mike she usa

Journal of Hydrology (2007) 334, 73–87

ava i lab le a t www.sc iencedi rec t . com

journal homepage: www.elsevier .com/ locate / jhydrol

Assessment of the effects of DEM gridding on thepredictions of basin runoff using MIKE SHEand a modelling resolution of 600 m

R.F. Vazquez a,*, J. Feyen b,1

a Centro de Investigacion y Tecnologıa Agroalimentaria de Aragon, Unidad de Suelos y Riegos, Avenida deMontanana 930, 50059 Zaragoza, Spainb Division Soil and Water Management, Department of Land Management and Economics, K.U.Leuven, Celestijnenlaan200 E, 3001 Heverlee, Belgium

Received 7 February 2006; received in revised form 26 September 2006; accepted 1 October 2006

00

do

*

1

KEYWORDSDEM;TOPOGRID;MIKE SHE;Catchment distributedmodelling;Resolution

22-1694/$ - see front matte

i:10.1016/j.jhydrol.2006

Corresponding author. TelE-mail addresses: raulfvazFax: +32 16 329760.

r ª 200.10.001

.: +34 97quezz@y

Summary A 586-km2 catchment was modelled with the distributed hydrologic modelMIKE SHE. Coarse digital elevation models (DEMs) having a 600-m resolution and griddedfrom a set of elevation points geographically distributed with a much finer resolution wereused in the modelling with the purpose of investigating potential effects of the DEM gen-eration methods on (i) model parameter values; (ii) adequacy of model global predictions;and (iii) the evaluation of internal state predictions. To address these aspects, this paperdescribes the DEM gridding methods, assesses the accuracy of the DEMs and examines sys-tematically the sensitivities of parameter values and predictions of the distributed modelwith respect to the DEMs. Three types of gridding methods were applied. Methods type Iwere based on the use of the MIKE SHE interpolation tool (Bilinear algorithm) for process-ing input elevation data distributed about the periphery of the gridded DEM cells. Inputelevation data distributed about the centre of the gridded DEM cells were processed ingridding methods type II. The third type was based on the use of the TOPOGRID algorithmthat considers landscape features, such as digitised streams, to improve the drainagestructure of the gridded DEMs. A multi-criteria protocol was applied for assessing the ele-vation quality of DEMs and their suitability for hydrologic purposes. It was found that thequality of the DEM products of the MIKE SHE interpolation tool were poorer. The indepen-dent calibration of the assembled hydrologic models revealed (i) important variations of

6 Elsevier B.V. All rights reserved.

6 716324; fax: +34 976 716335.ahoo.co.uk, [email protected] (R.F. Vazquez), [email protected] (J. Feyen).

mailto:[email protected]



74 R.F. Vazquez, J. Feyen

model predictions; and (ii) from average to important variations of effective parametervalues, as a function of the different DEMs. A multi-criteria protocol analysing dischargetime series, peak flows and piezometric levels showed that model performance is in broadterms in agreement with the elevation and slope quality of the DEMs.ª 2006 Elsevier B.V. All rights reserved.

Introduction

Digital Elevation Models (DEMs) are important tools inhydrologic research and water resources management owingto the relevance that geo-morphological features intrinsicin the DEMs have for the simulation of important water flowprocesses such as surface runoff, evaporation and infiltra-tion. However, DEMs, as source of spatially distributedground elevations, are not free of errors and limitations.DEM square-grid structures have limitations for handling dis-continuities in elevation and representing adequately all ofthe landscape features. Indeed, either triangulated irregu-lar networks (TIN) or contour lines should be preferred forrepresenting a surface for hydrologic purposes (Wise,2000; Vivoni et al., 2005). However, square-grid DEMs arestill widely used for hydrologic purposes owing mainly totheir simplicity and computational efficiency.

In this context, the referred limitations of grid DEMs forhandling discontinuities in elevations and representingappropriately landscape features are reduced by decreasingas much as possible their grid size (Walker and Willgoose,1999; Wise, 2000). Particularly, in catchment distributedmodelling using grid DEMs, research has enabled to recom-mend the use of DEM grid sizes smaller than 50 m for ade-quate flow pathway analysis at the hillslope scale (Saulnieret al., 1997a; Beven and Freer, 2001). Thus, using the dis-tributed code TOPMODEL (Beven et al., 1995), Zhang andMontgomery (1994) selected a 10-m grid size for the ade-quate simulation of geomorphic and hydrologic processesin two small catchments (0.3-km2 and 1.2-km2). Beldring(2002) used a 10-m DEM for modelling a 6.2-km2 catchment.Braud et al. (1999) modelled a 5.47-km2 mountainous catch-ment with the ANSWER code (Beasley et al., 1980) using a30-m grid size. Guntner et al. (1999) applied TOPMODELon a well-monitored 40-km2 catchment considering a 50-mgrid size.

The use of DEM grid sizes smaller than 50 m is howevernot always possible in catchment distributed modelling.This is in part due to the lack of world-wide data with theappropriate resolution. Other important reason is relatedto computational efficiency, which is sensitive to the num-ber of horizontal and vertical (modelling) computationalunits and, as such, to the size of the modelled catchment.In this respect, Xevi et al. (1997) and Christiaens and Feyen(2002) modelled a 1-km2 well-monitored experimentalcatchment with the code MIKE SHE (Refsgaard and Storm,1995) considering a grid size of 100 m. Refsgaard (1997)and Madsen (2003) considered grid sizes larger or equal to500 m for modelling a 440-km2 catchment. Refsgaard andKnudsen (1996) modelled a 1090-km2 catchment with a1000-m grid size using MIKE SHE and Jain et al. (1992) mod-elled with the same hydrological code the 820 km2 Kolarcatchment in India with grid sizes ranging from 500 to4000 m.

The use of such coarse grid sizes in catchment distrib-uted modelling implies important spatial scale differencesamong the scale to which the physical structure of thehydrologic codes were obtained, the scales to which the dif-ferent data are collected and the coarse scales to which thehydrologic codes are applied (Bergstrom and Graham, 1998;Vazquez et al., 2002; Vazquez, 2003). The following aretherefore important issues that are related to the impactof grid scale on the predictions of catchment modelling:

(i) what is the adequate grid resolution for achievingaccurate model predictions, while keeping computa-tional times under reasonable limits?Prior sensitivity analyses demonstrated that using (moreor less) different data for the same modelling variablelead to significant differences in both effective parame-ter values and model performance (Vazquez et al., 2002;Vazquez and Feyen, 2003b).(ii) Given that geomorphologic features intrinsic in theDEMs (i.e. elevation, slope, curvature, etc.) are impor-tant for the simulation of flow processes such as surfacerunoff, infiltration and evaporation, and provided thatdifferent DEM accuracies are expected from the applica-tion of different DEM gridding methods, do the effectiveparameter values reflect the differences of these DEMgeneration methods when using a coarse modellingresolution?(iii) Is the adequacy of global predictions affected by dif-ferent DEM generation methods? and(iv) Is the evaluation of internal state predictionsaffected by the DEM generation methods?

The assessment of the first of these grid-scale issues willdemand the consideration of various aspects such as param-eter error, model structural error and data (input and eval-uation) measurement error. In this context, Vazquez et al.(2002), after using 300, 600 and 1200-m modelling gridsizes, found that an acceptable compromise between accu-racy of model predictions and computational (i.e. running)time was reached when using a grid size of 600 m for themodelling of the Gete catchment (Belgium) with the MIKESHE model. This study did not consider model structural er-ror owing to the lack of access to the structure of the MIKESHE model (access limitations linked to the commercial nat-ure of the software). However, the main conclusions of thereferred study were based on parameter calibration, theevaluation of internal state predictions and a brief assess-ment of data measurement error concerning piezometricdata (for evaluation).

With regard to the other grid-scale issues, previous stud-ies have used topographically driven codes such as TOPOG(Vertessy et al., 1993) and TOPMODEL for examining the ef-fects of both the scale of the input elevation data and theresolution of the gridded DEMs on model performance

Assessment of the effects of DEM gridding on the predictions of basin runoff using MIKE SHE 75

(Zhang and Montgomery, 1994; Wolock and Price, 1994;Lane et al., 2004) and on the effective values of saturatedhydraulic conductivity (Saulnier et al., 1997a,b). However,published modelling studies using MIKE SHE with coarseDEMs have not addressed explicitly either of these topics(Refsgaard and Storm, 1995; Refsgaard and Knudsen, 1996;Xevi et al., 1997; Refsgaard, 1997; Feyen et al., 2000; Chris-tiaens and Feyen, 2002; Madsen, 2003). Furthermore, dis-cussion about the incidence of different gridding methodson either the quality of the DEM products, the model globalpredictions, the evaluation of internal state predictions orthe effective parameter values of the hydrologic model isnot common in previous publications about distributed mod-elling of catchments.

