Lack of data is one of the main limitations for hydrological modeling. However, it is often used as a justification for over simplifying, poorly performing models. If we want to enhance our understanding of hydrological systems, it is important to fully exploit the information contained in the available data, and to learn from model deficiencies. In this paper, we propose a methodology where we systematically update the model structure, progressively incorporating new hypotheses of catchment behavior. We apply this methodology to the Alzette river basin in Luxembourg, showing how stepwise model improvement helps to identify the behavior of this catchment. We show that the most significant improvement of the evolving model structure is associated to the characterization of antecedent wetness. This is improved accounting for interception, which affects vertical storage distribution, and accounting for rainfall spatial heterogeneity, which influences storage variations in the horizontal dimension. Overall, our results suggested that, due to the damping effect of the basin, the description of fast catchment response benefits more from spatially distributed information than that of slow catchment response.
 The main challenge of hydrological modeling is to express the response of a catchment in terms of its state variables and characteristics. This calls for a solution of the “closure” problem, which requires specifying the conditions that allow the closure of the undetermined system of balance equations [Reggiani et al., 2000]. A major obstacle to this challenge is the lack of appropriate measuring techniques, which hampers the identification of the mechanisms underlying the rainfall-runoff transformation [Beven, 2006]. To circumvent this problem, internal catchment behavior has to be inferred by other means.
 One approach is to build a model of the catchment and to identify it with the natural system. This approach considers the model as a “virtual laboratory” [Weiler and McDonnell, 2004] which is used to perform experiments in a controlled environment. This approach has been widely used in hydrology to illuminate aspects of catchment behavior that are difficult to measure and that are poorly understood, such as the sensitivity of catchment response to soil properties [e.g., Herbst et al., 2006], land cover change [e.g., Eckhardt et al., 2003], and rainfall input [e.g., Schuurmans and Bierkens, 2007; Arnaud et al., 2002]. While the advantage of this approach is that everything about the model is known and controllable, what limits its reliability is the unconditioned identification of the model with the natural system, which implicitly disregards the role of conceptual model errors.
 An alternative approach is to infer internal catchment behavior using real data. Clearly, the use of real data should place more confidence in the outcome of a study. However, more often than not, the only measured variable available to test model assumptions is catchment discharge. Several authors indicated that in this case the essential features of catchment behavior can be reliably estimated with only a handful of parameters [Beven, 1989; Jakeman and Hornberger, 1993; Hsu et al., 1995; Young and Parkinson, 2002]. Under such a complexity constraint it is difficult to explore the internal catchment behavior beyond very simplistic descriptions, and support models that retain enough empirical realism to be useful in addressing complex environmental problems [Vachè and McDonnell, 2006]. Indeed, most studies comparing complex versus simpler models [Perrin et al., 2001; Carpenter et al., 2001; Refsgaard and Knudsen, 1996; Michaud and Sorooshian, 1994] including those related to the Distributed Model Inter-comparison Project [Reed et al., 2004] could not detect significant differences in performance between different types of models.
 Most of these studies, however, evaluated model performance with respect to discharge data alone based on a single statistical summary. This implies a loss of information which affects both the level of complexity that a model can support without incurring in identifiability problems [Wagener et al., 2003], and the assessment of relative merits of individual models in an inter-comparison study.
 In the absence of direct measurements of individual processes, we propose a methodology to better explore internal catchment behavior based on a combination between the “top-down” approach to model development, and a multiobjective approach to model evaluation. The top-down approach, initially introduced by Klemes  and recently reformulated by Sivapalan et al. , is based on a deductive philosophy which traces back the “causes” that are implicit in the overall “effect” produced by a system. In hydrological modeling, this approach would start from a simple structure which is progressively expanded based on its shortcomings in reproducing overall catchment behavior [e.g., Jothityangkoon et al., 2001]. We use this approach in the process of “fingering down into the (smaller-scale) processes from above (i.e., catchment scale)” [Sivapalan et al., 2003]. The multiobjective approach to model evaluation [Gupta et al., 1998] is based on the consideration that a single measure of performance does not fully exploit the information contained in the data, and establishes that model performance should be evaluated with respect to several indicators. We use this approach not only to evaluate single models, but also to compare the performance of competing models [Fenicia et al., 2007]. The combination of the two approaches provides a powerful tool to guide model improvement, and to evaluate different hypotheses on catchment behavior.
 We apply this methodology to the Alzette catchment in Luxembourg, where we test different model concepts first in a lumped and then in a distributed mode. In line with the top-down approach, we use successful model modifications to learn new aspects about the actual catchment behavior.
 One aspect of our results relates to the importance of the interception process. This process, while accounting for an important component of the water balance, is often neglected in model applications [Savenije, 2004]. With respect to discharge simulation, some authors found that the introduction of interception in a model structure improved model performance [e.g., Zhang and Savenije, 2005], others experienced deterioration of model performance [Lindström et al., 1997].
