Sei sulla pagina 1di 12

Agronomie 22 (2002) 217228 INRA, EDP Sciences, 2002 DOI: 10.

1051/agro:2002001

217

Parameterisation P. Durand et al. of hydrological models

Review article

Parameterisation of hydrological models: a review and lessons learned from studies of an agricultural catchment (Naizin, France)
Patrick DURANDa*, Chantal GASCUEL-ODOUXa, Marie-Odile CORDIERb
a INRA,

Unit Sol et Agronomie de Rennes-Quimper, 65 route de Saint-Brieuc, 35042 Rennes Cedex, France b IRISA-INRIA, Universit Rennes 1, Campus de Beaulieu, 35042 Rennes Cedex, France (Received 6 February 2001; revised 14 November 2001; accepted 27 November 2001)

Abstract This paper aims to present, primarily to non-hydrologists, the problems of parameterisation in hydrological modelling, and the attempts made to solve them. First, the main approaches used in hydrological modelling are presented, with special emphasis given to the number of parameters required and their physical meaning. The semi-distributed models, based on simplified physical laws and a functional spatial description of the catchment, to limit the parameter set, are taken as an illustration of the problems of parameterisation in hydrology, through studies performed on a cultivated catchment in Brittany (Western France). These studies showed: (i) the effect of spatial and temporal resolution on the parameter values and their physical meaning; (ii) the fact that many different parameter sets can produce equally acceptable discharge predictions, and (iii) the interest of a Bayesian approach to incorporating additional information in the calibration process. As a perspective, exploration of other modelling approaches is suggested to look for new solutions to the parameterisation problem, putting emphasis on the modularity of the models and on the qualitative approach. modelling / simulation / calibration / parameter / hydrology

Rsum La paramtrisation des modles en hydrologie : synthse et enseignements tirs de ltude dun bassin versant agricole (Naizin, France). Cet article, destin surtout aux non hydrologues, prsente les problmes poss par la paramtrisation des modles hydrologiques et les rponses apportes. Les principales approches de modlisation hydrologique sont dcrites, essentiellement sous langle du nombre de paramtres considrs et de leur sens physique. Les modles semi-distribus, bass sur une simplification des lois physiques et des schmas de spatialisation du bassin versant, pour restreindre le jeu de paramtres, sont utiliss pour illustrer le problme de la paramtrisation des modles en hydrologie, partir des travaux mens sur le bassin versant de recherche de Naizin (Morbihan, France). Nous montrons ainsi : (1) lincidence de la rsolution spatiale et temporelle sur la valeur des paramtres et leur sens physique ; (2) la non-unicit du jeu de paramtres permettant une prdiction acceptable des dbits ; (3) lintrt de lapproche baysienne pour mieux contraindre les jeux de paramtres et formaliser la pluralit des jeux de paramtres possibles. Dautres approches de modlisation saffranchissant partiellement du problme de paramtrisation sont esquisses, bases sur : (1) une architecture plus modulaire des modles, permettant une identification des paramtres dcouple du modle ; (2) une approche qualitative de la modlisation, raisonnant laccessibilit des paramtres, lusage et la transparence du modle. modlisation / simulation / calibration / paramtrisation / hydrologie

Communicated by Daniel Wallach (Castanet-Tolosan, France) * Correspondence and reprints durand@roazhon.inra.fr

218

P. Durand et al.

1. INTRODUCTION

Historically, hydrological modelling was aimed at predicting discharge as a function of the catchment physiography and climate, essentially for operational purposes such as the design of hydraulic structures. Simple empirical models were usually quite adequate for such an objective. Progressively, as the impact of human activities on water resources quantity and quality has become the main issue for hydrologists, process-based models have been developed [1, 13, 23, 25, 26]. Models stopped being mere operational tools and became an essential part of the heuristic development of hydrology as an environmental science. As a consequence, hydrological modelling studies have not only consisted of model development and applications: numerous papers have been published on fundamental modelling issues such as validation and testing, coherence, consistency and parameterisation [7, 8, 25, 26, 29, 32, 36, 37, 39, 40]. We feel that it could be worth sharing this expertise with scientists from other environmental sciences confronted with similar problems. The difficulty of modelling catchment hydrology lies primarily in that the response of the system is strongly controlled by its heterogeneity, and this heterogeneity cannot be completely known and described. The transformation of a rainfall signal into a discharge signal involves processes that are, as a rule, non-uniform and nonlinear in time and space. The initial conditions and the dynamics of the rainfall event will activate or hinder a particular process (e.g., infiltration excess overland flow), involving retroactions or threshold effects. This will occur in parts of the catchment that will contribute, or not, to the overall response of the system [2]. This heterogeneity has a deterministic component (topography, soil types, etc.) and a stochastic one (distribution of macropores, microrelief, etc.). On top of that, the output variable that is usually measured and simulated (stream discharge or chemistry) is most often damped out and contains very little information on the internal catchment processes [18, 34]. Given that, the hydrological modeller has the choice between two alternatives: either he tries to describe the complexity of the system, or he focuses on a good simulation of the output response. Both the alternatives are fraught with acute parameterisation problems. These difficulties are well illustrated by the case of the hydrological modelling of agrosystems. In such systems, the objectives of the modelling approach are seldom restricted to the simulation of the discharge: crop production and/or water quality issues are most often