Therefore, in contrast to previous published work, thisarticle presents an assessment of different methods forgridding coarse-DEMs (up-scale gridding) and the potentialeffects of these methods on the referred grid-scale issues,namely, adequacy of global predictions, effective parame-ter values and the evaluation of internal state predictions.In line with the conclusions of previous work (Vazquezet al., 2002), the current assessment involves the use of a600-m grid size. The assessment is furthermore based onthe application of the MIKE SHE model on the data of theGete catchment (Belgium).

Materials

The study site

The study site, the Gete catchment (586 km2), located tothe east of Brussels-Belgium (Fig. 1), comprises the sub-catchments of the Grote Gete (326 km2) and the Kleine Gete(260 km2). The elevation of the area varies from approxi-mately 27 m in the northern part to 174 m in the southernpart. Land use is mainly agricultural with some local for-ested areas. The local weather is characterised by moderatehumid conditions. Nine soil units can be distinguishedaccording to the legend of the Belgian soil map (Vander

Main flow direction

5 10 15 20 25 30 35 40 45

45

50

40

35

30

25

20

15

10

5

55

Q-Grote Gete

Figure 1 Location of the study site and distribution of the calibratFeyen, 2003a).

Poorten and Deckers, 1994; Vazquez, 2003), e.g. loamy soils(Aba, Ada and Adc), sand–loamy soils (Lca, Lda and Ldc),clay soils (Eep and Uep) and soils with stony mixtures(Gbb). The dominant soil type in the catchment is the Abasoil unit. The reader is referred to Vazquez et al. (2002,2003) and Vazquez and Feyen (2003b) for additional detailsabout the description of the catchment.

The hydrologic code

The MIKE SHE code (Refsgaard and Storm, 1995) was consid-ered for the integral modelling of the study site. MIKE SHE isa well-known deterministic-distributed code that has beenused and described in a wide range of applications (e.g.,Refsgaard, 1997; Jayatilaka et al., 1998; Feyen et al.,2000; Christiaens and Feyen, 2002; Vazquez and Feyen,2003a). MIKE SHE integrates the entire land phase of thehydrologic cycle and can model interception, actual evapo-transpiration (ETact), overland flow, channel flow, flow inthe unsaturated zone, flow in the saturated zone and ex-change between aquifers and rivers. MIKE SHE applied at acatchment scale implies the assumption that smaller scaleequations are valid also at the larger scale; thus, it performsan upscaling operation using effective parameter values.The MIKE SHE model uses a square-grid modelling structure.Consequently, square-grid DEMs were used in this research.

The input elevation data

The DEMs were gridded from a set of elevation spot heights(Zsource), available from the Flemish Spatial Data Infrastruc-ture (OC-GIS Vlaanderen-Belgium). There are no publisheddetails about the methods used to create these elevationspot heights. However, according to staff at the NationalGeographical Institute of Belgium (NGIB), these data werederived originally from digitised contour lines from a1:50000 topographic map and finally arranged in a non-orthogonal but regular grid mesh with an approximate reso-lution of 40 m in the X-direction and 30 m in the Y-direction.

Q-GeteQ-Kleine Gete Catchment

outlet

Multi-Site observation wellMulti-Site stream stationSplit-Sample observation wellSplit-Sample stream station

Coordinates: simulation grid(600x600 m² model) 50 55

Legend:

N

Flemish region

Walloon region

Gete

France

North sea TheNetherlands

catchment

ion and evaluation wells and stream stations (after Vazquez and


Inspection of the numerical characteristics of Zsource re-vealed that the information was recorded to the nearestmeter. This constitutes by itself an important factor ofuncertainty associated with Zsource that should be taken intoaccount when assessing the elevation quality of the griddedDEMs, as owing to this uncertainty, Zsource is not a true rep-resentation of the topography of the study site. Further-more, digitised streams and the topographical catchmentdivide, as derived from the 1:50000 scale maps, wereavailable.

Methods

DEM gridding methods

The selection of a gridding method depends principally onthe spatial distribution of the input data and resolution ofthe output grid. In common hydrologic-model applications,DEMs are normally gridded by means of interpolation algo-rithms for estimating elevations, so that they have a resolu-tion that is finer than or similar to the average resolution ofthe input data (i.e. down-scaling operation). The use ofcoarse grid DEMs for hydrologic purpose demands, on thecontrary, an up-scaling gridding operation to generate DEMswith a resolution that is much coarser than the average res-olution of the input data.

In this work the spot elevation data were gridded accord-ing to five up-scaling methods. The first method ((A)) usesthe spot-based Bilinear (Bi) interpolation algorithm, avail-able as a pre-processing MIKE SHE tool (DHI, 1998). Thisinterpolation tool uses up to a maximum of four points(one per quadrant) to estimate the elevation at every grid-ded cell corner. The points are the nearest to the cell cor-ner and are selected on the basis of (i) their distance to thecorner and (ii) a user-defined searching radius around thecorner. The method then estimates, the centre value forevery cell as the average of the four corner values. There-fore, depending on both the density of the elevation dataand the modelling resolution, up to a maximum of 16 datapoints may contribute to the estimation of the cell centrevalue. Considering the approximate resolution of Zsource(i.e. 40 m · 30 m), the searching radius was given a value

(much) greater than 50 m ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffið40Þ2 þ ð30Þ2

qm, that is, long

enough to use 16 elevation points. For DEM coarse resolu-tions, this up-scaling method uses information from pointslocated around the gridded cell corners, that is, far awayfrom the cell centre where the elevation is finally derived.

The other four approaches, methods (B), (C), (D) and (E),were implemented with the aid of algorithms that are avail-able for gridding DEMs in common geographical informationsystem (GIS) packages, namely, ARC/INFO (ESRI, 1996) andIDRISI (Clark Labs, 1998). These methods are commonlyused for downscaling operations, but they are not necessar-ily the most appropriate to carry out the up-scaling griddingof coarse DEMs. Therefore a systematic evaluation of theirDEM-products is needed. For methods (B) and (C) a TIN sur-face was produced with ARC/INFO on the basis of a linearinterpolation process. In method (B), the TIN surface wasfurther interpolated into a regular lattice (TIN to latticetransformation) with a 600-m resolution. In general, thistransformation involved a linear interpolation along the

edges of the TIN triangles (positioned around the cell cen-tre) to determine the elevation of the lattice pixels. Thus,unlike method (A), method (B) uses input elevation dataconcentrated around the cell centre where the elevationis estimated. In method (C), contour lines were derived withARC/INFO from the TIN surface. These contours were thenutilised as elevation input for the MIKE SHE Bi interpolationutility. Thus, alike method (A), method (C) uses informationfrom elevation spot heights located at the periphery of thegridded cell.

For methods (D) and (E), the ARC/INFO module TOPO-GRID (ESRI, 1996) was used. TOPOGRID is a finite differenceinterpolation technique that is based on the specialisedinterpolation approach ANUDEM (Hutchinson, 1989) that ischaracterised by its computational efficiency, so that itcan handle large data sets, and by a drainage enforcementalgorithm that uses landscape features, such as digitisedstreams, to improve the structure of DEM products forhydrologic purposes. TOPOGRID imposes interpolation con-straints to remove spurious sinks that result in a more cor-rect drainage structure and representation of ridges andstreams. The use of TOPOGRID is therefore considered asone of the best current practices in DEM gridding, as it gen-erates a hydrologically consistent DEM (ESRI, 1996; Hutchin-son, 1989). The readers are referred to Hutchinson (1989),Hutchinson and Dowling (1991) and ESRI (1996) for a com-plete description of ANUDEM and TOPOGRID.