 A second important aspect relates to the question of whether spatial heterogeneity of rainfall has an impact on catchment discharge, and for which combination of processes. While virtual experiments [e.g., Arnaud et al., 2002] suggest that the catchment is sensitive to the spatial heterogeneity of rainfall, studies which employed real data [e.g., Obled et al., 1994] support an opposite conclusion.
 The results of this work, as well as our considerations on the potential benefit of distributed information in catchment modeling, will be further analyzed in the discussion section of the paper.
2. Study Site and Data Description
 The study site is the Alzette river basin upstream of the Ettelbrück streamgauge. The catchment is mostly located in the Grand Duchy of Luxembourg and covers an area of about 1090 km2 (Figure 1). The lithology is complex and heterogeneous. The northern part is dominated by an impervious schist formation, the central part is characterized by marls, highly responsive to rainfall, and the southern part by sandstone and limestone, which constitute the main infiltration zone and the main groundwater reservoir of the basin. Land cover is composed of forest (34%), agriculture (23%), pasture (31%) and urban areas (12%).
 The catchment is instrumented with several recording raingauges. For this study, we selected 8 raingauges based on the available length of record (Figure 1). Rainfall and discharge data have been calculated at an hourly time step, while the daily potential evaporation (including all forms of evaporation) has been determined by the Hamon equation [Hamon, 1961] with data measured at Findel (Luxembourg airport). The daily potential evaporation has been disaggregated into hourly values using a sine curve distribution between the hours of sunlight. In total, five years of data were available for model evaluation, from September 2000 to August 2005. The last three years have been used for calibration and the first two for validation, as hydrograph peaks are somewhat larger in the calibration period.
 Within the catchment area, several plot scale experiments have been carried out to investigate small-scale catchment processes and their eventual link to larger scale catchment behavior. Of interest to this study are canopy interception measurements executed on a 0.06 ha experimental plot located in a central location within the basin (Huewelerbach catchment), at 380 m altitude. The plot is equipped with more than 80 pluviometers spaced at about a three-meter distance. Measurements were available for one hydrological year (2003–2004) and showed that the annual canopy interception rate is about 13% of the total incident precipitation. Beech litter interception was measured with a special forest floor interception device at ground level filled with forest litter [Gerrits et al., 2007; De Groen and Savenije, 2006]. The box is equipped with weight sensors that allow continuous measurement of moisture storage. Preliminary results showed that litter interception (during the month of November) amounts to about 35% of the throughfall, indicating that the total annual evaporation from interception (by canopy and forest floor) is more than 43%.
 Interception data were not used directly for model evaluation, both because the forest floor interception data extend over a too short a period of time, and for the difficulty of up-scaling the data to the entire basin. However, they were used as an indication of the order of magnitude of the interception process under the given climate and vegetation conditions.
3. Model Description
 We started modeling the catchment with a basic structure, based on the Flex model described by Fenicia et al. . The model is of the lumped conceptual type and is composed of several interconnected boxes representing key zones of catchment response. The basic version, which we named FlexB, is characterized by three reservoirs: the unsaturated soil reservoir (UR), the fast reacting reservoir (FR) and the slow reacting reservoir (SR) (Figure 2). Rainfall input is first averaged with the Thiessen polygon method and then routed through the model. The model does not include a separate description of interception, hence rainfall (R) equals effective rainfall (Re) which reaches UR. Infiltration into the soil (Ru) is evaluated based on a rainfall excess model that uses a distribution function of soil moisture storage capacity over the catchment (equations (1) and (2)).
 The coefficient of runoff Cr is a function of the storage in UR, Su, and depends on the parameters Sfc, representing the maximum storage of UR, and a shape parameter β. When Su is filled to capacity, excess rainfall is routed to FR. The part of rainfall that does not infiltrate either reaches SR through preferential recharge (Rs) or is routed as runoff (Rf) to FR (equations (3) and (4)).
 Percolation is linearly related to the relative soil moisture content Su/Sfc with a maximum value of Pmax (equation (5)).
 As there is no interception, transpiration from vegetation and evaporation from interception are combined into a lumped evaporation term. This is often done in modeling, although conceptually it is considered an erroneous operation [Savenije, 2004]. In this approach, the actual total evaporation EUR depends linearly on the relative soil moisture content until this ratio exceeds the threshold Lp, after which it equals the potential evaporation Ep (equation (6)).
 The fluxes entering FR and SR are lagged through triangular transfer functions of linearly increasing weights (Figure 2) that are defined by the parameters Nlagf and Nlags, which determine the number of time steps in the transformation routine. The fast and slow discharges Qf and Qs are calculated through a linear relation between storage and discharge (equations (7) and (8)).
