that for tile drained lowland systems, especially nitrate concentrations differ strongly between the major flow routes.Although they also showed that flow route nitrate concentrations may vary over time and uncertainties are substantial, the clear concentration differences make rough estimates of the contribution of individual flow routes to catchment discharge possible.Inversely, catchment nitrate concentration dynamics can be an important verification for simulated flow route contributions in a hydrological model of a lowland catchment ("getting the right answers for the right reasons", Kirchner, 2006).
Upscaling approaches for solute transport often focus on travel time distributions (e.g.Botter et al., 2009 andVan der Velde et al., 2010b).Travel time distributions can be defined at all scales and implicitly account for flow routes by combining the spectra of travel times contributing to each flow route in an overall travel time distribution.However, both Botter et al. (2009) and Van der Velde et al. (2010b) showed that a flow simulation with accurate contributions of flow routes is paramount for representing the dynamics in travel time distributions that is needed to describe solute transport dynamics.In Van der Velde et al. (2010b) we used a spatially distributed groundwater model to upscale fieldscale flow route contributions to catchment scales.However, because of very long calculation times and the large number of parameters, groundwater models are inflexible and often inaccurate in simulating fluxes of specific flow routes at catchment scales.Hence, Van der Velde et al. (2009) proposed a new upscaling approach for hydrology in lowland catchments (from here on called the Lowland Groundwater-Surface water Interaction model, LGSI-model).We assumed that each flow route (i.e., ditch and stream drainage, overland flow, and tube drain flow) starts to discharge if the groundwater level exceeds a flow route-specific threshold groundwater level at that location, and that the magnitude of the flux depends on the groundwater level.The contribution of a flow route to the total catchment discharge is calculated by integration over all groundwater levels in the catchment, described by a groundwater depth distribution.Van der Velde et al. (2009) showed that each storage volume of groundwater in the saturated zone corresponds to a unique groundwater depth distribution.They also showed that the relation between storage and groundwater depth distribution can be defined at any spatial scale and thus satisfies Sivapalan's (2003) criterion: it "easily connect scales, and can also be easily scaled".
However, in order to measure this relationship between storage and the groundwater depth distribution at catchment scales relevant to water management authorities, many groundwater depth time series are needed throughout the catchment.This makes this approach laborious and in our previous paper (Van der Velde et al., 2009) we had to resort to spatially distributed transient groundwater modeling to derive this relationship.A workable alternative would be to have a dense network of groundwater monitoring wells on a small area within the catchment, and observe the groundwater levels frequently for a limited time period.Obviously some sort of upscaling is then needed to use this data to characterize the behavior of the entire catchment.We introduce here a nested-scale model setup combined with a probabilistic Monte Carlo approach to achieve this.This approach is tested by evaluating the model uncertainty in discharge predictions during an 8 year period and by evaluating uncertainty in simulated nitrate concentrations dynamics during a month with three successive storm events.
The objectives of this paper are twofold.Firstly, we test whether the LGSI-model can accurately describe all individual flow route fluxes at the field scale.This would increase our confidence in the ability of the LGSI-model to simulate flow route fluxes accurately at the catchment scale where Hydrol.Earth Syst.Sci., 15, 913-930, 2011 www.hydrol-earth-syst-sci.net/15/913/2011/

Materials and methods
This paper combines the nested-scale measurements introduced by Van der Velde et al. (2010a) and the upscaling approach described in Van der Velde et al. (2009).Therefore, we offer a brief summary of the relevant information (Sects.2.1 and 2.2), and refer to both papers for detailed background information.The LGSI-model is first applied to the field-site discharge and groundwater level measurements, which is described in Sect.2.3.Section 2.4 introduces a catchment model conditioned on catchment discharge and a groundwater level time series and a nested-scales model that combines the field-site model and the catchment model.Section 2.5 introduces validation strategies for both the catchment model and the nested-scales model to assess the value of nested scale monitoring.

