Case Studies

Effect of Single and Multisite Calibration Techniques on

the Parameter Estimation, Performance, and Output of a
SWAT Model of a Spatially Heterogeneous Catchment
Olkeba Tolessa Leta 1; Ann van Griensven 2; and Willy Bauwens 3
Downloaded from ascelibrary.org by University of Hong Kong on 10/02/16. Copyright ASCE. For personal use only; all rights reserved.

Abstract: Although the soil and water assessment tool (SWAT) is a physically based hydrologic simulator, it has many parameters
that cannot be measured directly in the field, but must be obtained through a model calibration process. Model calibration is thus an essential
task to obtain the optimal parameter values, which match simulations with observations as closely as possible. This study used the Zenne
River Basin (Belgium) as a case study, which experiences high spatial heterogeneity in terms of geological formation, groundwater recharge,
and rainfall-runoff responses. Therefore, the objectives of this paper were to calibrate the SWAT model on the basis of different calibration
techniques and identify which technique is suitable for such a heterogeneous basin so that the calibrated SWAT can be used as a tool for
integrated management of the Zenne River Basin. Prior to calibration, the sensitive parameters were identified on the basis of a detailed
sensitivity analysis (SA) of the Latin hypercube one-factor-at-a-time (LH-OAT) technique and increased sampling intervals. Then, SWAT
was calibrated by using single-site calibration (SSC) at the watershed outlet; sequential calibration (SC), calibration from upstream to
downstream; and simultaneous multisite calibration (SMSC), where data of two flow gauging stations were simultaneously used in a single
calibration. It was found that at least 200 sampling intervals should be considered for the LH-OAT SA method to obtain converged rankings
of SWAT parameters. In addition, to well capture the spatial variability of heterogeneous catchment and achieve stable sensitivity ranking,
simultaneous multisite SA technique is important. Streamflow findings suggested that the SC and the SMSC techniques provided very good
results and significantly improved model performance, but the SSC results were merely satisfactory. Though the results of SC and SMSC
techniques were similar, the SMSC was selected over SC because it simultaneously handles the entire catchment spatial variability
by assigning different parameter values and allows data information communication among stations in a single calibration. It was thus
concluded that simultaneous multisite calibration should be considered for catchments with a high spatial variability like the Zenne River
Basin. DOI: 10.1061/(ASCE)HE.1943-5584.0001471. © 2016 American Society of Civil Engineers.
Author keywords: Calibration; Multisite; Streamflow; Soil and water assessment tool (SWAT); Zenne River Basin; Belgium.

Introduction Vansteenkiste et al. 2013). Furthermore, distributed hydrologic

simulators can help decision makers to better understand environ-
Spatially distributed hydrologic simulators can be used for model- mental related problems and design appropriate mitigation mea-
ing different catchment processes, including evapotranspiration, sures (Debele et al. 2008; Leta et al. 2014; Shrestha et al. 2013b).
interception, infiltration, percolation, surface runoff, groundwater As such, they are useful tools for integrated river basin management
flow, soil erosion and sediment transport, and nutrients transport. studies and decision making (Andrews et al. 2011; Migliaccio and
To represent these processes, spatially distributed hydrologic sim- Chaubey 2007; Moriasi et al. 2007).
ulators primarily use physically-based mathematical equations. As a result of the advances in techniques for data acquisition and
Such simulators also can be used for investigating the impacts of measurement, the use of spatially distributed, physically based, and
land use change, climate change, agricultural activities, and various thus complex hydrologic simulators has gained increased attention
management practices on the catchment processes (Bae et al. 2011; for integrated river basin management and decision support (Ajami
Li et al. 2009; Mukundan et al. 2013; Narasimhan et al. 2010;
et al. 2004; Arnold et al. 2010). A typical example of such simu-
Santhi et al. 2006; Schilling et al. 2008; Singh et al. 2005;
lators is the soil and water assessment tool (SWAT) (Arnold et al.
1998). SWAT is a semidistributed and physically based simulator
1 that is used widely for continuous hydrological, soil erosion, sedi-
Postdoctoral Research Fellow, Water Resources Research Center,
Univ. of Hawaii at Manoa, 2525 Correa Rd., HIG 217, Honolulu,
ment, and nutrients transport modeling (Arnold et al. 2012;
HI 96822 (corresponding author). E-mail: [email protected] Debele et al. 2008; Gassman et al. 2007; Santhi et al. 2001, 2008).
2
Professor, Dept. of Hydrology and Hydraulic Engineering, Vrije The simulator considers the interactions among water, soil, and
Universiteit Brussel (VUB), Pleinlaan 2, 1050 Brussels, Belgium; Associ- plants in the upland catchments in a quite detailed way (Gassman
ate Professor, Dept. of Water Science and Engineering, UNESCO-IHE et al. 2007). However, simulators such as SWAT have many param-
Institute for Water Education, Westvest 7, 2611 AX, Delft, Netherlands. eters that cannot be measured directly in the field, but should be
3
Professor, Dept. of Hydrology and Hydraulic Engineering, Vrije obtained through a model calibration process (Hogue et al.
Universiteit Brussel (VUB), Pleinlaan 2, 1050 Brussels, Belgium.
2000; Li et al. 2010). Even if some parameters of spatially distrib-
Note. This manuscript was submitted on January 11, 2016; approved on
August 2, 2016; published online on September 30, 2016. Discussion per- uted hydrologic simulators could be measured, scaling problems
iod open until February 28, 2017; separate discussions must be submitted still would have to be solved (Nossent 2012). Therefore, model cal-
for individual papers. This paper is part of the Journal of Hydrologic En- ibration and parameter estimation are essential tasks to obtain
gineering, © ASCE, ISSN 1084-0699. the optimal parameter values that can reasonably match simulations