 In total the model has nine parameters: a shape parameter for runoff generation β (−), the maximum UR storage Sfc (mm), the runoff partitioning coefficient D (−), the maximum percolation rate Pmax (mm/h), the threshold for potential evaporation Lp (−), the lag times of the transfer functions Nlagf (h) and Nlags (h), and the timescales of FR and SR: Kf (h) and Ks (h).
4. Model Evaluation
 The issue of model evaluation has triggered a stimulating discussion in literature [Wagener et al., 2003]. From the large variety of existing methods for model evaluation, we adopted the multiobjective optimization approach proposed by Gupta et al. . This approach evaluates a model simultaneously by a number of objective functions. The objective functions are defined so as to assess the model performance with respect to different aspects of the system's behavior, with lower values indicating better performance. The trade-off between the different performance indicators is visualized by the Pareto-optimal front, which identifies the maximum performance that can be achieved by the model, given the available data. Improvements in performance can therefore be easily tracked by a shift of the Pareto optimal front toward the origin of the objective function space [Fenicia et al., 2007].
 The selection of performance measures can be based on different considerations, including knowledge of the errors involved in the modeling process, or, as in this case, arbitrary judgment related to the purpose of modeling [Gupta et al., 1998]. The performance measures selected for this study are the Nash and Sutcliffe coefficient CNS and the correlation coefficient R. These functions increase with model performance, having a maximum value of one. Because optimization generally deals with a minimization procedure, the functions are subtracted from one, as shown in equations (9) and (10), yielding the objective functions FNS and FCC.
 Where n is the number of observations, i accounts for the time steps, and Q is discharge. The subscripts s and o indicate the simulated and observed values, while the overbar indicates an average value over the period of observation. FNS has been chosen mostly for the purpose of communication, as it is a performance measure that is commonly used in modeling. This function is largely influenced by peak flow prediction errors due to the use of squared residuals. FCC has been selected to complement FNS with a relatively uncorrelated performance criterion. This function reaches a minimum value of zero when for all i: (Qs,i − s) = α(Qo,i − o), with α as a positive constant. It therefore indicates the agreement between the dynamics of the two hydrographs, disregarding the difference in their absolute values. In this case, it is useful to assess the time-shift between observed and simulated time series and serves the purpose of evaluating whether the timing is well captured by the model. The evaluation of this aspect is directly related to some of the modeling objectives (see section 5.2).
 As a sampling algorithm to estimate the Pareto-optimal front, the MOSCEM-UA (Multi Objective Shuffled Complex Evolution University of Arizona) algorithm [Vrugt et al., 2003a] has been used. The algorithm requires the selection of a number of parameters: the maximum number of iterations (set at 20000), the number of complexes (set at 10), and the number of random samples that is used to initialize each complex (1000). Parameter bounds have been chosen based on the results obtained with a preliminary run of the MOSCEM-UA on a very large parameter space.
 Calibration results for the FlexB model structure are summarized in Figure 3. The figure shows the Pareto-optimal front associated to the FlexB model, and of other models that will be introduced further down. The Pareto-optimal front shows the trade-off between the selected modeling objectives. Each point on the front may not be considered better than any other point on the same front, since moving from one point to another corresponds to an improvement in one objective function and a deterioration in the other. Moreover, it is not possible to find points outside the front that assume better (lower) values for both objective functions. Therefore the Pareto-optimal front marks the best performance that can be reached with a model with respect to the selected modeling objectives.
 The hydrograph generated by the FlexB Pareto-optimal model with best FNS value is represented in Figure 4. This model is characterized by a CNS of 0.87, which is a relatively high value. Despite that the performance of the FlexB model is relatively good, we investigated several modifications to this model structure to check whether they resulted in significant performance improvement. From this analysis, in line with the top-down approach, we tried to gain new insights into catchment behavior.
5. Proposed Modifications
5.1. Lumped Mode
 We tested a number of possible model improvements, first in a lumped, and subsequently in a distributed mode. In the lumped mode, we evaluated the effect of alternative model adjustments to account for threshold processes in runoff production. In the FlexB model, the rainfall that produces fast runoff (routed to FR) depends merely on the value of Cr. Hence we tested the effect of three alternative hypotheses: (1) Cr is zero if UR is below a certain threshold URt (mm). Should this happen, all rainfall infiltrates into UR. This hypothesis was implemented in the FlexURt model structure. (2) Cr is zero if rainfall (R) is below a threshold Rt (mm/h). This hypothesis was implemented in the FlexRt model structure. (3) We included an interception reservoir (IR) (Figure 5). The rainfall reaching IR fills the storage until the threshold Imax (mm) is reached. The amount that exceeds this threshold determines the effective rainfall Re, which is subsequently routed through the same model as in FlexB. During periods of no rain, water evaporates from IR at a rate EIR equal to the potential evaporation Ep. EUR, now to be regarded as transpiration from vegetation, is independent from the evaporation from interception EIR, and is represented by equation (6), where Lp now represents the threshold for potential transpiration. We called this model FlexI.