Nested experimental setup
The measurements for this study were performed in the Hupsel Brook catchment in the eastern part of The Netherlands (Fig. 1a) (52 • 06 N; 6 • 65 E).The size of the catchment is 6.5 km 2 , with surface elevations ranging from 22 to 36 m above sea level.At depths ranging from 0.5 to 20 m an impermeable marine clay layer is found (Van Ommen et al., 1989).
The unconfined aquifer consists of Pleistocene aeolian sands with occasional layers of clay, peat, and gravel (see Wösten et al. (1985) for more details).
The Hupsel Brook catchment is drained by a straightened and deepened main brook and by a dense artificial drainage network of ditches and tube drains.The spacing between the ditches averages 300 m (Fig. 1a) and approximately 50% of the area is tube drained (plastic perforated flexible tubes).
The Hupsel Brook catchment has a semi-humid sea climate with a yearly precipitation of 500 to 1100 mm and a yearly estimated evaporation of 300 to 600 mm.
Within the Hupsel Brook catchment, discharge was measured at three nested spatial scales: (1) the entire catchment of approximately 6.5 km 2 , (2) a sub-catchment of 0.4 km 2 and (3) a 0.009 km 2 field site located within the subcatchment (Fig. 1b).From August 2007 through December 2008, discharge was measured every 15 min for both catchment scales.Continuous surface water nitrate concentrations were measured at the outlet of the entire catchment with a Hydrion-10 multi-parameter probe (Hydrion BV Wageningen, the Netherlands).Monthly average nitrate concentrations of tube drain effluent were measured at 20 locations in the catchment with Sorbi-Samplers (Rozemeijer et al., 2010a).
The tube drained field site of 0.9 ha has a drain spacing of 14.5 m.Along a 43.5 m stretch inside the deep easterly ditch (Fig. 1c), we built in-stream reservoirs with separate vessels to capture tube drain discharge.The in-stream reservoirs collected overland flow and groundwater influx through the stream bed.Thus we separated the tube drain flow from the combined flux of overland flow and groundwater flow.The discharge of both flow routes was measured with 5 min intervals for November 2007 through December 2008.During that period we also manually measured phreatic groundwater levels at 31 locations within the field-site every week.Pressure sensors in 15 piezometers along drain 1 (Fig. 1c) recorded phreatic levels every 10 min.A meteorological station of the KNMI (Royal Dutch Meteorological Institute) bordering the field-site measured hourly rainfall and evapotransipration derived with the Makkink relation (Makkink, 1957).
Overland (q ov ), q grw,ov = (m−1)•u r ex for u < 0 r ex : Exfiltration resistance [T] Groundwater q grw,ov = 0 for u > 0 flow (q grw ) Tube drain flow q dr = D dr −u r dr for u < D dr D dr : Tube drain depth [L] (q dr ) q dr = 0 for u > D dr r dr : Tube drain resistance

Model concepts
The LGSI-model (Van der Velde et al., 2009) essentially consists of point-scale expressions of flow route fluxes (tube drain flow, overland flow, groundwater flow, direct rainfall, and evapotranspiration) and storages (saturated storage, unsaturated storage, and surface storage).A point in the catchment starts to generate a flux for a certain flow route when its groundwater level exceeds a threshold specific to that flow route.The magnitude of this flux (except for evapotranspiration) is directly proportional to the difference be-tween the groundwater level and the threshold level.Upscaling of fluxes and storages is achieved by integrating the point-scale expressions over all groundwater depths within a model area.This distribution of groundwater depths was found to approximate a normal distribution.Van der Velde at al. (2009) showed that because of the local nature of groundwater flows towards the nearest ditch or depression in lowland catchments, the overall pattern is a repetition of similarly shaped water tables within fields.The shape within a field depends on the soil type, distance between ditches and drains, surface elevation and the stored water volume.Hence, the catchment-scale distribution of groundwater depths is a summation of the distributions within all fields which approaches a normal distribution via the central limit theorem.