In method (D) a 20-m DEM was initially obtained througha common (i.e. downscaling) TOPOGRID application. Re-sampling methods were then applied on the 20-m DEM forobtaining coarse 600-m DEMs, namely, the Nearest Neigh-bour (Nn), the Bilinear (Bi) and the Cubic Convolution (Cu)methods (ESRI, 1996; Hanselman and Littlefield, 1998).The Nn algorithm assigns the value associated with the clos-est cell centre on the input grid to the re-sampled cell. TheBi interpolation algorithm uses a weighted average deter-mined by the values of the input cells at the four nearestcell-centres and their weighted distance to the centre ofthe re-sampled cell. Cu is a surface-fitting algorithm thatuses the 16 nearest input cell centres and their values todetermine the re-sampled cell value. Depending on the res-olution of the input grid and similarly to method (B), thesere-sampling methods use elevation data concentratedaround the coarse cell centre. The three DEM products ofthe re-sampling methods were compared with each otherfor selecting a DEM product representative of method (D)to be used in the hydrologic modelling. This assessmentwas based on the analysis of the cumulative distributionsof DEM elevations, slopes, etc, as explained later in thetext. The assessment of DEM elevation quality indicated amarginal difference in detriment of the product of the Nnre-sampling technique (DEM(D_Nn)) and in favour of theproduct of the Cu re-sampling technique (DEM(D_Cu)).Therefore, DEM(D_Cu) was selected as representative ofmethod (D) for the hydrologic modelling. This selectionwas also supported by the fact that more input data (16 in-put cell centres) are included in the Cu method than in theother re-sampling methods for determining every re-sam-pled cell centre.

For method (E), the coarse 600-m DEM was obtained di-rectly from Zsource without the need of a re-sampling meth-od through the application of TOPOGRID, within an up-


scaling context. In contrast to the other methods, TOPO-GRID does not use input elevation data exclusively concen-trated either about the centre or the periphery of thegridded cell, but instead, it uses simultaneously distributedelevation data and stream information to enforce a morenatural drainage structure in the DEM products.

Thus, depending on whether the gridding method uses in-put elevation data distributed about the gridded cell centreor about the gridded cell periphery, the gridding methodsare further classified in this work as belonging to class typeI (i.e. methods (A) and (C)), class type II (i.e. methods (B)and (D)) and class type III (method (E)). In the forthcomingsections of this paper, the quotation ‘‘DEM(method/type)’’stands for the DEM product of a particular gridding methodor type.

DEM quality assessment

The assessment of the quality for hydrologic use of thecoarse DEM products was based on (i) the comparison ofthe DEMs and Zsource, (ii) the comparison of the DEMs witheach other through their hydro-geomorphic properties,and (iii) the analysis of plots such as spatial distribution ofpits, drainage patterns, derived contours, etc. The set ofhydro-geomorphic properties comprised: drainage patterns,catchment areas, slopes and hillslope shading and the com-putation of the topographic index (k). These analyses weredone considering the domain defined by the digitised catch-ment boundary.

The discrepancies between the coarse DEM elevations(ZDEM) and the input elevation data (Zsource) were condensedinto summary statistics, such as minimum, mean and maxi-mum values, standard deviation and the root mean squarederror (RMSE)

RMSE ¼

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiPni¼1ðresÞ

2i

n

sð1Þ

where (res)i, is the residual, the arithmetic difference be-tween a coarse DEM elevation (ZDEM) and Zsource at the i-thtest location of interest; and n is the number of test loca-tions. The RMSE analysis included 150 test-spots that weredefined pseudo-randomly, that is, 120 spots were selectedrandomly whilst the remaining 30 spots were chosen subjec-tively to include terrain zones of hydrologic and geomorpho-logic interest, such as depressions, spots near water coursesand uplands. The comparison of the coarse DEMs and Zsourceincluded also the analysis of the cumulative distributions ofresiduals computed at the same test locations as in theRMSE evaluations. Furthermore a cumulative distributionof DEM elevations was calculated for every DEM product.

The derivation of drainage patterns and the automaticdelineation of catchment (i.e. drainage) areas are affectedby spurious sinks (i.e. local minima), as sinks interfere withthe (natural) connection of drainage patterns. Thus, sink re-moval was carried out prior to the assessment of these geo-morphic relationships. Because zones of natural depressionstorage are very difficult to account while using coarse res-olutions, such as the one considered in this study, all theidentified sinks were treated as artificial depressions result-ing from random errors induced by the gridding interpola-tion. Thus, all of the identified sinks were removed. In

doing so, a method based on the drainage enforcementalgorithm by Jenson and Domingue (1988) was applied. Withthis algorithm, the downslope flow direction for each cell isdefined after inspecting the eight cardinal directions andidentifying the neighbouring cell of lowest elevation. Thealgorithm fills the sinks by assigning to them the elevationof their lowest neighbour and then assigns flow directionsto these flat cells towards a neighbouring cell that has a pre-viously assigned flow direction.

The Jenson and Domingue (1988) algorithm was also usedto derive the drainage patterns, accounting for the flow thataccumulates at each cell of the DEM for a uniform and unit-value rainfall. A threshold value (10 flow units) was then ap-plied to produce the drainage networks through a reclassifi-cation process that accounted for cells with higheraccumulated flow. The catchment areas were automaticallydelineated also through the Jenson and Domingue (1988)algorithm by using the information on the flow directionfor each cell and the grouping of cells draining to a singleoutlet or seed cell. This seed pixel was defined explicitlymatching the coordinates of the Gete station (Fig. 1). Thederived catchment areas were compared to the digitisedcatchment boundary.

Although some of the assumptions on which the interpre-tation of the topographic index (k) is based are not fulfilled(mainly because of the coarse resolution used in thisresearch), it was calculated to briefly assess the potentialrunoff sources (Beven et al., 1995; Quinn et al., 1995;Ambroise et al., 1996). For a particular grid cell, it is calcu-lated as

ki ¼ lnai

tanðbiÞ

� �ð2Þ

where ai is the area draining through the cell per unit lengthof contour [L]; and tan(bi) is the local surface slope [–] ofthe cell. The multiple direction flow sharing algorithm byQuinn et al. (1991) was used to determine the downslopeflow pathways and for distributing to the downslope cellsa proportion of the accumulated contributing area. Thisalgorithm is more suitable than that of Jenson and Domin-gue (1988) for representing flow on divergent hillslopesand for larger grid resolutions (Quinn et al., 1991, 1995).The calculation of k enabled to compare the drainagepatterns derived with both methods, namely Quinn et al.(1991) and Jenson and Domingue (1988).

Sensitivity analyses on the coarse DEMs

The grid-scale issues of interest defined in the objectives ofthis manuscript were examined through a Multi-Calibration(MCal) test, in which several models differing only in theDEM input data were subjected to identical calibrationand evaluation. After model calibration, the differences inthe sets of effective parameter values were assessed.During model evaluation, the differences in model perfor-mance were characterised through the analysis of the modelresiduals (i.e. differences between predictions andobservations).

In addition, an analysis of the effects of the sink removaloperation was performed considering the DEM(A) represent-ing gridding type I and DEM(B) representing gridding type II.


Additionally, this analysis was extended to include DEM(C)for investigating the modelling consequences of the mis-match that was observed between the derived catchmentoutlet and the location of the Gete station, which is de-picted later in the text.

Hydrologic model of the study site

The profile definition of the river tributaries was based oninterpolation/extrapolation of a few measured profiles.Drains were specified in the model set-up to improve thesimulated hydrograph shape and to account for the small ca-nals and ditches present on a scale smaller than the model-ling resolution. The drainage depth (zdr) and the reciprocaltime constant (Tdr), a sort of drainage coefficient, were cal-ibrated because they influence mainly the velocity of thedrainage and the peak and recession of the hydrograph.The spatial extent of the soil units and their vertical proper-ties could be assessed considering two soil databases (Feyenet al., 2000). Parameters for describing the flow through thesoil system were calculated with pedo-transfer functions(PTFs). Despite the uncertainties associated with the PTFsand the soil databases, the soil parameters were not in-cluded within the calibration process to avoid consideringtoo many parameters (for 9 soil units) during modelcalibration.

The complex geology of the studied system comprisesnine geological units (Vazquez and Feyen, 2003a), of whichonly two are underlying completely the area of the catch-ment. The geology was incorporated in the three-dimen-sional groundwater model of MIKE SHE. A sensitivityanalysis demonstrated that the model could be simplifiedfurther to six geological units without influencing the globalresults noticeably (Vazquez et al., 2002). The model in-cludes five upper geological units on top of the low-perme-able Palaeozoic rocky basement. The hydrogeologicparameters of these units were tuned during the calibrationprocess. Owing to the lack of appropriate measurements,the groundwater divide was assumed coincident with thetopographical divide and the aquifers were given no-flowboundary conditions.