 All these model modifications involved the introduction of one extra parameter. The models were evaluated with the multiobjective approach described above. The results of the proposed model modifications are represented in Figure 3. If the model modification resulted in improved performance, the associated Pareto-optimal front would shift toward the origin of the axes. Despite that all proposed hypotheses are plausible, only the FlexI model structure resulted in improved performance. The calibration routine set the optimal values of the parameters URt and Rt to zero. Hence the hypotheses represented by the FlexURt and FlexRt model structures were not confirmed by the data, while the hypothesis contained in the FlexI model structure helped to better explain catchment dynamics. In line with the top-down approach, this suggests that the interception process is in this case an important mechanism affecting catchment response. We note that both FlexB and FlexI can reproduce a correct water balance. In FlexI, evaporation from interception EIR was compensated by a reduction in transpiration EUR, as the calibrated values of LP were larger for FlexI than for FlexB. Table 1 summarizes the contribution to the water balance of the output fluxes from FlexB, FlexI, and of other models that will be introduced further down. It can be observed that IR evaporates about 30% of the total rainfall volume.
Table 1. Contribution to the Total Water Balance of the Output Fluxes of the FlexB, FlexI, FlexId, and FlexId,URd Pareto-optimal Models With Best FNS
 The hydrograph of the FlexI model with the best FNS value is represented in Figure 4. It is possible to observe that while the simulation of peaks is similar to the one of FlexB, the simulation of low flows is improved. The FlexB model produces secondary peaks during low flows that are not observed in reality, while the FlexI model is closer to observations. Apparently, the rainfall generating these small peaks in FlexB is in reality intercepted or added to the soil moisture (see Figure 4). While all proposed model modifications could in principle take this into account, the introduction of an interception store could act selectively on this aspect, without deteriorating model performance in other hydrograph regimes. In section 8.2 we will provide a more extended discussion on role of the interception process.
5.2. Distributed Mode
 While the previous model versions used average rainfall collected over the catchment, we subsequently tested whether accounting for the spatial heterogeneity of the rainfall could improve model results. We subdivided the catchment in nT Thiessen polygons (Figure 1), to be regarded as model units, and subsequently stepwise introduced a distributed description of model components, starting with the interception reservoir, followed by the soil moisture reservoir and other system components. In doing so, we kept the rainfall for each Thiessen polygon as observed, without averaging. In this way the dynamics of the observed rainfall was maintained.
 The first modification consisted of considering for each of the nT model units an interception reservoir (IR) of equal size. While in the other model versions rainfall was first averaged and subsequently routed through the model, in this case an individual, but equally sized, interception reservoir was assigned to each unit (Figure 5). The outputs of the different interception reservoirs are combined and weighted according to the area of each unit. Subsequently, the routing of the effective rainfall through the model proceeds as in the FlexI model, in a lumped manner. The different interception reservoirs have the same maximum threshold Imax and are characterized by the same evaporation rate EIR. Therefore this model does not require more parameters than the FlexI model. However, at any time, each interception reservoir can have a different storage, depending on the associated rainfall history. We called this model structure FlexId.
 Subsequently, we distributed also the unsaturated reservoir (UR). Similarly, the distribution concerns only the reservoir internal states, and not the parameters. Output fluxes from UR are averaged and further processed in a lumped mode. We called this model structure FlexId,URd.
 While the previous structures did not use any information on the specific location of the individual raingauges, the subsequent model structures make use of this information. Through the digital elevation model of the study area, we calculated the average drainage distance to the outlet L (km) of each model unit, and we assumed a non linear relation between this distance and the lag time associated to the fast and slow reacting reservoirs:
 Where j (1, …, nT) indicates a model element, Lmax (km) is the maximum of L1, …, nT, and Tf (h), δf (−), Ts (h), and δs (−) are the parameters characterizing Nlagf and Nlags, previously defined. The model structures that involve the distribution of lag times associated to FR, and to FR plus SR required one additional parameter each. These models were named FlexId,URd,LFd and FlexId,URd,LFd,LSd respectively.
 Distribution of the internal states of FR and SR was not considered because this would not bring any improvement. In fact, a set of linear reservoirs in parallel characterized by the same timescale would behave like a single linear reservoir. As we tried to maintain a low model complexity in terms of the parameter used, we did not attempt to distribute model parameters (besides internal states) for different catchment elements. While it is in principle possible to account for the heterogeneity of catchment characteristics in a conceptual and parsimonious way [Refsgaard and Storm, 1996], there is not much guidance on how to deal with this problem. Addressing this aspect would require an analysis that goes beyond the scope of this paper.