Field site
Sub-catchment Catchment A dr /A tot 1.0 0.7-0.9 (4)0.4-0.6 (4) r ex 0.1-10 d (3) A s /A tot 0 0.0054-0.0066 (4)0.009-0.011 (4)r ditch 500-4000 d (3) A tot 7700-9000 m 2 0.36-0.48Km 2 (4) 6.0-7.3Km 2 (4) u ditch 1.05 m (5) (1) Soil parameter estimates from Wösten et al. (2001) D dr 0.75-0.95m (5) (2) DEM r dr 100-300 d (3) (3) Rough estimates (4) Topographic maps and field survey b dr 0-0.14 d −1 (3) (5) Field-site measurements This reasoning is valid as long as the scale of the features that create the shape of the water table is far smaller than the model area.The catchment-scale relation between storage and the normal distribution was formalized by the Groundwater Depth Distribution curve (GDD-curve) that describes the relationship between the spatial average groundwater depth and the spatial groundwater depth standard deviation.In summary: saturated storage is calculated by a water balance.Each storage volume corresponds to a distribution of groundwater depths via the GDD-curve.From this distribution all fluxes and unsaturated and surface storage are calculated through simple threshold expressions that relate a water table to a flux or storage.
The model parameters of LGSI-model can be subdivided into process-specific parameters that describe fluxes and storages as a function of the local groundwater level and scale-specific parameters that describe the spatial distribution of groundwater depths, the total catchment area, the tube-drained area, and the area occupied by the surface water network.In Table 1 all point-scale process formulations, the GDD-curve, and their parameters are introduced.Appendix A lists the abbreviations and variables used in this paper.The complete set of LGSI-model equations is summarized in Appendix B. The LGSI-model is a fast calculating process model that calculates flow route discharges for a decade on hourly basis within a few seconds.This is a huge advantage over fully distributed models (e.g.Rozemeijer et al., 2010b), and allows for extensive parameter estimation by Monte Carlo simulation as will be demonstrated in this pa-per.However, the model is less suited to evaluate the effects of measures that affect the shape of the groundwater table, since this shape is derived from measurements and not calculated from physical principles.

Model extensions for the field site
To apply the LGSI-model to the field site, the basic setup needed to be extended to explicitly include groundwater flow out of the field into the deep ditch, lateral groundwater flow into the field from adjacent fields, and a time-variant flow resistance of the tube drains.
The effects of the single deep ditch to the east of the field on the water table within the field are large compared to the field-site area and hence are not well represented by the assumption of a normal distribution of groundwater depths.To account for groundwater flow towards this ditch, we introduce a new discharge term that approximates the groundwater flux to the deep ditch as a function of the average groundwater depth in the field, u(t) [L]: with u ditch [L] the depth of the ditch relative to the mean surface elevation of the field site and r ditch [T], the resistance of the field-site to groundwater flow towards the ditch.The lateral groundwater inflow, L f [LT −1 ], was assumed constant throughout the simulation period.The total discharge measured by the in-stream reservoirs of the field experiment, Q res [LT −1 ], can now be calculated < 1 cm (0.2-0.8 p) * (0.1-0.9 p) * * The lower and upper quantile of the modeled groundwater depth distribution that is assumed to envelope the measured groundwater depth at the field site.The GE gives the maximum average difference between the measured groundwater depth and the modeled envelope.by all the water that enters the surface water except for the tube drain flux: with Q ov the flux by overland flow, P Q rainfall on ponded surfaces (including the ditch) and E Q evaporation from ponded surfaces (all LT −1 ).
During the experimental period, we observed a strong decline in the drainage effectiveness of the tube drains.At the beginning of the experiment the tube drains were cleaned by pressure flushing as is common practice in the Hupsel Brook catchment.This pressure flushing is repeated every two years.We hypothesized that the tube drains slowly get clogged in periods with substantial discharge and that in dry periods without discharge, aeration and oxidation of the clogging material inside the tube drains reduces the resistance.Similar behavior was also observed by Bentley and Skaggs (1993).The following simple empirical relation was adopted to account for the tube drain resistance change as a function of discharge:

Probabilistic parameter estimation
A parsimonious process model as the LGSI-model necessarily suffers from equifinality (parameter non-uniqueness; Beven and Freer, 2001) stemming from parameter uncertainty, the lumped nature of the parameters, the subjectivity introduced by including and excluding processes, the chosen process formulations, and the many different types of measurements that the model needs to describe.We dealt with equifinality by generating many combinations of parameters in a Monte Carlo procedure (GLUE, Beven and Freer, 2001).
In this study we introduce three LGSI-models: a field-site model, a catchment model, and a nested-scales model.For each model, random parameter values were generated from prior uniform and independent distributions between predetermined parameter ranges.All parameter ranges are listed in Table 2. Parameter sets were qualified behavioral when the model satisfactory described the measured data and all behavioral parameter sets were considered equally probable.The criteria that divide the parameter space in behavioral and non-behavioral parameter sets for each of the models are listed in Table 3.This procedure was continued for each model until an ensemble of 500 behavioral parameter sets was found.