The input time step, during which no change in boundaryconditions occurs (stress period), was taken as one day, inrecognition of the lack of more precise meteorological data.MIKE SHE requires crop potential evapotranspiration (ETp)data for modelling ETact. Modelling ETact was done by meansof the Kristensen and Jensen (1975) approach. ETp datawere estimated in turn by means of the Kc–ET0 method thatuses crop coefficients (Kc) and crop reference evapotranspi-ration (ET0). ET0 time series were estimated with the Foodand Agriculture Organisation (FAO) Penman-method 24(FAO-24). Locally-applicable parameter values were avail-able for the components of the FAO-24 method (Vazquezand Feyen, 2004). The estimates produced by the FAO-24method, using these locally-applicable parameter valuesfor the various components of the method, were equivalentto the estimates of the FAO-56 Penman-Monteith methodusing standard parameter values recommended by theFAO-56 report for the different components of the method(Allen et al., 1998). No locally-applicable parameter valueswere available for the components of the FAO-56 method(Vazquez and Feyen, 2004). The effective values for the

parameters that MIKE SHE uses to estimate ETact (Kristensenand Jensen, 1975) were taken from a previous sensitivityanalysis (DHI, 1998; Vazquez and Feyen, 2003b). Theseparameter values were kept constant throughout the mod-elling analysis. The reader is referred to Doorenbos and Pru-itt (1977), Vazquez and Feyen (2003b, 2004) for a completedescription of the Penman FAO-24 approach and the set-upof the catchment model.

Model calibration and performance assessment

In principle, when sufficient data are available, physicallybased distributed models do not need to be calibrated.However, this type of models must be calibrated to improvetheir predictions because of sub-grid variability of parame-ter values, model structure uncertainties and data (inputand evaluation) uncertainties. A good calibration processshould aim therefore to reduce as much as possible themodel error, that is, parameter uncertainties (i.e. obtainingappropriate grid-scale effective parameter values) andmodel structure uncertainties (i.e. developing the mostaccurate catchment model) so that the total modellingerror is mainly composed by the data measurement error(which is usually unknown and scale dependant). With thispurpose in mind, and in line with their nature, distributedmodels must be evaluated against distributedmeasurements.

In this study, care was taken to avoid violating physicalconstraints during the model calibration (Vazquez andFeyen, 2003a). Particularly, the availability of piezometricdata defined the calibration period as from the 1st of Janu-ary 1985 until the 31st of December 1986 and the main eval-uation period from the 1st of January 1987 until the 31st ofDecember 1988 (Split-Sample test). Additional evaluationperiods of variable length were however considered in thescope of a Multi-Window (MW) test (Vazquez, 2003; Vazquezand Feyen, 2003b), as depicted later in the text.

Owing mainly to the high computational requirementsand the large number of model parameters, the calibrationof distributed hydrologic models is not a trivial activity. Inthis research, a considerable computational work was re-quired since every simulation lasted about one hour andan average of 400 simulations were needed to reach everymodel calibration in the framework of the Multi-Calibration(MCal) test. A conventional calibration by fit process was ap-plied. It was based on a ‘‘trial and error’’ procedure, inwhich the influence of the various model parameters wasexamined step by step through a multi-criteria performanceprotocol. In general, for each DEM product the model wascalibrated and evaluated using a Split-Sample (SS) proce-dure against basin-wide daily discharge measurements andwater levels for 12 observation wells, with screens in differ-ent geological layers (Fig. 1). To investigate how well thecalibrated model was able to simulate internal variables, aMulti-Site (MS) evaluation test was also performed for twointernal discharge stations and 6 observation wells thatwere not considered during the calibration process (Fig. 1).

The general protocol used for assessing the model per-formance consisted of two complementary components:(i) the analysis (statistical and graphical) of different timeseries properties; and (ii) the evaluation of a set of multi-objective statistics. The inspected time series simulation


properties included: (i) cumulative flow volumes; (ii) mod-elled versus observed discharge maxima; and (iii) high flowextreme value statistics.

Unrelated or slightly correlated statistics were preferredfor the set of multi-objective statistics. However, as anexception, two statistics that are correlated were consid-ered in this research for getting a quick overview of themodel performance simulating peak flows. These statisticsare the coefficient of efficiency (EF2) (Legates and McCabe,1999), that is frequently used for an estimation of the total(combined systematic and random) average error and thecoefficient of determination (CD) (Loague and Green,1991) that is related to the EF2 but is particularly usefulto assess the simulation of peak values (Vazquez, 2003).These statistics are defined as follows:

EF2 ¼ 1�Pn

i¼1ðOi � PiÞ2Pni¼1ðOi � OÞ2

ð3Þ

CD ¼Pn

i¼1ðOi � OÞ2Pni¼1ðPi � OÞ2

ð4Þ

where, Pi is the i-th simulated value, Oi is the i-th observedvalue, O is the average of the observed values and n is thenumber of observations. The optimal value of EF2 is 1.0and the feasible range of variation is �1 < EF2 6 1.0, whilefor CD, these parameters are 1.0 and 0.0 < CD < +1, respec-tively. For a more even assessment of the simulation of bothhigh and low flows, the multi-objective set of statistics wasalso applied on the logarithmic transformation of the ob-served and the predicted variables (Vazquez, 2003).

Primarily, the drainage depth (zdr) and the reciprocaltime constant (Tdr) were calibrated against the overall dis-charge of the catchment. Next, the values of the hydrogeo-logic parameters of the five upper most geological unitswere tuned against the outflow discharge of the catchment.When the overall discharge was reasonably well simulated,the hydrogeologic parameters of each unit were tuned fur-ther to improve the agreement between the predicted andobserved piezometric levels in the calibration wells. Previ-ous modelling experiences (Feyen et al., 2000; Vazquezet al., 2002; Vazquez and Feyen, 2003b) showed that aftertuning the hydrogeologic parameters for improving the pie-zometric predictions, the model performance simulatingoverall discharges was diminished and, as a consequence,an additional tuning of zdr, Tdr and the horizontal conductiv-

0.0

0.2

0.4

0.6

0.8

1.0

0tan( ) [--]

]--[ aera tne

mh ctac f

o n

oitcarF

20-m(A)(B)(C)(D)(E)

DEMs after sink removal

1 2 3 4 5

Figure 2 Cumulative frequency distributions of (a) DEM local slopspot heights (_rand).

ity of the loamy Quaternarian (Kw) and the clayey sandLandeniaan (Ln) units was necessary for improving the sim-ulation of the overall discharge. Thus, in this research a fur-ther tuning of these parameters was carried out to improveas much as possible the prediction of the overall dischargebut trying at the same time to reach a modelling compro-mise to avoid affecting significantly the prediction of piezo-metric levels.

For the extreme value analysis (EVA), independent Peak-Over-Threshold (POT) values were extracted from the totaldischarge (Qt) using independency criteria based on the dif-ferences among the recession constants of the hydrologicsubflows: overland flow (Qov), interflow (Qin) and baseflow(Qbs) (Willems, 2000; Vazquez and Feyen, 2003a). The EVAwas performed from the 1st of January 1984 until the 31stof December 1995. An exponential distribution fitted rea-sonably well the observed data beyond an optimal thresholdQt equal to 6.4 m3 s�1 (Vazquez and Feyen, 2003a).

Results

DEM quality assessment

The analysis of the cumulative distributions of DEM eleva-tions and the elevation statistics did not reveal a clear dif-ference among all of the inspected DEMs. Fig. 2a shows thecumulative distributions of slope (tan(b)) as a fraction ofthe catchment area and as a function of the DEMs after sinkremoval. The plot also includes the corresponding distribu-tion for the 20-m product of the TOPOGRID algorithm,considered in this research as hydrologically consistent(Hutchinson, 1989; ESRI, 1996) and, as such, as a reference.Fig. 2b illustrates the cumulative distributions (Fr(res 6RES)) of residuals for a particular value of interest (RES),calculated at the 150 pseudo-random test locations as afunction of the DEMs after sink removal.

Fig. 2a shows the considerable smoothing of the DEMsthat took place as a consequence of increasing the resolu-tion of the elevation data (from about 20 m to 600 m) andas such indicates the generalised low elevation quality ofall of the coarse DEMs used in this research. Nevertheless,Fig. 2a and b show that the products of methods (B) and(D) described the catchment’s elevation slightly better thanthe DEM products of methods (A), (C) and (E). Additionally,

0.0

0.2

0.4

0.6

0.8

1.0

0RES [m]

<ser(rF

]--[ )S

ER (A)_rand

(B)_rand(C)_rand(D)_rand(E)_rand


4 8 12 16

e; and (b) residuals among DEMs and pseudo-random elevation


these figures clearly show three different sets of cumulativedistributions. One of the sets groups the DEM products ofmethods (A) and (C). A second set groups the DEM productsof methods (B) and (D). Finally, the DEM product of method(E) constitutes a third set. This is consistent with the threedifferent types (I, II and III) of DEM gridding methods thatwere used in this work and indicates that these DEM griddingmethods affected the adequacy of DEM elevations.