considered. The agricultural landscape, especially in the European temperate zone, is usually heterogeneous in space and time, due to the distribution of the crops and the succession of the agricultural practices. It is also structured by the limits of the fields, the drainage network, the paths, etc. Usually, the modelling exercise will aim at assessing the effect of changing agriculture activities on water quantity and quality. This usually involves the definition of scenarios. In such a framework, the predictive ability of the model is critical, as well as its sensitivity to the internal or external changes introduced to the scenarios. This requires a careful analysis of the parameterisation of the model. In this paper, some issues of the parameterisation of hydrological models will be reviewed for the non-hydrologist reader and illustrated in the case of an agricultural catchment. The different modelling approaches will be presented and compared in terms of parameter identifiability. Then the paper will focus on two models that have been applied to this agricultural catchment with particular attention on the issue of parameterisation, using three different approaches [4, 14, 21]. First, a sensitivity analysis of the parameters to time and space resolution will show the limits of the physical significance of the parameters. Second, the concept of equifinality, i.e., the fact that different parameter sets may provide equally acceptable simulations will be illustrated using a Monte Carlo approach. Third, this equifinality will be systematised using a Bayesian approach, called GLUE. Finally, the interest of new hydrological modelling approaches for agricultural catchments will be discussed.

2. THE MAJOR TYPES OF HYDROLOGICAL MODELS

Most of the models used in hydrology fit into one of the following categories (Fig. 1). Black box, empirical models (BBMs): the relationship between the input signal and the output signal is derived from long series of observations, using statistical (time series analysis, multivariate regression) or computing (neural network) tools [15, 28]. Lumped conceptual models (LCMs): the catchment is considered as a whole, and the system is modelled using a conceptual, metaphorical vision, e.g. a cascade of reservoirs [19, 35] or a simplified physical system [31].

Parameterisation of hydrological models

219

DPM

Figure 1. The main modelling approaches in hydrology. x stands as undefined function, f as conceptual relationship, as physically-based equation.

Semi-distributed, physically-based models (SDMs): the catchment is divided into functional subparts, which can be areas that are homogenous with respect to one or more characteristics (land use, soil type...), or structural elements (slope, drainage network...), or classes of points having the same hydrological behaviour (hydrological similarity concept). The fluxes are computed using physical equations that are simplified thanks to some assumptions and solved analytically [12]. Distributed, process-based, numerical models (DPMs): the catchment is described using a grid (regular or not, square or not). The fluxes are computed using the partial differential equations of the physics of porous media and hydraulics, solved by a numerical scheme (finite differences or finite elements) [1, 6]. There are also hybrid models, that include features of two of the above types. For example, some models are fully distributed, but based on simplified equations [4]. The parameterisation of the BBMs is a question that is not specific to hydrological models and will not be treated here (for an application to hydrology, see [27]). The LCMs usually have few parameters (two to five). The parameter values are usually fixed through a calibration process. However, some studies have sought to derive parameter values for ungauged catchments, most often on a regional scale, based on generic features of the

catchments [29]. In the case of calibrated parameters, the main problem is to define the domain of applicability of a particular parameter set, owing to the uncertainty in the physical meaning of the parameters. If the calibration data set is limited, it is generally possible to get a very good fit, and the model prediction will be very accurate as long as the application does not differ from the calibration conditions. However, the quality of prediction is likely to drop when the model has to cope with conditions outside the range of the calibration. If the parameters are calibrated with a data set that encompass a large variety of conditions, even including changes within the catchment, the validity domain will be larger, but the goodness of fit will be generally poorer. If the model is used for investigative purposes, such as assessing the effect of a change within the system, the best approach will be to calibrate with a limited data set, and then to look at the lack of fit between simulations and observations after the change. The alternative that consists of doing two calibrations, before and after the change, and then comparing the parameter values of the two sets, is probably less rigorous because (i) the interpretation of the physical meaning of the parameters is difficult, and (ii) the calibration procedure, either manual or automatic, does not guarantee that the final parameter set is unique and optimal (see part 3). The DPMs usually have numerous parameters, sometimes variable in space or time. The developers of DPMs claim that this type of model does not require calibration, since the parameters have a physical meaning and are measurable. Beven [7, 10] has demonstrated that this is not really the case, essentially for three reasons. Firstly, the equations used have been established on a small scale for homogeneous systems: there is no proof that the same equations, with the same parameters, apply to the grid scale, owing to the high heterogeneity of the system and to different sources of non-linearity. Secondly, even if the equations are valid, effective parameters must be defined on the grid scale; there is not, to date, a general method for deriving these effective parameters from point measurements that are often very variable. Thirdly, it is, in practice, impossible to measure all the parameters required for each grid square. Consequently, DPMs are most often calibrated, at least for several parameters. This calibration is difficult since these models are over-parameterised in a system simulation sense, i.e., the amount of information in the observed data is not sufficient to completely determine the calibration, and the interactions between parameters are important. In the process of the calibration, one can suspect that the physical meaning of the parameter is

220

P. Durand et al.

partially lost. In spite of these limitations, the performance of todays computers allow an increasing use of these models, and they seem to be satisfactory in different systems with limited calibration. The SDMs, that are also widely used, appear to be a good compromise between LCMs and DPMs. The number of parameters is often limited, and the process-based approach gives a physical meaning to them, even if they cannot be directly measured, in general. However, the dangers of the calibration mentioned above still apply, and a careful analysis of the parameterisation is needed. This type of model is used here to illustrate different aspects of this analysis.