Field-site model of flow route fluxes
The measured groundwater levels at the field-site were converted to field-site average groundwater depths, standard deviations of groundwater depths, and volumes of ponds on the soils surface to comply with the variables of the LGSI-model.The measured absolute groundwater levels at 31 locations within the field (Fig. 1c) were interpolated to a groundwater table for the entire field.Subsequently, this groundwater table was subtracted from a detailed DEM (5 × 5 m resolution) and all groundwater depths were grouped into a groundwater depth distribution.The volume of negative groundwater depths of this distribution quantifies the volume of ponds on the field.A mean and standard deviation of the groundwater depth distribution and the volume of ponds were calculated for all 57 weekly field-site groundwater depth surveys.Continuous groundwater level measurements in the 15 groundwater wells around the tube drain 1 (Fig. 1c) were used to interpolate between the weekly field average groundwater depths in order to create a continuous field average groundwater depth time series.Figure 2 shows the five sequential steps that were followed to generate an ensemble of 500 Behavioral Parameter Sets (BPS-FS).The relations between model equations and measured data determined the specific order of these steps.The parameter distributions of BPS-FS were analyzed for parameter sensitivity and the model results were analyzed to quantify the uncertainty in flow route contributions to the total discharge owing to equifinality.

Catchment and nested scales model
The catchment model was conditioned on catchment discharge and a single groundwater level time series.The resulting ensemble of 500 Behavioral Parameter Sets is referred to by BPS-C.Table 2 lists the parameter ranges from which the parameter sets were generated and Table 3 list the criteria that devides the parameter space in behavioral and non-behavioral parameter sets.The ranges for the processspecific parameters were equal to those used for the fieldsite model.Only the ranges for the scale-specific parameters that described areas (i.e. the catchment area, the tube drained area, and the area of the surface water) were different.
The groundwater level time series, on which the catchment model was conditioned, was not considered representative for the dynamics of the groundwater storage of the entire catchment.From field experience, we estimated that at any one time, at least 10% of the catchment area had shallower groundwater and another 10% had deeper groundwater than the single observed level: the observed groundwater level thus was assumed to be within the 0.10 and 0.90 percentile but allowing for an average exceedence of 1.0 cm (GE, Table 3).
We attempted to constrain the uncertainty in flow route contributions of the catchment model by combining information from measurements from the field site, discharge measurements of a small sub-catchment, and discharge measurements at the catchment outlet in a nested-scales model setup.This nested-scales model consist of thee LGSI-models representing each of the scales: field-site, sub-catchment, and catchment.These models are connected by assuming that the parameters that describe the discharge response to groundwater depth are scale invariant (process-specific parameters, see Table 1), while the parameters that describe the spatial distribution of groundwater depths are assumed scale-specific (Table 1).The underlying hypothesis is that the differences between the observed hydrographs at the three scales are primarily an effect of a different spatial distribution of groundwater depths and resulting different active drainage areas.model setup (BPS-N).For the catchment-scales new constant drainage depth and drainage resistance were estimated, because the drainage depth and the time variant drainage resistance of the field site were specifically estimated for the three drains of the field site.Three drains probably do not represent the drain populations at larger scales.Hence, for the sub-catchment and catchment scale, we assumed a constant spatially averaged drainage resistance and drainage depth.The flow route contributions to discharge during the entire field-site monitoring period and the parameter distributions for BPS-C and BPS-N were compared to assess the added value of introducing nested-scale measurements.