 The performance of the models in response to distributed rainfall input is represented in Figure 6. From the figure, it is possible to see that the Pareto-optimal front shifts progressively toward the origin of the axes, showing which model structural modifications help to better capture catchment dynamics. FlexId performs better than FlexI, indicating that interception, which is a threshold process, is particularly sensitive not just to rainfall volumes, but also to specific rainfall patterns. FlexId,URd performs better than FlexId, which shows that the non-linear partitioning of effective rainfall into infiltration and fast runoff is also strongly affected by rainfall patterns. FlexId,URd,LFd performs only slightly better than FlexId,URd. Apart from model structural errors, this may also indicate that the time to peak of the catchment does not depend much on the area where the rainfall is concentrated. This may be due to the shape of the catchment, which is not particularly elongated, or because in the study area the dominant direction of moving rainstorms is orientated perpendicularly to the direction of the main river, or because other factors, such as for instance lithology, may be more important than the distance of the rainfall event from the catchment outlet. FlexId,URd,LFd,LSd performs like FlexId,URd,LFd, indicating that the distribution of lag times associated to slow flow does not result in any improvement. An explanation for this can be that slow processes, which are characterized by a large stock to flux ratio, are much less sensitive to rainfall spatial heterogeneity than fast processes. This point, as well as our considerations on the tradeoff between model complexity and model performance, will be further analyzed in the discussion section of the paper.
Figure 4 shows the hydrographs of FlexId and FlexId,URd corresponding to the Pareto-optimal models with best FNS. The hydrograph of subsequent structures are not reported as they do not introduce significant improvements, both with respect to objective functions, and in terms of visual inspection. It is possible to notice that while the introduction of an interception reservoir in lumped mode did not produce any improvement on hydrograph peaks, the introduction of distributed interception and the subsequent implementation of distributed soil moisture result in an improvement of peak runoff. This shows that the response of the catchment also during peaks is particularly sensitive to a correct description of the antecedent moisture condition, which depends on the spatial distribution of rainfall and interception.
6. Model Validation
 In order to check if the modifications introduced did not lead to an over-parameterization of the model structure, model performance was evaluated for a two years validation period (September 2000–August 2002). Model validation was performed by running the models FlexB, FlexI, FlexId, and FlexId,URd with the parameter values of the respective Pareto-optimal fronts, using the time series of the validation period as input. Successive model structures were not considered, as they as they did not lead to significant improvements.
 The results are shown in Figure 7. We can see that the stepwise model improvement is confirmed by the validation, in the sense that the clusters of points corresponding to the different structures are progressively shifting toward the origin of the axes. This indicates that the structural modifications do not add unnecessary degrees of freedom to the model.
 Interestingly, the points computed for the validation period do not preserve the arc shape that is displayed during calibration (Figure 7). This is due to the presence of errors both in the model and in the data. Estimating these errors is a difficult task. However, the fact that an optimal model for the calibration period is not optimal during the validation period is a clear indication of their presence. In validation, FlexI did not perform better than FlexB in terms of CNS. This may also be an effect of errors. The other models display better performance in both objective functions.
 While the effect of errors may put into question the role of optimization in hydrology, we still think that this approach is a powerful tool to evaluate model performance and compare different types of models. In the absence of information about the nature of the errors, it makes in fact sense to rely on data, and try to develop models that can closely reproduce them.
 With respect to the estimation of plausible parameter values, optimization may not be a good approach. Concentrating the selection of parameter values on the optimal models only may be too restrictive. Because of the presence of errors, in fact, a wider range of parameters may potentially be as good in representing the system. For this reason, the sensitivity analysis presented in the next section is not restricted to the Pareto-optimal models, but includes a larger number of potentially good models.
7. Sensitivity Analysis
 Sensitivity analysis deals with the question to what extent model outputs are affected by model parameters. Parameters with poor sensitivity cannot be well identified within the parameter space, and this is clearly a problem in hydrological modeling. If a model parameter cannot be identified, it means that there is little confidence in the model's correspondence with reality [Kleissen et al., 1990]. A way to deal with this problem is to reduce model complexity to a level that can still be supported by the data [e.g., Young and Parkinson, 2002]. The main disadvantage of this approach is the risk of oversimplified conceptualization of the processes, which may not be useful for the purpose of modeling. An alternative approach is to find new (orthogonal) constraints [Freer et al., 2004; Vaché and McDonnell, 2006], either through new measurements, or, as shown here, through an improved use of the available information.
 Traditionally, the RSA is based on random sampling of the parameter space. This method however, is not efficient as it requires a large number of runs to cover all portions of the parameter space. Hence more efficient sampling strategies have been recommended [Tang et al., 2007]. In this case, the sensitivity analysis was performed exploiting the same parameter sets generated by the MOSCEM-UA algorithm. The MOSCEM–UA algorithm integrates random sampling and Markov-Chain sampling, which allows a more efficient exploration of the parameter space in the region of the optima.