3.2. Description of the models Two models were used. The first one, TOPMODEL [12], is a SDM and the second one, the hydrological module of a TNT model [5], is intermediate between a SDM and a DPM. It is fully distributed but simplifies water transfer using some of the simplifying assumptions of TOPMODEL. The basic assumptions and the parameters of the two models are now presented. TOPMODEL is the semi-distributed model most widely used as a heuristic tool. This model uses a topographic index as a measure of hydrological similarity, which makes it possible to simulate a simplified spatially distributed response of the catchment: all the pixels that have the same index have the same hydrologic response. This concept of similarity is based on three major assumptions related to the dynamics of the shallow groundwater: first, the groundwater flow is represented adequately by a succession of steady states; second, the recharge to the groundwater is uniform in the catchment; third, the local hydraulic gradient can be considered as constant and equal to the local surface gradient. This topographic index is derived from a Digital Elevation Model (DEM, also called DTM, digital terrain model), i.e., a matrix of terrain elevation at each node of a square grid. Among the various indexes proposed, the most widely used is the ratio between the cumulative upslope area and the local slope gradient, assimilated to the hydraulic gradient, and called hereafter Beven-Kirkbys index. Detailed descriptions of the model are given elsewhere [12, 38]. Four lumped parameters are usually calibrated in this model: M (in metres) is a coefficient of the exponential decrease of hydraulic conductivity with depth; SKO (in metres by time step) is the hydraulic conductivity at the soil surface; SRMAX (in metres) is the maximum storage in the root zone, which is used to control actual evapotranspiration rates; CHV (in metres by time step) is the flow velocity in the channel network. Most often, the parameters are calibrated for a period, and a validation run is performed on another period. The objective function used is a measure of the goodness of fit between observed and simulated discharge. The most frequently used is the Nash-Sutcliffe efficiency criterion [33] defined as follows: Efficiency = 1 (((Qobs Qsim)2) / ( (Qobs Qmoy)2)) where Qobs, Qsim and Qmoy are observed, simulated and mean discharge, respectively. The efficiency is 1 for a perfect fit, and negative if the simulation is worse than the constant mean discharge. The other outputs of the model, i.e. the proportion of overland flow, the extension

3. HYDROLOGICAL MODELLING OF THE KERVIDY-NAIZIN CATCHMENT

3.1. The Kervidy-Naizin catchment The experimental catchment of Naizin, 12 km2 in area, and the subcatchment of Kervidy, 5 km2, are located in Brittany, France. Extensive studies in soil science, geochemistry, hydrology and agronomy have been performed on this site [16]. The climate is humid and temperate. The bedrock is Brioverian Shales, with a weathered layer of very variable thickness (10 m in average). The slopes are gentle, usually less than 5%. The soil texture is silty loam with a depth varying from 0.5 to 1.5 m. The soil distribution consists of an upslope well drained domain and a poorly drained lower ground that covers 16% of the total surface. Land use is essentially intensive farming. The surface is equally covered by maize, cereal and pasture. The average annual precipitation over the last 30 years was 909 mm. Rainfall is distributed fairly evenly during the year (maximum and minimum monthly means are 116 mm in January and 45 mm in July, respectively). The annual discharge at the Kervidy outlet varied from 135 mm to 685 mm between 1992 and 1998. The stream may dry out during the summer period for one or two months. In this climatic and geological context, the fluxes in the river are controlled by the seasonal and event dynamics of the shallow groundwater. During winter, the groundwater table is at the soil surface in the lower ground. It is close to the surface (one to a few metres) on the slopes.

Parameterisation of hydrological models

221

of the saturated areas or the groundwater table, allow further checking of the consistency of the modelling. Since the parameters are supposed to have a physical meaning, and therefore to be independent of the period considered (provided the system remains intrinsically invariant), the model is considered as acceptable if the use of the calibrated parameter set for a distinct period, preferably featuring conditions outside the range of the calibration period extremes, does not result in a dramatic decrease of the efficiency criterion. The hypothesis that the local hydraulic gradient in the groundwater is equal to the surface slope is roughly acceptable on the Kervidy-Naizin catchment and justifies the use of TOPMODEL. This model has been applied in different ways and for different purposes to the Kervidy catchment: short term dynamics of discharge and saturated areas, prediction of the location of hydromorphic soils, seasonal dynamics of groundwater. In each application, the model was considered as consistent with the major hydrological processes of the catchment, although in detail some discrepancies between simulated and actual dynamics were noted. The TNT model (Topography-based Nitrogen Transfer) has been developped to study the effect of the spatial distribution of agricultural nitrogen input and of its biotransformations on the overall nitrogen fluxes at the outlet of an agricultural catchment. Basically, the hydrological module is a fully distributed version of TOPMODEL, adapted to solute transfer. It is based on the same assumptions, except the uniform recharge, which is not necessary in a distributed mode. The adaptation consisted, first, of expressing the water storage absolutely and not as a deficit, as in TOPMODEL; second, in defining a threshold of groundwater storage above which the groundwater re-enters the root zone (subsurface flow), thus allowing the groundwater nitrate to participate in biotransformations in the valley bottom [4, 5]. The model has six parameters: M, SRMAX and SKO, as in TOPMODEL; T0 (the lateral saturated transmissivity), Max_nappe (Max groundwater, the upper limit of the groundwater store, in m of water) and Seuil (the threshold for subsurface flow). The parameterisation of these two hydrological models was studied following three approaches: (1) a sensitivity analysis of the parameters to space and time resolution of the modelling, and the consequence on their physical meaning; (2) a sensitivity analysis using a Monte-Carlo procedure, allowing the parameters to vary randomly inside physically sensible limits, then trying to reduce the number of acceptable parameter sets by