Model validation
The BPS-C and the BPS-N were both validated for their ability to predict the catchment-scale discharge.For the validation we chose the period 1994-1995 for its high quality discharge data without data gaps and obvious measurement errors, and 1996-2001 for its episodes of extremely high discharges that are outside the discharge range of the calibration period.
The impact of uncertainty in flow route contributions to discharge on solute transport is demonstrated for the period of 16 March to 10 April 2008.For this period measured nitrate loads are compared with loads calculated by both BPS-C and BPS-N.For each flow route constant concentrations were estimated from measurements reported in Van der Velde et al. (2010a): tube drain flow (72 mg L −1 ), overland flow (9 mg L −1 ) and direct rainfall (3 mg L −1 ).The groundwater flow concentration was estimated to be 40 mg L −1 , based on the results of Rozemeijer et al. (2010c).Although constant nitrate concentration do not reflect the complexity of all processes affecting nitrate, Rozemeijer et al. (2010c) showed that tube drain nitrate concentrations remained relatively constant during this period and are not much affected by rainfall events.Also they showed that the spatial variability of nitrate concentrations is far larger than the temporal variation of individual flow routes at a single location.Therefore, constant flow route concentrations are a reasonable approximation for the one month period proposed for this comparison Note, that this approximation is only applied to illustrate the role of flow route fluxes in solute transport simulations and is not presented as a solute transport model.

Field-site model results
Van der Velde et al. ( 2009) reported a decrease in the simulated groundwater depth variance as the groundwater depth increased in their catchment-scale groundwater model.For the field site this finding is corroborated by observations for a wet and a dry day (Fig. 3).Approximating the groundwater depth distributions by normal distributions introduced only small errors (Fig. 3).It shows that the extent of the features that dominate the shape of the water table such as tube drains, soil heterogeneity and micro topography are far smaller than the size of the field site.Consequently the central limit the-orem guides the overall distribution of groundwater depths towards a normal distribution even for this single field site.Figure 4a shows the measured groundwater depth means and standard deviations for the 57 weekly groundwater depth surveys.The grey band in Fig. 4a represents the results of all behavioral parameter sets (BPS-FS) for the GDD-curve.This grey band is particularly narrow between average groundwater depths of 0.5 and 0.8 m (some of the data points are even outside it), indicating that the model results for storage and discharge are very sensitive to the GDD-curve in this range of average groundwater depths.GDD-curves outside the grey band, although they may closely fit the observed GDD-data, did not yield behavioral models for some of the other criteria such as storage or discharge.In Fig. 4b we plotted the measured ponding volumes and the ponding curves of all BPS-FS.Because the measured ponding volumes are relatively uncertain (they are difficult to measure and we have only a few measurements) we allowed for a larger curve error (CE, Table 3).This resulted in the grey band in Fig. 4b.
Figure 5a shows the measured and modeled spatially averaged groundwater depth.The model results are accurate, but some of the moderate groundwater level peaks are underestimated.This also caused an underestimation of the tube drain discharge (Fig. 5b) during these moderate groundwater level peaks.Overall, Fig. 5 shows that the LGSI-model is able to accurately describe the average groundwater depth, tube drain flow, and reservoir discharge (Eq.2) simultaneously.All three time series were simulated with a Nash-Sutcliff (NS) coefficient (Nash and Sutcliff, 1970) exceeding 0.8.
A comparison of the LGSI-model results with the results of Rozemeijer et al. (2010b), who used the same dataset and a fully distributed HydroGeosphere (Therrien et al., 2009) model to simulate flow routes during a single discharge

Nested-scales model results
The discharges at all three nested-scales could be accurately described by parameter sets that share the same values for the process-specific parameters and differ only in scale-specific parameter values (Fig. 6).This result supports our hypothesis that scale effects in lowland hydrology can be attributed to scale differences in the shape of the groundwater table.These scale differences were quantified by the GDD-curves of the individual scales.Figure 7 shows the inferred ensemble of GDD-curves for the three nested scales.The differences between the field-site GDD-curve and the GDD-curve of the two catchment scales are much larger than those between the GDD-curves of both catchment scales.This is consistent with the small differences in the shape of the hydrographs between both catchment scales.
The individual flow route contributions (the median contribution of BPS-N) to discharge are shown in Fig. 8.The relative overland flow and groundwater flow contributions to discharge increase with increasing scale, at the expense of tube drain discharge.Table 4 gives the 10-90 percentile estimates of flow route contributions for the entire simulation period (this period equals the field-site measurement period).The uncertainty in the flow route contributions at the field site is constrained by many different types of measurements (see also Fig. 5).At the larger scales fewer measurements were available and consequently the uncertainty is much larger.In Table 4 we also compared the uncertainty of the flow route contributions calculated by BPS-N and BPS-C (the model constrained by catchment discharge only).The uncertainty in groundwater flow, overland flow, and direct rainfall is significantly reduced by introducing nested scale measurements.In contrast, the uncertainty of the tube drain discharge could hardly be reduced because the field-site tube drain depth and tube drain resistance could not be transferred to larger scales.We re-estimated the tube drain-specific parameters for the sub-catchment and catchment-scale by assuming them to be equal for both scales, but even this assumption did not reduce the uncertainty.