The outcomes of methods (B) and (D) (gridding type II) in-cluded greater numbers of sinks than the outcomes of meth-ods (A) and (C) (gridding type I). No sink was identified in theoutcome of method (E), because sinks were automaticallyremoved by TOPOGRID (ESRI, 1996). Visual inspection ofthe spatial distribution of sinks indicated that the coarseDEM gridding methods (in particular methods (B) and (D))did not rightly model the transition from the hillslopes tothe floodplains in the regions where depressions occurred.

Given the limitations for describing appropriately thetopography of the study site associated with the 600-m res-olution, it may be concluded that the drainage networks de-rived through the Jenson and Domingue (1988) approachagreed acceptably well with the digitised watercourses.However, some considerable differences of detail, such asartificial branch discontinuities were noticed (Fig. 3a).Fig. 3b shows that, particularly for DEM(C), the outlet ofthe derived drainage network does not match the digitisedoutlet (i.e. Gete station). This is likely to affect the perfor-mance of the respective hydrologic model because the sim-ulated discharges are evaluated at the location of the Getestation rather than at the derived catchment outlet.

Furthermore, owing to this mismatch, the derived catch-ment area for DEM(C) was much smaller than the digitisedboundary despite the sink removal pre-processing. As away of mitigating this mismatch and ensuring a drainagearea covering the extent of the digitised boundary, an addi-tional smoothing of the coarse DEM(C) was performed with a3 · 3 mean filter (Clark Labs, 1998). As a consequence, thedrainage area for DEM(C) was significantly improved, sincethe location of the predicted catchment outlet was shiftedto the position of the Gete station but to the cost of losinginformation after smoothing (Fig. 3c). Thus, the effects ofthe additional smoothing of DEM(C) on the hydrologic pre-

Legend:

DEM(A) after sink removal

Significant difference with respect to the digitised streams

N

Derived catchment

DEM(C) after sink r

Digitised stream networkDEM-derived stream network

outlet

Figure 3 Maps showing (a) the spatial distribution of derived dralocation of the predicted catchment outlet and the position of thebefore and after (C_mf3) smoothing through the mean filter (3 · 3

dictions were investigated by means of a sensitivity analysisconsidering the smoothed DEM into the hydrologic modelset-up. The model was parameterised using the effectiveparameters that were obtained with the DEM prior thesmoothing process.

The analysis of the spatial distributions of k revealed anacceptable agreement between the higher k values and thedigitised river network. In this respect, the k distributions(Quinn et al., 1995) were consistent with the drainage pat-terns derived through the Jenson and Domingue (1988) ap-proach, even with the situation depicted in Fig. 3b forDEM(C). Furthermore, this analysis suggested a smoothergeneration of runoff for the DEM products of methods (A),(C) and (E) with respect to the outcomes of methods (B)and (D), which is in agreement with their slope characteris-tics (cf. Fig. 2a). The cumulative distributions of k evolvedas a function of the three different types (I, II and III) ofgridding methods inspected in this work. This study in-spected the cumulative distributions of the two componentsof k, namely ln(1/tanb) accounting for land–surface slopeand ln(a) accounting for land–surface shape. The distribu-tions of ln(1/tanb) were consistent with the slope distribu-tions shown in Fig. 2a. The distributions of ln(a) indicatedno special concentrations of either convex or concave land-scape features in the DEMs.

Hydrologic modelling

The results of the hydrologic modelling are presented withregard to the main grid-scale issues that are addressed inthis article.

Do the sets of effective parameter values reflect the dif-ferences of the DEM generation methods?

Table 1 lists the main effective parameter values in rela-tion to the three types of DEM gridding methods. Further-more, the parameters are classified in two main groupswith respect to the presence of artificial sinks in the DEMs.The table lists only the effective values of the loamy Qua-ternarian and the clayey sand Landeniaan layers, whichhave a considerable influence on the modelling of thegroundwater flow, as well as the aquifer–river interaction.During the model calibration, the hydrogeologic parameters

0.0

0.2

0.4

0.6

0.8

1.0

0.0 0.5 1.0 1.5 2.0 2.5tan( ) [--]

Fra

ctio

n o

f ca

tch

men

t ar

ea [

--]

(C)(C)_mf3

DEM(C) after sink removal

Gete station

emoval

Catchment boundary

inage network for DEM(A); and (b) the mismatch between theGete station. (c) Cumulative frequency distributions of DEM(C)pixels smoothing mask).

Table 1 Main effective parameters in relation to the DEM generation methods

DEM properties Model parameter Geological unit Coarse-DEM gridding method

Type I Type II Type III

A C B D E

After sink removal (DEMsNOsink) zdr (m) �0.46 �0.30 �0.43 �0.43 �0.50Tdr (s

�1) 6.20 · 10�8 1.55 · 10�7 6.80 · 10�8 8.50 · 10�8 5.90 · 10�8

Ksat (m s�1) Kx (m s�1) Quaternarian 7.70 · 10�7 8.00 · 10�6 2.00 · 10�7 2.30 · 10�7 2.00 · 10�7

Landeniaan 6.00 · 10�6 5.50 · 10�6 5.50 · 10�6 6.00 · 10�6 6.75 · 10�6

Kz (m s�1) Quaternarian 4.24 · 10�7 4.80 · 10�6 9.00 · 10�8 5.75 · 10�8 9.40 · 10�8

Landeniaan 1.50 · 10�6 1.54 · 10�6 3.58 · 10�6 3.90 · 10�6 4.39 · 10�6

Sy (–) Quaternarian 0.17 0.21 0.20 0.22 0.23Landeniaan 0.41 0.41 0.39 0.43 0.41

Including artificial sinks (DEMssink) zdr (m) �0.40 �0.20 �0.22Tdr (s

�1) 7.00 · 10�8 1.65 · 10�7 1.30 · 10�7

Ksat (m s�1) Kx (m s�1) Quaternarian 2.00 · 10�6 4.00 · 10�6 1.00 · 10�7

Landeniaan 7.00 · 10�6 9.00 · 10�6 9.30 · 10�6

Kz (m s�1) Quaternarian 1.74 · 10�6 2.40 · 10�6 1.00 · 10�7

Landeniaan 1.19 · 10�6 1.80 · 10�6 3.07 · 10�6

Sy (–) Quaternarian 0.20 0.20 0.20Landeniaan 0.19 0.34 0.30

zdr = Drainage level, Tdr = Reciprocal time constant, Ksat = Saturated hydraulic conductivity, Kx = Horizontal hydraulic conductivity, Kz = Vertical hydraulic conductivity, Sy = Specific yield.

Asse

ssmentoftheeffe

ctsofDEM

griddingonthepredictio

nsofbasin

runoff

usin

gMIKESH

E81

0102030405060708090100

Jan-85 May-85 Sep-85 Jan-86 May-86 Sep-86 Jan-87

yad

mm[ l la f

niaR

1-]

1

5

9

13

17

21

m[ egra

hcsiD

3s

1-] Rainfall

DEM(A)Observed

0102030405060708090100

Jan-85 May-85 Sep-85 Jan-86 May-86 Sep-86 Jan-87

yad

mm[ l laf

niaR

1-]

13579

111315171921

m[ egra

hcsiD

3s

1-]

RainfallDEM(B)Observed

0102030405060708090100

Jan-85 May-85 Sep-85 Jan-86 May-86 Sep-86 Jan-87Date

yad

mm[ llaf

niaR

1-]

13579

111315171921

m[ egra

hcsiD

3s

1 -]

RainfallDEM(E)Observed

Figure 4 Observed and calibrated hydrographs for the modelsrelated to DEM(A), DEM(B) and DEM(E), after sink removal.


were allowed to vary spatially considering factors such asthe extent of the main sub-catchments, the extent of thegeologic units and the location of the abstraction wells(Vazquez and Feyen, 2003a). For a particular modelled geo-logic unit, the values listed in Table 1 correspond to theeffective zone with the lowest hydrogeologic parametervalue.

In broad terms, the assessment of the sets of effectiveparameter values obtained using the DEMs after sink re-moval revealed that only an average variation of parametervalues took place as a function of the three types of DEMgridding methods. The set associated with DEM(C) has how-ever noticeably different values for zdr (lower absolute va-lue), Tdr (higher value) and the saturated hydraulicconductivity (Ksat) of the Quaternarian (Kw) layer (higher va-lue), parameters that are related to the simulation of sur-face processes within the modelled hydrologic system.The reason for obtaining this different parameter set islikely to be linked to the mismatch between the predictedcatchment outlet and the digitised catchment outlet (Getestation, cf. Fig. 3b) where the model prediction was finallyevaluated.