 The choice of objective functions can largely affect the outcomes of a sensitivity analysis. Most sensitivity analyses are based on the selection of a single statistical summary. This however, can underestimate the information content of the data, and consequently the influence of model parameters on model outcomes. Similar to the advantage of using multiobjective optimization, a multiobjective sensitivity analysis helps to better understand the role of model parameters on simulation. For consistence with the optimization procedure, we used the same (arbitrary) choice of objective functions selected for model optimization.
 The sensitivity analysis was performed on the FlexId,URd model, which we consider the final outcome of the model development process. Subsequent model modifications were discarded, as they did not result in significant improvements. In order to evaluate the constraints that the two objective functions put on model parameters, the analysis was performed stepwise. In a first step, parameter sensitivity was evaluated with respect to FNS only and was expressed determining the parameter samples that correspond to objective function values that are lower than the specified threshold FNS,T. Subsequently, parameter samples were further constrained by discarding the parameter sets with values of FCC that exceeded the threshold FCC,T. FNS,T and FCC,T were arbitrarily specified at 0.15 (corresponding to a CNS of 0.85) and 0.04 respectively. These values were selected in order to include the whole Pareto-optimal front and also sub-optimal models that are close to the Pareto-optimal frontier (Figure 6).
 The specification of a threshold value to separate between adequately and poorly performing models is at the heart of the RSA and of the GLUE approach [Beven and Binley, 1992]. This element of the analysis is based on the consideration that the position of optimal models in the parameter space is strongly dependent on errors in the data and in the model structure. Hence releasing the criterion of optimality in favor of a concept of “equifinality” [Beven, 1993] provides a better picture of the possible combinations that are equally well representative of catchment behavior. The results of the analysis are typically represented through scatterplots which are visually interpreted. Figure 8 displays parameter sensitivity with respect to FNS only (grey dots), and to FNS and FCC simultaneously (black dots). Interestingly, the constraint on FCC further reduces the sensitivity ranges for almost all parameters. The interception threshold Imax demonstrates poor sensitivity with respect to FNS only, while it is further constrained when both objectives are used. The range of sensitivity of this parameter (from 1 to 4 mm) corresponds to an interception rate of up to 30% of the total rainfall, which is consistent with literature values [Link et al., 2004] and observations in the catchment. The experiments performed in the study area showed a somewhat larger estimate of interception (see section 2). However, these measurements were executed on a land use type (beach forest) which is characterized by larger interception rates than the rest of the catchment. Moreover, due to the disparity of scales, the measurements are difficult to generalize to the whole catchment. Nlags shows no sensitivity at all with respect to both objective functions, showing that the processes related to this parameter may represent an unrealistic hypothesis, or that they cannot be assessed through the given combination of data and objective functions.
8.1. On the “Art” of Modeling
 The main objective of hydrological modeling is to explain the variability of catchment response in terms of the factors that may influence it. These factors include temporal and spatial variability of storage distribution, physical characteristics of the catchment, and forcing input. The relations that link the state variables to catchment response are often non-linear, hysteretic and scale-dependent, and their determination form one of the major problems that hydrological science is facing today [Beven, 2006].
 This contradiction is difficult to be solved with the current availability of observation techniques, which prevents us to shed light on the behavior of internal catchment processes. However, a possible way forward is to make better use of the data available, extracting in a more efficient way the information that it contains.
 In this study we identify a possibility to better explore internal catchment behavior in a combination between the top-down approach [Sivapalan et al., 2003] and a multiobjective approach to model evaluation [Gupta et al., 1998]. We show that the combination of these approaches forms a powerful tool to zoom-in from catchment to process scale, and to test the effect of different hypotheses on catchment behavior.
 We demonstrate that increasing model complexity does not necessarily improve model performance (section 5.1). On the other hand, we also show that an improvement of model performance can be obtained just by improved use of the available information, without increasing the number of parameters (section 5.2). Hence it is not the number of model parameters that determines the model ability of reproducing catchment response. Instead, it is the role of these parameters, the processes they represent, and their impact on catchment response.
 This shows that modeling is both and Art and a Science. The science lies in the use of fundamental scientific principles and the formality of analysis; the art accounts for professional experience, insight, creativity and intuition. The latter is particularly important in developing a perceptual and conceptual model that captures the main processes at play, while maintaining minimum levels of complexity. As Wagener et al.  points out “The modeler's task is to draw an inference from the type of failure that has occurred with respect to the hypothesis underlying the specific model component in order to develop an improved version”. This inference is what we call the art of modeling.
8.2. On Interception as a “Dominant Process”
 Several studies underline the importance of interception as one of the key processes affecting the water balance. Vegetation intercepts a significant part of the rainfall, and in humid and temperate climates the amount of water that returns to the atmosphere as a result of canopy interception is estimated to vary between 10 and 50 percent of the annual rainfall [Link et al., 2004, and references within]. If also forest floor interception is taken into account, the amount of intercepted rainfall can be much larger [Gerrits et al., 2007].