Parameter distributions
Figure 9 shows the distribution of the behavioral parameter sets of all process-specific parameters.Since the prior distributions were uniform, sensitive parameters are identified by markedly non-uniform distributions.For the field site model (BPS-FS) the most sensitive parameters are α, D dr , a dr , L f , and r ditch .The sensitivity of a dr signals the importance of accounting for increasing tube drain resistance after tube drain cleaning.The lateral inflow of groundwater, L f , which is the closing term for the water balance, could be determined accurately around 0.6 mm day −1 .Surprisingly insensitive parameters are θ s and u ET .The insensitivity of u ET signals that evapotranspiration reduction at our relatively wet (high groundwater tables) field site might not be very important.
As long as the value of u ET is less than two standard deviations (20 cm) below the lowest average groundwater water table, the modeled evapotranspiration reduction is small and u ET does not affect the model results.
At the catchment scale (BPS-C) on the other hand, u ET is the most sensitive parameter.At this scale this parameters closes the overall water balance by increasing or reducing evapotranspiration (at the catchment scale there is not net lateral groundwater flow).Also m and r ex are relatively sensitive as they directly control the discharge, which is the only calibration objective.All other parameters are insensitive and the uncertainty in flow route contribution is largely determined by the boundaries of the prior distributions.
For the nested-scales model (BPS-N) we sought processspecific parameter sets that can describe all three scales simultaneously.Figure 9 shows that the parameter distributions of BPS-N combine the constraints of the distribution of both BPS-FS and BPS-C.The added value of including nested-scale measurements to reduce parameter uncertainty is apparent for almost all process-specific parameters.
Figure 10 shows the distributions of the scale-specific parameters and again the reduction in parameter uncertainty by introducing nested-scale measurements is clear.Note that the distribution of σ min tends to high values exceeding the   preset boundaries in Table 2 that were determined from a detailed DEM.Under dry conditions, the groundwater level is almost parallel to the soil surface.Variations in the groundwater depth under such conditions emanate largely from local variations in the soil surface elevation that are too small to appear in the phreatic level.However, under dry conditions a few very deep incisions of the stream produce most discharge.Around these incisions, the groundwater depth necessarily decreases sharply to zero at the stream bank, and these deviating groundwater depths produce outliers from the normal distribution valid for the rest of the catchment.As a consequence, the calibration tried to increase the groundwater depth variation (and thereby the range of σ min ) under dry conditions to be able to generate low discharges.Because these discharges are low, the effect of underestimating discharges during dry conditions on the entire water balance of the catchment is small.In general, both Figs. 9 and 10 suggest that that the LGSImodel is highly over-parameterized when the model is only calibrated on discharge, because almost all parameters of BPS-C are insensitive (uniform distributions).However, nested-scale measurements and the assumption that scale differences originate only from differences in the shape of the groundwater table, increases the parameter sensitivity and allows most parameters to be conditioned.