When the DEMs including sinks were considered in theMCal analysis, it was observed however a noticeable influ-ence of the DEM gridding methods on the values adoptedby the effective parameters of the hydrologic models.These effects were especially important with respect tozdr, Tdr and Ksat–Kw (c.f. Table 1).

Looking at the effect of the sink removal operation (i.e.smoothing operation) on the sets of effective parametersvalues, Table 1 illustrates that, in general, higher zdr abso-lute values and lower Tdr values (i.e. higher drainage veloc-ities) are associated with the (smoother) DEMs after sinkremoval. This is for opposing to the smoothing (i.e. flatter-ing) of the DEMs by routing higher overland and interflowwater volumes more quickly (higher zdr absolute values)but, at the same time, controlling the magnitude of peaksthrough lower Tdr values. In some cases the accelerationof flow routing is accentuated by higher (horizontal and ver-tical) values of Ksat–Kw, such as for the products of the DEM-methods (A) and (B). For the Landeniaan (Ln) layer, which isthe most influential to groundwater flow and the aquifer–river interchange flow, the MCal analysis indicated thatcomparable values of Ksat were obtained for all the modelsindependently of whether the DEM products included sinks.Thus the effects of the sink removal smoothing were re-flected principally on the variation of zdr, Tdr and Ksat–Kw.

Is the adequacy of global predictions affected by differ-ent DEM generation methods?

Fig. 4 shows the observed and calibrated hydrographs forthe DEM products of the gridding approaches type I(DEM(A)), type II (DEM(B)) and type III (DEM(E)). The figureshows that, in general, the models have certain difficultiesfor rightly simulating the recession limbs and the subse-quent baseflow, especially in the periods January of 1985,February–March of 1986 and June–October of 1986. In gen-eral, the models tended to overestimate the peakflowevents. However, in broad terms, the analysis of the cali-brated time series of total discharge revealed that the mod-els related to the DEM products of gridding methods type II((B) and (D)) and type III ((E)) predicted better dischargeseries than the other models, regardless of the presence

of artificial sinks. These results are in agreement with thebetter quality, in terms of elevation, slope and land-sur-face, of the DEM products of gridding methods types II andIII (cf. Fig. 2a).

The Multi-Window (MW) analysis included several evalua-tion periods of different length. The MW test indicated thatthe discharge performance of the models related to the DEMproducts of gridding methods type II ((B) and (D)) were themost acceptable in the different periods of analysis, regard-less of the presence of sinks. Fig. 5 depicts yearly EF2 valuesfor the period (1984–1995) as a function of both DEMsincluding sinks and DEMs after sink removal. Sink removalwas carried out by smoothing the DEMs. This caused a mod-ification of the original structure of the smoothed DEMs,particularly perceived when calculating the distribution ofslopes (smoother) and the drainage network topology. Con-sequently, the simulation of surface water flow dynamicswas modified with a faster water routing throughout theflatter (i.e. smoother) DEMs, characterised by a generaldeterioration of the discharge performance (especially formethod (A)), as depicted in Fig. 5b. Nevertheless, forDEM(C) the newer river network topology brought the gen-eral enhancement of the drainage network, as comparedto the digitised drainage network, which had a positive ef-fect on the performance of the model related to DEM(C).

In broad terms, the peak flows were reasonably wellsimulated within the calibration period independently ofthe DEM gridding method. However, in the main evaluation

DEMs including sinks

61014

18222630

3438

0.1 1.0 10.0 100.0Return period [years]

m[ egra

hcsiD

3s

1-]

Exp. Distr.Observed(A)(B)(C)


61014

18222630

3438

0.1 1.0 10.0 100.0Return period [years]

s ³m[ e

grahcs i

D1 -]

Exp. Distr.Observed(A)(B)(C)(D)(E)

Figure 6 EVA for the Gete station for the period (1984–1995) in relation to the DEM gridding methods and the removal of artificialsinks.


0.0

0.2

0.4

0.6

0.8

1.0

1984 1986 1988 1990 1992 1994Evaluation window

FE

2]--[

DEM(A) DEM(B) DEM(C)

DEMs including sinks

0.0

0.2

0.4

0.6

0.8

1.0

1984 1986 1988 1990 1992 1994Evaluation window

FE

2]--[

Figure 5 Yearly MW model performances (streamflow) for the Gete station in relation to the DEM gridding methods and theremoval of artificial sinks.


period (1987–1988) the adequacy of the simulated peakflows was inferior due to overestimation. The results ofthe Extreme Value Analysis (EVA) in the period (1984–1995) are illustrated in Fig. 6a for the DEMs including artifi-cial sinks and Fig. 6b for the DEMs after sink removal. Be-sides a generalised overestimation, Fig. 6a and b showthat removing the sinks by smoothing the DEMs producedhigher peaks, except for the product of method (C). Again,higher peaks after sink removal were caused mainly by low-er effective drainage levels (zdr, higher absolute values)that accelerate the routing of water throughout the flatterDEMs and evacuate higher overland and interflow volumesfrom the catchment. Despite lower Tdr effective valueswere obtained to mitigate the rising of the peaks, the pre-dicted peaks were higher after sink removal. Fig. 6b showsthat the highest overestimation of peaks was related tothe DEM product of the MIKE SHE Bilinear interpolation algo-rithm (method (A)). The peakflow predictions related to thegridding methods (C), (B) and (D) were comparable and bet-ter than the predictions related to the outcome of theTOPOGRID algorithm (method (E)).

Concerning the DEM(C), the analysis revealed that theimprovement of the discharge performance after the addi-tional DEM smoothing by the 3 · 3 mean filter is particularlynoticeable in the simulation of peakflows that are lowerafter the additional DEM smoothing in the period (1984–1995). This generalised improvement of the discharge per-formance is consistent with the improved location of thepredicted catchment outlet (after the additional mean filtersmoothing) with regard to the location of the Gete station.This illustrates the significant enhancement of the model

performance associated with the improvement of the DEMdrainage network topology despite the deterioration ofother DEM features such as the distribution of slopes (cf.Fig. 3c). This illustrates as well the necessity of assessingthe consequences of GIS operations such as DEM smoothingon both the structure of the DEMs and the associated modelperformance before accepting the model predictions asbeing valid.

Is the evaluation of internal state predictions affectedby the DEM generation methods?

Since distributed models should be evaluated against dis-tributed measurements by considering the predictions ofinternal state variables, this section illustrates the main dis-tributed results from the hydrologic modelling.

The Multi-Site (MS) analysis of the river discharge predic-tions for the two internal river stations that were not in-cluded in the calibration process suggested that all of themodels have marked difficulties to predict the distributeddischarge variables with reasonable accuracy, suggestingthat some processes such as flow through saturated andunsaturated zones may not be rightly modelled to this scale.Besides the coarse modelling resolution, the noticeableuncertainty attached to the input data that were used forconstructing the hydrologic model contributes probably ina greater proportion to these low model efficiencies. Inany case, the discharge predictions for these stations (cf.Fig. 1), related to the DEM products of the gridding methodstype II ((B) and (D)) and type III ((E)) were better.

Figs. 7 and 8 depict the piezometric level performance ofthe models using DEMs after sink removal for three wellsconsidered in the Split-Sample (SS) test and two wells used

V2TI-KU.PP2 (Landeniaan)

43

45

47

49

51

53

Jan May Sep Jan May Sep Jan May Sep Jan May Sep Jan

]m[

daeH

1985 1986 1987 1988

4038187 (Brusseliaan)

128

130

132

134

136

138


]m[

daeH

1985 1986 1987 1988

4048204 (Landeniaan)

104106108110112114116


]m[

daeH

Observed (A) (B) (C) (D) (E)

1985 1986 1987 1988

Figure 7 Predicted piezometric levels after model calibrationfor some of the wells considered in the Split-sample (SS) test as afunction of the DEM gridding methods (after sink removal).


in the MS test, respectively. These figures show that, in gen-eral, the prediction of the piezometric levels differed con-siderably among the wells and that in some cases therewas an important variation of the performance in relation

V2HG-BR.B5 (Landeniaan)

52

54

56

58

60

62


]m[

daeH

1985 1986 1987 1988

4038488 (Brusseliaan)

118

120

122

124

126

128


]m[

daeH

Observed (A) (B) (C) (D) (E)

1985 1987 19881986

Figure 8 Predicted piezometric levels after model calibra-tion for some of the wells considered in the Multi-site (MS) testas a function of the DEM gridding methods (after sink removal).

to the DEM gridding methods. Both tests revealed that, ingeneral, the piezometric performances associated withthe DEM products of gridding methods type II ((B) and (D))were poorer (cf. Figs. 7 and 8). This suggests that the mod-els related to these DEMs were characterised by lower base-flow and higher overland and interflow predictions, whichfinally resulted in better global model performances.