physical interpretation; (3) a Bayesian approach for a multi-criteria calibration that restricts the range of acceptable parameter values. TOPMODEL was used for the first and the third approach, and the hydrological module of TNT was used for the second approach.

4. IDENTIFICATION AND PHYSICAL SIGNIFICANCE OF PARAMETERS IN HYDROLOGICAL MODELLING

4.1. Values of the parameters versus space and time resolution of the modelling 4.1.1. Method Aside from classical sensitivity analysis where the effect of each parameter, or a combination of parameters, on simulation results is described, it is important to study the possible influence of other model features, in particular those that can affect some major assumptions [24, 25]. In the case of TOPMODEL, the time step adopted for the simulations is one of those, because of the quasi-steady state assumption: if a steady state approximation for a few minutes seems reasonable, is it still true for one day? Another question is the effect of the resolution of the grid used to derive the topographic index, which influences the estimation of the local slope and of the drainage area included in the index, and so the index distribution class. This analysis was published by Bruneau et al. [14]. The influence of the grid size was analysed using four DEMs independently calculated from scanned 1/10 000 contour data. Their grid sizes are respectively 100, 70, 50 and 30 metres. The flow pathways were defined on the catchment and different topographic variables were computed for each cell (slope, cumulative upslope area, Beven-Kirkby index). The influence of the time step was analysed from a study of the variations in mean and maximum values of precipitation intensity and discharge, according to the time step. Seven steps were distinguished: 30 minutes, one, two, four, seven, twelve and twenty-four hours. The influence of time and space resolution on the calibrated parameters was analysed. The optimised values were determined automatically using a Rosenbrock optimisation scheme that minimises the objective function, here the sum of square errors. A grid size of 30 m was selected to analyse the effect of time resolution, and

222

P. Durand et al.

time step variations. When the grid size increases, the topographic index distribution is shifted towards higher values, because the local slope is underestimated, which means that the saturated area for a given average catchment wetness increases, resulting in higher discharge peak, due to more overland flow. The increase of SKO tends to compensate for this: increasing the hydraulic conductivity decreases the soil saturation. SRMAX (maximum storage in the root zone) increases rapidly with time step then levels off, while it fluctuates without a clear trend when grid size varies, showing a particularly high value for the 100 m. Since the calibration period was a rainy period in December and January, SRMAX affected the first storm events only, decreasing the amount of hydrologically efficient rainfall. Apparently, the automatic calibration procedure used this parameter artificially to decrease the peak discharge due to large grid resolutions. The same is true for CHV (channel flow velocity): its effect is to delay the flow created by the rainfall and to distribute it over the next time steps. The low optimised value obtained for the larger grid size resulted in a decrease of the peak discharge. The variation of this parameter with time is the same as the variation of SKO, i.e., a sharp decrease, approximately opposite to the increase of SRMAX. Clearly, there is here an interaction between the three parameters, the decrease of SKO (that increases the peak discharge) being compensated by an increase of SRMAX and CHV. Moreover, it should be noted again that over a time step of one or two hours, the steady state assumption probably becomes false for the storm flow periods. In theory, the physical meaning of these three parameters is as clear as for the first one. In practice, the automatic calibration process, as well as the use of poor input variable resolution, resulted in a loss of this meaning. This sensitivity analysis shows the necessity of being very cautious when attaching too much physical meaning to parameters, especially after an automatic optimisation procedure and when the model is applied at the margin of its domain of validity. The drawback of using a simplified model, with few parameters, is that these parameters partly integrate different hydrological processes and so depend on space and time resolution, input data and initial conditions. A good way to limit this dependency is to use these parameters in a limited range of space and time resolution and to consider them rather as effective parameters than as measurable physical properties, as shown by Beven [9]. The next example chosen will

Figure 2. Variations of the optimised calibrated values of the TOPMODEL parameters versus time step (up) and grid size (down), for a grid size of 30 m and a time step of 1 h, respectively (after Bruneau et al., 1995). M: coefficient of the exponential decrease of hydraulic conductivity with depth; SKO: hydraulic conductivity at the soil surface; SRMAX: maximum storage in the root zone; CHV: flow velocity in the channel network.