Model validation
The BPS-N parameter sets performed slightly better than BPS-C for both validation periods.For the period 1994-1995 the BPS-N yielded an average NS-coefficient of 0.90, against 0.85 for the BPS-C.For the validation period 1996-2001 the average NS-coefficients were 0.79 for BPS-N and 0.73 for BPS-C.Although these may seem minor improvements in model performance when weighted against the efforts involved in the nested-scale monitoring, these model improvements are especially apparent in extreme discharge situations beyond the range of the calibration dataset.Figure 11 shows the validation results of the four most extreme discharge events during the period 1996-2001.From all events it is clear that the BPS-N much better predicted discharge than BPS-C (particularly the magnitude of the peaks), with far smaller uncertainty ranges.During the discharge event in Fig. 11c, which are the highest discharges measured in the past 20 years, the measured maximum discharge was more than twice the maximum discharge of the calibration period.During this period some of the BPS-C overestimated the discharge by a factor 3, while the BPS-N all predicted discharges close to the measured discharge.
The model improvement achieved by conditioning the model on nested-scale measurements is also apparent from the comparison between simulated and measured nitrate concentrations and nitrate loads (Fig. 12).There are hardly any differences between simulated discharges of BPS-N and BPS-C (Fig. 12a), as both parameter sets have been cal-ibrated on this discharge time series.However, both parameter sets simulate different combinations of flow routes volumes that yield this overall discharge.The effects of flow route contributions on solute transport become apparent when each flow routes receives a unique concentration as demonstrated in Figs.12b and c.Especially during peak discharges the spread of concentrations is much smaller for the parameter sets calibrated on the nested scale measurements (BPS-N).Hence, from the observed dynamics in nitrate concentrations it can be concluded that the flow route volumes of BPS-N are a more realistic representation of the flow process during high flow conditions.Figure 12c shows that conditioning on nested-scale data reduced the uncertainty in cumulative load estimates by roughly 50% and thus demonstrates the crucial role of accurate understanding of flow route discharges for predicting solute loads towards downstream surface water bodies.

Conclusions
Detailed and unique flux measurements at a pasture field site allowed us to formulate and calibrate our parsimonious LGSI-model.Even the very non-linear process of saturated overland flow was adequately simulated.An exceptional feature of this model is that the model concepts were designed around the available measurements (Van der Velde et al., 2009).Consequently, the parameters that describe the discharge and storage processes could all be conditioned on measurements.This yielded a field-site model that accurately described both storage and fluxes simultaneously.
The combined nested-scale measurement and model setup made it possible to combine discharge information of the field scale, a small sub-catchment, and the entire catchment.We demonstrated that the differences between hydrographs at the three scales could all be described by only changing the Groundwater Depth Distribution (GDD) curve, even though the hydrographs were markedly different.This result supports our hypothesis that scale effects on discharge in lowland catchments are primarily an effect of differences in the spatial distributions of groundwater depths between different scales.Still, the range of GDD-curves that yielded good model results for the catchment scales was wide.This emphasizes the importance of spatially distributed groundwater depth monitoring to further condition these GDDcurves.This should lead to an even more solid foundation of the relation between storage and discharge, and hence to more reliable results.
Tube drain effluent is the most important route for nitrate towards the surface water network in lowland catchments.We were able to measure tube drain discharge at the field site and we concluded that almost 80% of the water (Van der Velde et al., 2010a) and 92% of the nitrate (Rozemeijer et al., 2010c) was transported by tube drains.It is by no means trivial to extrapolate these field-scale findings to the entire catchment.Our combined nested-scales observation and modeling approach could narrow down our estimates of the contribution of tube drain discharge to the discharge of a sub-catchment of 0.4 km 2 to 34-61% of the total discharge.For the entire catchment of 6.5 km 2 , we estimated that 25-50% of the discharge originated from tube drains.These results not only demonstrate that we need to be careful extrapolating field experiment results to entire catchments but also show that nested-scale measurements are essential to understand and quantify the flow route contributions to the discharge of a catchment.
In this paper we demonstrated the potential of combined nested-scale monitoring and modeling for the Hupsel Brook catchment.However, many of our findings can be generalized.First of all, we showed that detailed field-site measurements of storage and flow routes provide the processunderstanding that is needed to develop a model structure that adequately describes the catchment-specific flow routes.Secondly, we demonstrated that the combination of a relatively short period of nested-scale measurements with nested-scale models significantly constrains uncertainty in the contributions of groundwater flow, overland flow, and direct rainfall into surface waters.This reduction in flow route flux uncertainty significantly reduces uncertainty in nitrate load estimates as demonstrated by a simple flow route mixing approach.Finally, we showed that conditioning parameter sets on nested-scale measurements considerably improves discharge predictions compared to parameter sets constrained on discharge only.Model calibration on nested-scale measurements may not yield models with better calibration-statistics than models that are calibrated on catchment discharge alone, but the nested-scales model approach yields models that are able to predict (peak) discharges during validation periods more accurately.Improved quantifications and predictions of solute loads and peak discharges make the efforts involved in nested-scale monitoring worthwhile.

GDD-curve
Groundwater Depth Distribution curve.
Curve that relates the spatial standard deviation of the groundwater depth to its spatial average.