These results illustrate the importance of carrying out anevaluation of distributed models using distributed measure-ments for the evaluation of simulated internal state vari-ables. The variability of these results illustrates howeverthe inherent difficulties in doing so, including incommensu-rability issues due to the fact that data errors (input and/orevaluation) are usually unknown and scale dependent. Itshould be noticed that Figs. 7 and 8 provide evidence forrejection of all models, unless consideration is given tothe scale issue (i.e. incommensurability) of comparingpoint-scale elevation and piezometric measurements versus600-m grid predictions.

Conclusions

Three types of gridding methods were applied to producecoarse DEMs (600-m resolution) for the modelling of theGete catchment with the MIKE SHE model. The first typeof gridding method uses input elevation data distributedabout the periphery of the gridded DEM cells (i.e. methods(A) and (C)) and was implemented with a MIKE SHE pre-pro-cessing tool for interpolation. The second type uses inputelevation data distributed about the centre of the griddedcells (i.e. methods (B) and (D)). The third type is based onthe TOPOGRID (ESRI, 1996) algorithm that uses landscapefeatures, such as digitised streams, to improve the drainagestructure of the DEM product (i.e. method (E)).

A protocol, examining the accuracy of DEM elevations,evaluating geomorphic relationships and predicting hydro-logic conditions in hillslopes, was applied in this work forcharacterising the quality of the coarse DEM products forhydrologic use. The protocol revealed that, for the particu-lar characteristics of the study site and the elevation inputdata, gridding methods type II ((B) and (D)) produced coarseDEMs with higher elevation accuracy, followed by TOPO-GRID and finally by the MIKE SHE tool for interpolation (grid-ding methods type I). Correspondingly, the Multi-Calibration(MCal) analysis revealed a better performance (for outletdischarges and peakflows) of the hydrologic models relatedto gridding methods type II, regardless of the presence ofspurious sinks.

Thus, this study revealed that, in general, the DEMproducts of the gridding methods type II are more appro-priate for the current coarse modelling resolution. In thiscontext, the assessment of the model performance re-vealed a congruence with the predictions of overland flowgeneration from the topographic index analysis, that is,higher runoff production induced by the DEM products ofgridding methods types II and III and smoother runoff pro-ductions related to the DEM products of gridding methodstype I ((A) and (C)).

Some of the piezometric results suggested however apotential underestimation of base flow associated to theDEM products of gridding methods type II that could not


be studied further owing to the lack of baseflow measure-ments or estimates. The present research could thereforebe extended to investigate in the future a protocol forestimating total hydrograph subflows and assessing in thisway the performance of the hydrologic models simulatingsubflows. The subflow analysis may have the potential ofimproving the simulation of certain processes that other-wise might be simulated wrongly at the current 600-mresolution.

The Multi-Site (MS) test indicated moreover that all ofthe hydrologic models predict distributed state variableswith even lower performances than the performances cor-responding to calibrated variables. This test illustrated fur-thermore the importance of using distributed observationsof streamflow and piezometric levels for evaluating themodel performance simulating internal state variables.The significant variability of the model performance resultsindicated however the inherent difficulties in achievingthis distributed evaluation, including incommensurabilityissues, owing among other factors to data uncertaintyand scale-dependent aspects that affected the direct com-parison of point-scale observations versus 600-m grid pre-dictions. It is important therefore to complement in thefuture the current analysis by including in the distributedevaluation protocol estimated intervals of data uncertaintythat could enable accounting not only for data errors butalso for discrepancies between point-scale measurementsand grid-scale predictions. In this respect, the analysis ofthe procedures that were followed up for deriving dis-charge observations from the rating (i.e. level versus dis-charge) curves and the analysis on the discrepanciesamong the input elevation data (Zsource), the DEM eleva-tions (ZDEM) and the ground levels utilised for monitoringthe observation wells (Zmonitor) are likely to play an impor-tant role in the estimation of the referred data uncertaintyintervals.

The MCal test using DEMs without spurious sinks revealedan average influence of the gridding methods on the effec-tive values adopted by the parameters of the hydrologicmodels, except for the DEM (C) that has a drainage outletlocated in a different position with respect to the locationof the catchment outlet.

Artificial sinks were removed from the DEM outcomes ofgridding methods types I and II. This assessment revealedthat the products of gridding methods type II include ahigher amount of artificial sinks than the outcomes of typeI. Comparing the conditions after the sink removal opera-tion with the conditions observed prior the referred oper-ation, the MCal analysis after sink removal showed atendency for obtaining zdr, Tdr and Ksat–Kw parameter val-ues incrementing the overland and interflow volumes andaccelerating the routing of these hydrograph components,but controlling at the same time and as much as possiblethe magnitude of peaks. This general tendency is to com-pensate the deterioration of model predictions due to thesmoothing (i.e. flattering) of DEMs caused by the sink re-moval and explains as well the variation of baseflow pre-diction noticed through the analysis of piezometric levelsto compensate for the changes in overland and interflowvolumes.

Despite the multi-objective and systematic approach tomulti-calibration, the trial and error methodology that

was used in this research is based on the concept ofattaining a single optimum set of parameters. However,given the high dimensionality of the parameter spaceassociated with the distributed model of the study site,it is likely that this parameter space was not adequatelysampled with the consequent risk of having identified onlya local optimum rather than a global optimum. Further-more, the calibration of distributed models is usually fac-ing the risk of parameter equifinality, that is, several setsof parameter values that give acceptable fits to the cali-bration data might be scattered widely in the parameterspace, as a result of errors in the data and model struc-ture, besides parameter interactions (Beven and Freer,2001).

Thus, an important future activity is to define predictionlimits for estimating the degree of confidence on the cur-rent hydrological modelling by taking into account in thescope of a joint deterministic-stochastic framework theuncertainties in data, model structure and parameters. Inthis context, the current research should be understood asa preliminary sensitivity analysis aiming to reduce as muchas possible the geomorphologic data uncertainty by analy-sing in a complementary way the accuracy of DEMs andthe associated model performances.

Finally, the reader should be aware that some of the re-sults obtained in this research, in particular about the out-comes of the DEM gridding methods, may be modelstructure, modelling resolution and catchment specific.The main objective of this manuscript is to communicateto the reader the importance of assessing the quality ofthe topographic input data (i.e. identifying intrinsic errors)to avoid negative consequences on the hydrologicalmodelling.

Acknowledgements

This work was possible thanks to research grants from theOSTC (Belgian Federal Office for Scientific, Technical andCultural Affairs, project CG/DD/08C), the InteruniversityProgramme in Water Resources Engineering (IUPWARE,KULeuven-VUBrussel) and the Katholieke Universiteit Leu-ven (postdoctoral project PDM/03/188, awarded to thefirst author). The completion of this article was achievedin the framework of the Research and Development con-tract of the first author funded by the Instituto Nacionalde Investigacion y Tecnologıa Agraria y Alimentaria (INIA,Spain) and the Centro de Investigacion y Tecnologıa Agro-alimentaria de la Diputacion General de Aragon (CITA-DGA, Spain). The authors would like to thank those whohave supported us and helped to clarify our way through-out this continued research. Special thanks go to LucFeyen, Patrick Willems and Prof. Keith Beven for their con-structive suggestions.

References

Allen, G.R., Pereira, L.S., Raes, D., Martin, S., 1998. Cropevapotranspiration-guidelines for computing crop water require-ments. FAO Irrigation and Drainage Paper, 56, Food andAgriculture Organization, Rome, 290pp.


Ambroise, B., Beven, K.J., Freer, J., 1996. Towards a generalisationof the TOPMODEL concepts: topographic indices of hydrologicsimilarity. Water Resources Research 32 (7), 2135–2145.

Beasley, D.B., Huggins, F., Monke, E.J., 1980. ANSWERS: a modelfor watershed planning. Transactions of the ASAE 23 (4), 938–944.

Beldring, S., 2002. Multi-criteria validation of a precipitation-runoffmodel. Journal of Hydrology 257, 189–211.

Bergstrom, S., Graham, L.P., 1998. On the scale problem inhydrologic modelling. Journal of Hydrology 211, 253–265.