 Notwithstanding the important role of interception in the hydrological cycle, this process is often disregarded in hydrological models. Among them: the HBV model as described by Bergström , the REW model [Reggiani and Rientjes, 2005], the Topkapi model [Liu and Todini, 2002], and the HYMOD model [Vrugt et al., 2003b]. These models combine evaporation from interception and transpiration from vegetation into an “evapotranspiration” term. Savenije  argues that this is a conceptual mistake, as interception has different dynamics from transpiration. The timescale of interception is short, as this process ends generally within a period of one day after rainfall [De Groen and Savenije, 2006], while transpiration has a much longer timescale due to the larger storage to flux ratio involved. The energy balance into the ground is in fact much smaller than on the surface, where due to air turbulence rapid rates of evaporation occur.
 Among the reasons for neglecting this process there certainly is the need to keep the model as simple as possible, so as to reduce the number of calibration parameters. Incorporating complexity into a model determines parameter identification problems and does not necessarily result in increased model performance [e.g., Beven, 2002]. Hence there is a need to focus on the “dominant processes” that mostly affect catchment response [Grayson and Blöschl, 2000].
 In some cases the implementation of interception has been related to water balance considerations [Wagener et al., 2004]. However, hydrological models can easily produce a correct water balance compensating for structural errors by parameter readjustment [Andréassian et al., 2004]. This is also shown by our work, where the introduction of evaporation from interception was compensated by a reduction of the transpiration flux. The effect of interception on the efficiency of discharge simulation is controversial. Some authors found that the introduction of an interception component in a model structure improved model performance [Zhang and Savenije, 2005], others experienced deterioration of model performance [Lindström et al., 1997]. In both works, model performance has been assessed through CNS only.
 In this study, we tested several threshold mechanisms for the production of runoff, and we demonstrated that the inclusion of the interception process improves model performance significantly. Hence in this case, interception, besides being conceptually important, is also a “dominant process” that affects catchment response. We note that in the model the interception component separates from the precipitation the effective part that contributes to soil moisture storage or runoff production. Hence while in the literature interception is often defined as a process whereby precipitation is intercepted by vegetation and subsequently evaporated before reaching the ground, in the present case, a more appropriate definition for this case is the one given by Savenije  which considers interception as all the water that is intercepted by the wetted surface (including canopy, understorey, bare soil, litter, roads and build-up areas) and subsequently returned to the atmosphere during or shortly after a rainfall event.
 We link the improvement to a better description of antecedent wetness conditions. In the absence of an interception process, all rainfall is stored in the catchment or produces runoff. However, in reality, the infiltration and runoff production are threshold processes, which are highly sensitive to the wetness of soil surface and of the first centimeters of the soil. A more accurate representation of the storage in the soil and its spatial distribution, therefore, yields a better description of catchment response under variable wetness conditions.
 The divergence of opinions between this and other studies may be due to climatic conditions. In very wet or energy constrained conditions, a split between transpiration and interception may not improve model performance, while in drier conditions the effect of this process on the hydrograph will be more evident. Also, the importance of specific processes is strongly dependent on model evaluation criteria. In this case we built two performance measures to describe model agreement with observations (the Nash-Sutcliffe criterion and the correlation coefficient) and we showed that their combined use helps to better constrain parameter values. The use of additional orthogonal information to test model assumptions, such as, for instance, isotope concentration [Vaché and McDonnell, 2006], may shed more light on the relative importance of the interception process.
8.3. On the Effect of Rainfall Spatial Distribution on Catchment Discharge
 Spatial rainfall patterns can be highly heterogeneous. In general, short duration high intensity rainfall, typically generated by convective thunderstorms, has a very low spatial correlation, whereas the spatial correlation of long duration rainfall is larger. Whether or not spatial heterogeneity of rainfall has an impact on catchment discharge and for what reason, is a problem that has been often addressed in hydrology and that is still poorly understood. The question has been explored directly or indirectly in relation to several issues: the comparison of lumped versus distributed models, the eventual benefit of high density raingauge networks, the advantage of radar derived rainfall data at high spatial resolution, and the improved understanding of catchment internal processes.
 There is much less work that addressed this question using real data, comparing model results with real observations. Interestingly, the majority of these studies seems to point into the opposite direction. While a correct estimate of rainfall volume has a significant effect on model predictions [Andréassian et al., 2001; Sun et al., 2000; Obled et al., 1994] an accurate description of the rainfall spatial pattern does not always appear to be necessary to explain catchment behavior. Most studies comparing distributed versus lumped models [Carpenter et al., 2001; Perrin et al., 2001; Refsgaard and Knudsen, 1996; Michaud and Sorooshian, 1994] including those related to the Distributed Model Inter-comparison Project [Reed et al., 2004] could not detect significant differences in model performance. Obled et al.  conclude that distributed rainfall estimates do not improve discharge predictions, hypothesizing that “the catchment has such a damping capacity that it does not require this type of information”. Smith et al.  support the hypothesis that only catchments with marked rainfall variability and little filtering effect will benefit from distributed rainfall information.