a time step of one hour to analyse the effect of space resolution (Fig. 2). 4.1.2. Results The hydraulic conductivity decay rate, M, increases slowly with time step, and is hardly affected by the grid size variation. This parameter has two physical meanings: Saulnier [38] has shown that it was closely linked to the depth beyond which the water movements are negligible, while Ambroise et al. [3] have shown how it controls the shape of the recession curve. It seems therefore quite logical that this parameter is little affected by space and time resolutions. The variations of the three other parameters are more marked, for complex reasons that will not be given in detail. Only some illustrations will be provided to point out the effect of the calibration and of the space and time resolution on the physical meaning of the parameters. SKO (hydraulic conductivity at the soil surface) increases with increasing grid size, while it fluctuates with

Parameterisation of hydrological models

223

describe in more detail the problem of adjusting the parameters by calibration. 4.2. The concept of equifinality 4.2.1. Definition and methods The calibration of a model, either manual or automatic, aims at determining the set of parameters that maximises an objective function. In hydrology, this objective function is usually a measure of the goodness of fit between observed and simulated discharge. Implicitly, this suggests that there is a unique set of parameters that fulfils this condition. In practice, even in the case of models with few parameters (three to five), some interaction between parameters can occur, and different sets of parameters will give equally acceptable results. A sensitivity analysis performed for the six parameters of the TNT hydrological module is provided as an illustration of this problem, called equifinality by Beven and Binley [11]. The analysis was conducted in two steps: first, a set of simulations was run using a MonteCarlo procedure, allowing the parameters to vary randomly and independently within physically sensible limits; then, the physical interpretation of the model was used to explore the interactions between the parameters. 4.2.2. Results The results of the analysis for the main parameters are presented in Figure 3, where the parameter values are plotted versus the efficiency of the discharge simulation for the 2000 Monte Carlo runs. The most striking result is that for each value of a single parameter within the range of possible values, there is at least one (and often much more) combination of parameters that gives a reasonable fit (Nash criterion above 0.85). The same type of result was found by Beven and Binley [11] for TOPMODEL. The main reason for this is the compensation between the parameters. For example, Figure 3e shows that the values of M and T0 that produce acceptable simulations are linked. The product M.T0 is the integral of the transmissivity for the whole range of saturation deficit. When this product is too high, the Nash-Sutcliffe criterion stays around 0.36, corresponding to an almost constant and very low discharge: the material is too permeable and no rapid flow occurs. Further reasoning on the physical meaning of the parameters made it possible to reduce the degree of freedom of the variations of the parameters. For example, it is very likely that M

(linked to the thickness of permeable material, as noted in the previous section) and max_nappe (thickness of the aquifer) are linked. Actually, Figure 3f shows that if max_nappe is much higher than M, the groundwater table remains too deep to produce rapid flow, and the Nash-Sutcliffe criterion is again close to 0.36. Given these interactions between the parameters, the modeller is given an alternative: either he reduces the number of parameters, which will make the calibration easier and the equifinality less problematic, but also, which will need the introduction of more assumptions and will make the model more rigid; or he tries to put more constraint and more information in the calibration procedure, to reduce the number of acceptable parameter sets. The next part illustrates the second alternative.

4.3. GLUE, a Bayesian approach to parameterisation 4.3.1. Method The Generalised Likelihood Uncertainty Estimation (GLUE) approach was introduced in hydrology [11, 20, 22] to define the performance of the range of possible models or parameter sets in terms of some likelihood measures. GLUE is based on a Monte Carlo approach. A large number of model runs are made, each with random parameter values selected from probability distributions. The acceptability of each run is assessed by comparing predicted to observed discharges in terms of likelihood. A wide range of likelihood measures may be used as appropriate to an application or data set. The only constraint is that the value of the likelihood measure should increase monotonically with increasing performance of the model. A threshold of acceptability is defined. Runs that achieve a likelihood below this threshold are then considered to be non-behavioural and have their likelihood values set to zero. They are removed from the subsequent analysis. The threshold value, as well as the type of likelihood measure, are chosen by the operator: this choice is of primary importance and one of the interests of the method is to focus the attention on this step of the analysis. Following the rejection of non-behavioural runs, the likelihood weights of the retained runs are rescaled so that their cumulative total is 1.0. At each time step the predicted output from the retained runs are likelihood weighted and ranked to form a cumulative distribution of the output variable: the main output of the analysis is an estimation of the uncertainty of the model predictions.

224

P. Durand et al.

Figure 3. TNT parameter values versus efficiency (Nash-Sutcliffe criterion), for 2000 Monte Carlo simulations (after Beaujouan, 2001). M: coefficient of the exponential decrease of hydraulic conductivity with depth; T0: lateral saturated transmissivity at the soil surface; SRMAX: maximum storage in the root zone; Max_nappe: upper limit of the groundwater store.

This analysis focuses on parameter sets rather than the behaviour of individual parameters and their interactions. It is possible and recommended to incorporate into the analysis different types of variables, discharge of

course, but also the depth of the water table, the soil moisture, the extension of the saturated area in the catchment, etc.