Beven, K., Freer, J., 2001. A dynamic TOPMODEL. HydrologicalProcesses 15, 1993–2011.

Beven, K.J., Lamb, R., Quinn, P.F., Romanowicz, R., Freer, J.,1995. TOPMODEL. In: Singh, V.P. (Ed.), Computer Models ofWatershed Hydrology. Water Resources Publications, USA, pp.627–668.

Braud, I., Fernandez, P., Bouraoui, F., 1999. Study of the rainfall-runoff process in the Andes region using a continuous distributedmodel. Journal of Hydrology 216, 155–171.

Christiaens, K., Feyen, J., 2002. Constraining soil hydraulic param-eter and output uncertainty of the distributed hydrological MIKESHE model using the GLUE framework. Hydrological Processes 16(2), 373–391.

DHI, 1998. MIKE-SHE v.5.30 User Guide and Technical ReferenceManual, Danish Hydraulic Institute, Denmark, 50pp.

Doorenbos, J., Pruitt, W.O., 1977. Crop water requirements. FAOIrrigation and Drainage Paper, 24, Food and Agriculture Organi-sation, Rome, 156pp.

ESRI, 1996. ARC/INFO Online Documentation. Environmental Sys-tems Research Institute Incorporated, Redlands, CA, USA.

Feyen, L., Vazquez, R.F., Christiaens, K., Sels, O., Feyen, J., 2000.Application of a distributed physically-based hydrologic model toa medium size catchment. Hydrology and Earth System Sciences4 (1), 47–63.

Guntner, A., Uhlembrook, S., Seibert, J., Leibundgut, Ch., 1999.Multi-criterial validation of TOPMODEL in a mountainous catch-ment. Hydrological Processes 13, 1603–1620.

Hanselman, D., Littlefield, B., 1998. Mastering MATLAB 5: AComprehensive Tutorial and Reference. Prentice-Hall Inc., UK,pp. 638.

Hutchinson, M.F., 1989. A new procedure for gridding elevation andstream line data with automatic removal of spurous pits. Journalof Hydrology 106, 211–232.

Hutchinson, M.F., Dowling, T.I., 1991. A continental hydrologicalassessment of a new grid-based Digital Elevation Model ofAustralia. Hydrological Processes 5, 45–58.

Jain, S.K., Storm, B., Bathurst, J.C., Refsgaard, J.C., Singh, R.D.,1992. Application of the SHE to catchments in India–Part 2: fieldexperiments and simulation studies on the Kolar Subcatchmentof the Narmada River. Journal of Hydrology 140, 25–47.

Jayatilaka, C.J., Storm, B., Mudgway, L.B., 1998. Simulation ofwater flow on irrigation bay scale with MIKE SHE. Journal ofHydrology 208, 108–130.

Jenson, S., Domingue, J., 1988. Extracting topographic structurefrom digital elevation data for geographic information systemanalysis. Photogrametric Engineering and Remote Sensing 54(11), 1593–1600.

Kristensen, K.J., Jensen, S.E., 1975. A model for estimating actualevapotranspiration from potential evapotranspiration. NordicHydrology 6, 170–188.

Labs, Clark, 1998. Online Idrisi32 Help SystemClark Labs. ClarkUniversity, Worcester, MA, USA.

Lane, S.N., Brookes, C.J., Kirkby, M.J., Holden, J., 2004. A network-index-based version of TOPMODEL for use with high-resolutiondigital topographic data. Hydrological Processes 18, 191–201.

Legates, D.R., McCabe, G.J., 1999. Evaluating the use of ‘‘good-ness-of-fit’’ measures in hydrologic and hydroclimate modelvalidation. Water Resources Research 35 (1), 233–241.

Loague, K., Green, R.E., 1991. Statistical and graphical methods forevaluating solute transport models: overview and applications.Journal of Contaminant Hydrology 7, 51–73.

Madsen, H., 2003. Parameter estimation in distributed hydrologicalcatchment modelling using automatic calibration with multipleobjectives. Advances in Water Resources 26 (2), 205–216.

Quinn, P., Beven, K., Chevallier, P., Planchon, O., 1991. ThePrediction of Hillslope Flow Paths for Distributed HydrologicalModelling Using Digital Terrain Models. Hydrological Processes 5,59–79.

Quinn, P., Beven, K.J., Lamb, R., 1995. The ln(a/tanb) index: howto calculate it and how to use it within the TOPMODELframework. Hydrological Processes 9, 161–182.

Refsgaard, J.C., 1997. Parameterisation, calibration and validationof distributed hydrological models. Journal of Hydrology 198,69–97.

Refsgaard, J.C., Knudsen, J., 1996. Operational validation andintercomparison of different types of hydrological models.Water Resources Research 32, 2189–2202.

Refsgaard, J.C., Storm, B., 1995. MIKE SHE. In: Singh, V.P. (Ed.),Computer Models of Watershed Hydrology. Water ResourcesPublications, USA, pp. 809–846.

Saulnier, G-M., Obled, Ch., Beven, K., 1997a. Analytical compen-sation between DEM grid resolution and effective values ofsaturated hydraulic conductivity within the TOPMODEL frame-work. Hydrological Processes 11, 1331–1346.

Saulnier, G-M., Beven, K., Obled, Ch., 1997b. Digital elevationanalysis for distributed hydrological modelling: Reducing scaledependence in effective hydraulic conductivity values. WaterResources Research 33 (9), 2097–2101.

Vander Poorten, K., Deckers, J., 1994. A 1:250,000 Scale soil map ofBelgium. Proceedings of the Fifth European Conference andExhibition on Geographic Information Systems, EGIS ‘94, Utrecht1, 1007–1015.

Vazquez, R.F., 2003. Assessment of the performance of physicallybased distributed codes simulating medium size hydrologicalsystems. PhD dissertation ISBN 90-5682-416-3, Department ofCivil Engineering, K.U. Leuven, Belgium, 335pp.

Vazquez, R.F., Feyen, J., 2003a. Assessment of the performance ofa distributed code in relation to the ETp estimates. WaterResources Management 16 (4), 329–350.

Vazquez, R.F., Feyen, J., 2003b. Effect of potential evapotranspi-ration estimates on effective parameters and performance ofthe MIKE SHE-code applied to a medium-size catchment. Journalof Hydrology 270 (4), 309–327.

Vazquez, R.F., Feyen, J., 2004. Potential Evapotranspiration for thedistributed modelling of Belgian basins. Journal of Irrigation andDrainage Engineering 130 (1), 1–8.

Vazquez, R.F., Feyen, L., Feyen, J., Refsgaard, J.C., 2002. Effect ofgrid-size on effective parameters and model performance of theMIKE SHE code applied to a medium sized catchment. Hydro-logical Processes 16 (2), 355–372.

Vertessy, R.A., Hatton, T.J., O’Shaughnessy, P.J., Jayasuriya,M.D.A., 1993. Predicting water yield from a mountain ash forestcatchment using a terrain analysis based catchment model.Journal of Hydrology 150, 665–700.

Vivoni, E.R., Ivanov, V.Y., Bras, R.L., Entekhabi, D., 2005. On theeffect of triangulated terrain resolution on distributed hydro-logic modeling. Hydrological Processes 19 (11), 2101–2122.

Walker, J.P., Willgoose, G.R., 1999. On the effect of digitalelevation model accuracy on hydrology and geomorphology.Water Resources Research 35 (7), 2259–2268.

Willems, P., 2000. Probabilistic immission modelling of receivingsurface waters. PhD dissertation ISBN 90-5682-271-3, Depart-ment of Civil Engineering, K.U. Leuven, Belgium, 233pp.

Wise, S., 2000. Assessing the quality for hydrologic applications ofdigital elevation models derived from contours. HydrologicalProcesses 14, 1909–1929.


Wolock, D.M., Price, C.V., 1994. Effects of digital elevation modelmap scale and data resolution on a topography-basedwatershed model. Water Resources Research 30 (11), 3041–3052.

Xevi, E., Christiaens, K., Espino, A., Sewnandan, W., Mallants, D.,Sorensen, H., Feyen, J., 1997. Calibration, validation and

sensitivity analysis of the MIKE-SHE model using the Neuenkir-chen catchment as case study. Water Resources Management 11,219–239.

Zhang, W., Montgomery, D.R., 1994. Digital elevation model gridsize, landscape representation, and hydrologic simulations.Water Resources Research 30 (4), 1019–1028.

mike she usa

Environment

gridded dems

dem gridding methods

different dems

effects of dem gridding

gridded dem cells

dem products

dem generation methods

elevation quality of