 Our results show that the spatial distribution of rainfall is relevant to explain catchment dynamics. Most importantly, the largest improvements in model performance are obtained without increasing the number of parameters. Hence the improvement is not an attributable of increased model complexity. The interception and unsaturated soil model compartments, which are characterized by a non-linear threshold-like behavior, are particularly sensitive to rainfall spatial distribution. Interestingly, information on the spatial distribution of individual raingauges (introduced by assuming a distribution of lag times associated to raingauge location) did not produce significant improvements. Apart from model structural errors, this may indicate that, in this catchment, spatial heterogeneity of rainfall does not have a strong impact on peak timing.
 In contrast to previous work analyzing this problem using real data, our outcome clearly shows that catchment response is sensitive to spatial heterogeneity of rainfall. One reason that may explain this difference could lie in the different set of hypotheses underlying alternative model structures. Another explanation could be that previous studies analyzed this problem on smaller catchment areas, where the spatial heterogeneity of rainfall is less evident. In general, the effect of rainfall spatial heterogeneity in modeling catchment behavior is clearly linked to model assumptions, catchment characteristics (e.g., catchment area and shape), climate, raingauge location and density. The effect of these controls needs further investigation in future research.
8.4. On the Utility of Distributed Information for Catchment Modeling
 While several authors indicated that the catchment acts as a low-pass filter attenuating the variability of the input signal [e.g., Smith et al., 2004; Andréassian et al., 2001; Obled et al., 1994], we think that such damping effect is also related to the timescale of the processes affecting the rainfall-runoff transformation. Relatively fast processes, that is, those determining the fast response of the catchment to rainfall, demonstrate a higher sensitivity to the space-temporal variability of the input signal than relatively slow processes, such as those determining catchment response during dry weather periods. Our results show that substantial model improvement can be obtained through the distribution of interception, saturated soil, and to a minor extent the lag time of the fast reacting reservoir. All of these aspects contribute to the fast response of the catchment to rainfall. Interception is a process characterized by timescales in the order of several hours, and it is the first shackle in the chain of subsequent processes that determine catchment response. The storage level in soil determines the proportion of effective rainfall which contributes to fast catchment response. No improvement at all is registered on the lag time distribution associated to slow processes.
 This result relates to the debate on the relative merits of distributed versus lumped models. As mentioned before, there is little agreement on whether distributed information helps to improve model predictions. In this respect, we think that the modeling of fast processes, like those related to surface or near-surface runoff, benefits more from distributed information than the modeling of slow processes, such as those related to groundwater flow. For fast processes, the variability of the forcing data is not reduced by the filtering effect of the catchment and is reflected in the variability of catchment response. For slow processes, the variability of the input variables is likely to be filtered out significantly by the averaging effect of the basin, so that long duration rainfall with a low spatial variability can be used, as is often the case, for groundwater related processes.
 In order to better understand internal catchment behavior, without the shortcomings of “virtual experiments”, we adopted an approach that extracts information from real data in a more efficient way than is traditionally done. Our methodology is based on a combination of the “top-down” approach to model development, which is a framework for understanding catchment behavior based on data interpretation, and a “multiobjective” approach to model evaluation, which is based on the consideration that multiple measures of performance are needed to properly extract information from the data.
 The modeling started with a basic model structure applied to the Alzette catchment in Luxembourg. Subsequently, further refinements of model conceptualization were introduced and evaluated, initially in a lumped and then in a spatially distributed mode. We determined that model performance is particularly sensitive to the description of the state of wetness of the catchment. This may seem trivial, but we showed that the improved wetness strongly depends on the process of interception and on the distribution of model internal states in conjunction with distributed rainfall input. These results are of interest to ongoing discussions on which there is little consensus to date. In fact, the interception process, although accounting for an important component of the water balance, is often neglected in modeling application, particularly in relation to hydrograph simulation. Regarding the spatial heterogeneity of rainfall, while theoretical studies with artificial data show that it may have a considerable impact on catchment discharge, most applications using real data support an opposite conclusion.
 Our results contribute to the debate on the relative merits of lumped versus distributed models, showing that fast catchment response benefits more from distributed modeling than slow catchment response. Processes depending on a large stock to flux ratio (e.g., groundwater flow, transpiration, percolation) do not require information on the spatial distribution of rainfall, whereas fast processes, such as interception and surface runoff do. This is due to the damping effect of the basin, which filters the space-temporal variability of the input signal and is larger for slow processes than for fast processes.
 The first author thanks the National Research Fund of Luxembourg for financial support. The authors thank Jeff McDonnell for his suggestions, and Miriam Gerrits, Nico de Vos, Hugo Hellebrand and Tania Cherdantseva for useful discussions.