Parameterisation of hydrological models

225

4.3.2 Results The method was applied to TOPMODEL on the Kervidy-Naizin catchment combining both discharge observations and remote sensing data [21]. ERS-1 scenes provided an estimate of the saturated area by means of an index based on radar data treatment computed with a time step of 3 days, during a few months. The likelihood weight employed in this study was defined by the following likelihood measure: (1) L (i / X) = (1 2i / 2obs) N where L (i / X) is the likelihood for the parameter set i , 2i the variance of the errors with the parameter set i given the set of discharge observations X, 2obs the variance of observed data set and N a shaping factor. This likelihood measure is equal to a coefficient of determination or the Nash-Sutcliffe efficiency criterion when N = 1 [33], which is the case in the results presented here. Higher values of N accentuate the weight of the better simulations, which is better for finding an optimum set, but not necessarily for determining the uncertainty bounds. This likelihood measure does not exactly correspond to the definition used in likelihood theory. For a detailed discussion on this point, see [19] and [20]. In order to constrain the uncertainty better, additional data are incorporated into the likelihood estimates. The new likelihood distribution may be achieved through the application of the Bayes equation in the form: (2) L (i / Y) = L1 (Y /i) L0 (i) / C where L0 (i) is a prior likelihood measure for the parameter set i, L1 (i / Y) is the likelihood measure calculated for the simulation of observed variable Y by the parameter set i , L1 (i / Y) is the posterior likelihood for the parameter set i given the new observations Y, and C is a scaling factor. Again, the reader should refer to [20] for a detailed discussion on the validity of this formula to the likelihood function used. The parameter sets were evaluated with respect to the comparison of predicted and observed discharges through application of (1). Bayesian updating of likelihood was then performed through application of (2) with respect to the predicted area extent of saturation compared to the uncertainty range of actual saturated area extent as inferred from extrapolation of the limited saturation ground through radar data treatment. This procedure induces a rejection of parameter sets that produce non-behavioural responses in terms of saturated areas. The range of values of transmissivity that produce a good fit for discharge and a reasonable internal response is much narrower than when discharge only is considered

Figure 4. TOPMODEL parameter values versus discharge simulation efficiency (Nash-Sutcliffe criterion), and versus discharge and saturated area extension simulation efficiency after rejection of non-behavioural simulations (after [21]). M1 and M2: two coefficients of the exponential decrease of transmissivity with depth, for upper and deep layer, respectively; T0: lateral saturated transmissivity; SRMAX: maximum storage in the root zone.

(Fig. 4). The other parameters are not constrained to the same extent, indicating a lack of sensitivity of the predicted saturated area to these other parameters. This methodology is well adapted to the incorporation of additional data in the calibration process. In addition, it makes it possible to estimate the uncertainty of prediction of the different variables, once the parameter sets have been better constrained. The major limitation is that

226

P. Durand et al.

it requires considerable computer time, which makes it difficult to apply systematically, especially for fully distributed models.

5. PERSPECTIVES FOR NEW MODELLING APPROACHES Most of the problems of parameterisation presented above arise during the calibration process. A different modelling approach, more qualitative [17], could perhaps be used to make this calibration unnecessary or less important. This could be particularly adapted for models used in decision-making. The model properties that are valued in this approach are: insensitivity: the simulations must not be too sensitive to parameter variations that are within the range of uncertainty attached to an independent estimation of these parameters. In some cases, these parameters may be only known qualitatively; transposability: the model needs to be applicable to a new site (within its validity domain) without requiring new calibration or expensive data acquisition; transparency or explicativity: when the model is used to compare scenarios, the link between the simulation results, the forcing variables and parameters must be explicit and explanatory. The decision will come from the analysis of this link, and must be valid for an area and a period compatible with the policy-making. Of course, these properties cannot be completely fulfilled, and must be balanced in each particular application; availability of information: the model should use information (data, parameters) that can be made available for the user. If this information cannot be provided directly by reasonably simple measurements or observation, the model should provide an explicit way of estimating them. The qualitative modelling approach, which encompasses a wide range of models and applications, will not be described here. In short, the main idea is that a continuous, quantitative model may not be necessary when qualitative results are required and when the system is not mechanistically knowable. For example, to compare the effect of different agricultural practices on nitrogen pollution in a catchment, it may not be necessary to simulate quantitatively the whole nitrogen and water cycle to calculate nitrate concentration in the river at each time step.

We have begun to develop some qualitative models in the field of pollutant transfer in agricultural catchments that try to comply with these requirements. The example presented here deals with the transfer of pesticides, to study the effect of application date and molecule choice, as related to soil surface conditions and wetness of the catchment, on the loss of pesticide from the field. The inputs of the model are the rainfall time series and the characteristics of the pesticide applications in the field: amount, date, molecule; the outputs are time series of concentrations and fluxes of pesticides out of the field. Numerous processes interact: overland flow generation and transfer, leaching through the soil, volatilisation, degradation, retention by vegetated buffer strips. Given the lack of precise, quantitative knowledge on the physics and chemistry of this complex system, the relevant and simple relationships between the variables have been expressed in a simple way, and sometimes qualitatively, from established literature results. For example, the infiltrability threshold of the soil, required to estimate the partition between runoff and infiltration, varies in time and cannot be measured simply. Here, this parameter is initialised, as proposed by Lecomte [28], from expert rules based on readily observable data: vegetation cover and soil surface conditions (roughness and crusting stage). The evolution of the parameter with time can be calculated using simple function such as the sum of temperature (for the vegetation cover) and the amount of rainfall (for the soil surface conditions). Alternatively, this evolution can be inferred by direct observations, or simulated by another model, depending on the data available and on the accuracy expected. Using readily available parameters and data, this model makes it possible to compare different scenarios of pesticide application, linking explicitly the variables included in the scenarios and the simulation results. The model is still at the development stage. Basically, what is new in this approach is that the emphasis is put on the adequacy between the objectives of the modelling (i.e., to compare scenarios and identify the more significant factors of variations), the information available and the model structure. It is complementary to classical modelling approaches: the physical models can be used to investigate a process or a series of processes on a well-equipped study site, and the qualitative approach could use this knowledge to test scenarios in a variety of situations. The parameterisation issues addressed in the previous sections are not necessarily avoided in this type of model, but they are, at least, made explicit. For example, there is