Ponding-curve
The relation between spatially averaged groundwater depth and the volume of ponds and surface waters.BPS Ensemble of 500 Behavioral Parameter Sets BPS-FS, BPS-C, BPS-N : BPS for each of the three models: fieldsite, catchment and nested-scales model, respectively.Fractional rate with which r dr decreases during dry periods c dr [L 3 T −1 ] Threshold tube drain discharge: below this discharge r dr decreases, above this discharge r dr increases CDE [-] Cumulative discharge error, difference between cumulative measured and modeled discharge CE [-] Curve error, difference between measured data and the modeled GDDcurve or ponding-curve D dr [L] Tube drain depth e pot [LT −1 ] Potential evapotranspiration ET act [LT −1 ]  Actual evapotranspiration

A2 Symbols
Evaporation from ponded surface and surface waters fu [L −1 ] Normal distribution function of groundwater depths Fu [-] Cumulative distribution function of groundwater depths Fu −1 [-] Inverse cumulative distribution function of groundwater depths
dr >c dr } a dr − 1 {Q dr <c dr } b dr r dr (3) with a dr [-] the rate with which the drainage resistance, r dr [T], increases when the tube drain discharge is larger than threshold discharge c dr [L 3 T −1 ].The resistance decreases with fractional rate b dr [T −1 ] for discharges smaller than c dr .

Fig. 2 .----
Fig.2.Procedure for derivation of behavioral parameter sets for the fields-site and the nested-scales model.The first five steps yield behavioral parameter set for the field site (BPS-FS).All seven steps yield behavioral parameter sets for the nested scale model setup (BPS-N).The parameters are explained in Table1.

Fig. 3 .
Fig. 3. Map of field-site groundwater depth for a dry (A) and a wet (C) day and the corresponding groundwater depth distributions (B and D).

Fig. 4 .
Fig. 4. Measured and simulated groundwater depth distribution curve (GDD-curve) (A) and Ponding curve (B) of the field-site.The grey area represents the ensemble of behavioral parameter sets of the field-site model (BPS-FS).

Fig. 5 .
Fig. 5. Measurements and simulation results of the field site model (BPS-FS) for the spatially averaged groundwater depth (A), tube drain discharge (B), and reservoir discharge (combined flux of overland flow and groundwater flow, C) of the field-site.The black band gives the results of all behavioral parameter sets (BPS-FS).

Fig. 6 .
Fig. 6.Measurements and simulation results the nested-scales model for total discharges of the field-site (A), sub-catchment (B) and entire catchment (C).The black band gives the results of all behavioral parameter sets (BPS-N).

Fig. 7 .
Fig. 7. Bandwidth of behavioral GDD-curves for the three scales of the nested-scales model (BPS-N).The dark grey area indicates overlap between ensembles of GDD-curves.

Fig. 8 .
Fig. 8. Contributions of flow routes to total discharge.The displayed contribution is the median of contribution of all behavioral parameter sets (BPS-N).

Table 4 .
Calculated flow route contribution (0.1-0.9 quantiles) of BPS-N and BPS-C.The contribution is calculated over the period November 2007 through December 2008.

Fig. 9 .
Fig. 9. Process-specific parameter distributions for the three behavioral parameter sets: BPS-FS, BPS-C, and BPS-N.The grey-filled distributions were only used for the field-site sub-model within the nested-scales model.

Fig. 11 .
Fig. 11.Model validation results for extreme discharge events at the catchment outlet.The light grey band gives the results of the model calibrated on all nested-scale measurements, while the black band give the results of the model calibrated on catchment discharge only.

Fig. 12 .
Fig. 12. Simulation results of BPS-N and BPS-C for discharge (A), nitrate concentration via a flow route mixing analysis (B), and cumulative nitrate load (C).
r dr increases during wet periods b dr [T −1 ]

Table 1 .
LGSI-Model basics and parameters.The process formulations are point-scale model equations.The catchment-scale equations are obtained by integration over the groundwater depth distribution within appropriate integration bounds (see Appendix B).

Table 2 .
Estimated Process-specific and Scale-specific parameter ranges.

Table 3 .
The cutoff criteria for behavioral model runs.Model runs are assigned behavioral when they meet to all "goodness of fit" criteria.Expressions for the error terms are given in Appendix C.