Parameterisation of hydrological models

227

no global calibration in the sense of fitting observed for simulated data for the whole model. When calibration is required, it is performed for individual modules. Generally, the results of the simulations are not interpreted in terms of goodness of fit, but rather by trying to explain the differences between the model output and what is observed or expected.

Although these three directions are perfectly valid and interesting, we feel that a fourth one still requires further attention and creative investigation. This concerns investigating the adequacy of the architecture of the model, the amount and the quality of information available to implement and test it, and the amount and the quality of information really needed by the end user.
Acknowledgements: This paper was written following an invitation to the Toulouse meeting by D. Wallach. The authors are very grateful to him for the invitation and for the editing work he did on the paper.

6. CONCLUSION Hydrology is a science that has developed under the pressure of practical necessity, to give mankind the tools for managing and protecting water as a resource, and the tools to protect itself from the adverse effects of the uneven distribution of this resource in time and space. In that context, hydrological modelling is both a predicting tool, for operational purposes, and a heuristic tool, to test hypotheses and help to understand processes. This duality is responsible for the wealth of models and the variety of modelling approaches, which has made it necessary for the hydrologist to investigate, to question and to criticise his modelling practices. This paper has presented some examples of this questioning, through the problem of parameterisation. To make a long story short, the problem is not to build a model that fits the field data, but one that fits them for the right reasons [26, 30]. The major difficulties are common to many environmental sciences: the system cannot be described in all its complexity and heterogeneity, although taking this heterogeneity into account is compulsory in order to understand the systems behaviour; the processes cannot be completely put into equations, because the heterogeneity of the system causes nonlinearity on all the scales; the data available are generally incomplete, uncertain, and do not bear enough information to constrain adequately a complex model. Faced with these difficulties, the modellers have tried (i) to limit the parameter set by means of conceptualisation of the processes formulation and/or of the system description, (ii) to give to the parameters a physical meaning, so that they can be assessed independently of the model structure and of the study site and (iii) to develop the theoretical frame and the technical tools necessary to incorporate new information into the system within the calibration process, and to quantify the uncertainty limits of the prediction.

REFERENCES
[1] Abbott M.B., Bathurst J.C., Cunge J.A., OConnell P.E., Rasmussen J., An introduction to the european hydrological system systme hydrologique europen, SHE, 1: History and philosophy of a physically-based, distributed modelling system, J. Hydrol. 87 (1986) 4559. [2] Ambroise B., La dynamique du cycle de leau dans un bassin versant. Processus, facteurs, modles, Editions *H*G*A*, Bucarest, 1999. [3] Ambroise B., Beven K.J., Freer J., Toward a generalization of the TOPMODEL concepts: topographic indices of hydrological similarity, Water Resour. Res. 32 (1996) 21352145. [4] Beaujouan V., Modlisation des transferts deau et dazote dans les sols et les nappes. Dveloppement dun modle conceptuel distribu. Application de petits bassins versants agricoles, Thse cole Nationale Suprieure Agronomique de Rennes, 2001. [5] Beaujouan V., Durand P., Ruiz P., Aurousseau P., Modelling the effect of spatial distribution of agricultural practices on nitrogen fluxes in rural catchment, Ecol. Model. 137 (2001) 93105. [6] Beven K., Distributed models, in: Anderson M.G., Burt T.P. (Eds.), Hydrological Forecasting, Wiley and Sons Ltd., 1985, pp. 405435. [7] Beven K., Changing ideas in hydrology. The case of physically-based models, J. Hydrol. 105 (1989) 157172. [8] Beven K.J., Prophecy, reality and uncertainty in distributed hydrological modelling, Adv. Water Resour. 16 (1993) 4151. [9] Beven K.J., Linking parameters across scales: subgrid parametrizations and scale dependent hydrological models, Hydrol. Proc. 9 (1995) 507525. [10] Beven K.J., The limits of splitting: hydrology, Sci. Total Environ. 183 (1996) 8997. [11] Beven K., Binley A., The future of distributed models: model calibration and uncertainty prediction, Hydrol. Proc. 6 (1992) 279298.

228

P. Durand et al.

[12] Beven K., Kirkby M.J., A physically based, variable contributing area model of basin hydrology, Hydrol. Sci. Bull. 23 (1979) 419438. [13] Blsch G., Sivapalan M., Scale issues in hydrological modelling: a review, Hydrol. Proc. 9 (1995) 251290. [14] Bruneau P., Gascuel-Odoux C., Robin P., Mrot P., Beven K., Sensitivity to space and time resolution of a hydrological model using digital elevation data, Hydrol. Proc. 9 (1995) 6981. [15] Campolo M., Andreussi P., Soldati A., River flood forecasting with a neural network model, Water Resour. Res. 35 (1998) 11911197. [16] Cheverry C., Agriculture intensive et qualit des eaux, INRA Edition, 1998. [17] Coquillard P., David R.C. Hill, Modlisation et simulation dcosystmes : des modles dterministes aux simulations vnements discrets, Masson, 1996. [18] Durand P., Neal C., Neal M., Variations in stable isotopes and solute concentrations in small mediterranean montane catchments, J. Hydrol. 144 (1993) 283290. [19] Franchini M., Pacciani M., Comparative analysis of several conceptual rainfall-runoff models, J. Hydrol. 122 (1991) 161219. [20] Franks S.W., Beven K.J., Bayesian estimation of uncertainty in land surface-atmosphere flux predictions, J. Geophys. Res. 102 (1997) 2399123999. [21] Franks S.W., Gineste P., Beven K.J., Merot P., On constraining the predictions of distributed models: the incorporation of fuzzy estimates of saturated areas into the calibration process, Water Resour. Res. 34 (1998) 787797. [22] Freer J.B., Ambroise B., Beven K.J., Bayesian estimation of uncertainty in runoff prediction and the value of data: An application of the GLUE approach, Water Resour. Res. 32 (1996) 21612173. [23] Gascuel-Odoux C., Hubert-Moy L., Robin P., Walter C., Quelle chelle est pertinente pour dfinir les paramtres de fonctionnement dun modle hydrologique en milieu agricole intensif ? Report INRA Rennes, Sol-Agronomie, 1994. [24] Gascuel-Odoux C., Robin P., Walter C., Molnat J., Quelles rsolutions spatiales et temporelles sont pertinentes pour la modlisation hydrologique des bassins versants dordre 1, in: Cheverry C. (Ed.), Agriculture intensive et qualit des eaux, INRA Editions, 1998, pp. 251266. [25] Klemes V., Conceptualization and scale in hydrology, J. Hydrol. 65 (1983) 123. [26] Klemes V., Dilettantism in hydrology: transition or destiny? Water Resour. Res. 22 (1986) 77S188S. [27] Koutsoyannis D., Optimal decomposition of covariance matrices for multivariate stochastic models in hydrology, Water Resour. Res. 35 (1998) 12191229.

[28] Lecomte V., Transfert de produits phytosanitaires par le ruissellement et lrosion de la parcelle au bassin versant. Processus, dterminisme et modlisation spatiale, Thse de lcole Nationale du Gnie Rural et des Forts, 1999. [29] Makhlouf Z., Michel C., Relation entre les paramtres dun modle global pluie-dbit et les caractristiques physiques dun bassin versant breton, in: Le Barbe L., Servat E. (Eds.), Rgionalisation en hydrologie, application au dveloppement, ORSTOM, Bondy (F), 1995, pp. 531547. [30] de Marsily G., De la validation des modles en sciences de lenvironnement, in: Blasco F. (Ed.), Tendances nouvelles en modlisation pour lenvironnement, Actes de Journes du Programme Environnement, Vie et Socit du CNRS, Elsevier, 1999, pp. 375382. [31] Molnat J., Davy P., Gascuel-Odoux C., Durand P., Study of three subsurface hydrologic systems based on spectral and co-spectral analysis of time series, J. Hydrol. 222 (1999) 152164. [32] Moss M.E., Space, time and the third dimension (model error), Water Resour. Res. 15 (1979) 17971800. [33] Nash J.E., Sutcliffe J.V., River flow forcasting through conceptual models. 1. A discussion of principles, J. Hydrol. 10 (1970) 282290. [34] Neal C., Rosier T.W., Chemical studies of chloride and stable oxygen isotopes in two conifer afforested and moorland sites in the British Uplands, J. Hydrol. 115 (1990) 269283. [35] Normand M., Loumagne C., Cognard C., Ottle C., Taconet O., Vidal-Madjar D., Approche conceptuelle de la modlisation pluie-dbit prenant en compte ltat hydrique du sol. in: Cheverry C. (Ed.), Agriculture intensive et qualit des eaux, INRA Editions, 1998, pp. 169182. [36] Pilgrim D.H., Some problems in transferring hydrological relationships between small and large basins and between regions, J. Hydrol. 65 (1983) 4972. [37] De Roo A.P.J., Hazelhoff L., Heuvelink G.B.M., Estimating the effects of spatial variability of infiltration on the output of a distributed runoff and soil erosion model using Monte Carlo methods, Hydrol. Proc. 6 (1992) 127143. [38] Saulnier G.M., Information pdologique spatialise et traitements topographiques amliors dans la modlisation hydrologique par TOPMODEL, thse de Doctorat de lInstitut National Polytechnique de Grenoble, 1996. [39] Sorooshian S., Parameter estimation, model identification and model validation: case of conceptual models, in: NATO-Asi symposium, Recent advances in modelling hydrological systems, 1988. [40] Wood E.F., Sivapalan M., Beven K.J., Band L., Effects of spatial variability and scale with implications to hydrologic modelling, J. Hydrol. 102 (1988) 2947.

Potrebbero piacerti anche