© ASCE 05016036-1 J. Hydrol. Eng.

J. Hydrol. Eng., 05016036

 with observations. Only then the calibrated model can be used as a have been presented in the literature. This study will investigate the
tool for integrated river basin management and scenario analysis. effect of different calibration techniques on SWAT model parameter
During the calibration process, parameter values are adjusted estimations, outputs, and performance by using single-site, sequen-
until the model simulations consistently match the observations tial, and joint multiple-sites calibration techniques for hetero-
as closely as possible. Calibration has been usually practiced by geneous catchments.
minimizing the difference between simulations and observations, First, the results of a single-site calibration (SSC) technique will
e.g., through the use of the sum of squared residuals (SSR), as be discussed for the SWAT model of the Zenne River Basin. This
an objective function. Nevertheless, when the number of parame- method will use the measured daily streamflow data at the catch-
ters in a simulator is large because of a large number of catchment ment outlet. Next, the results for sequential calibration (SC) and
hydrological processes that are considered, calibration of all the simultaneous, multisite calibration (SMSC) techniques will be
parameters becomes a very challenging task. This is primarily be- shown, in which data of two streamflow observation sites are used.
cause of the over-parameterization of physically based simulators Finally, the results of the three calibration techniques will be com-
that lead to complex calibration processes, time consuming and pared and evaluated.
Downloaded from ascelibrary.org by University of Hong Kong on 10/02/16. Copyright ASCE. For personal use only; all rights reserved.

computationally intensive during model parameters optimization

(Cibin et al. 2010; Nossent and Bauwens 2012). In such
cases, a sensitivity analysis (SA) is an essential step to identify Materials and Methods
the most influential parameters that can potentially impact the
simulator’s outputs. Because SA plays a significant role in model SWAT Simulator
calibration and parameter estimation, uncertainty analysis, and sce-
SWAT is a basin-scale, physically based, and semidistributed
nario analysis (Pappenberger et al. 2008; Ravalico et al. 2010), it is
hydrologic simulator that operates at different time scales. The sim-
an initial step of the model calibration process, especially for a
ulator was originally developed to predict the impact of watershed
high-dimensional model. Then, calibration of a model can focus
management on water, sediment, nutrients, and agricultural and
on a limited set of influential parameters.
chemical yields (Arnold et al. 1998). SWAT also can integrate com-
The most widely used model calibration technique is a single-
plex basins with varying land use, weather, soils, topography, and
site calibration, in which observations (e.g., streamflow) from
management conditions over a long period of time. A basin is di-
a single gauging location (usually at the basin outlet) are used
vided into a number of subbasins that have homogeneous climatic
(Bannwarth et al. 2015; Shi et al. 2013). However, the applicability
conditions (Van Liew et al. 2005). Subbasins are further subdivided
of this technique for complex and spatially heterogeneous basins is
into hydrological response units (HRU) on the basis of a homog-
questionable because only one output information at the basin out- enous combination of land use, soil type, and slope value within
let is used for model constraint. Such a method may yield physi- subbasin (Arnold et al. 2011).
cally unrealistic parameter values, which might not appropriately SWAT uses a water balance equation that includes precipitation,
represent the variability of the processes within the basin. Alterna- surface runoff, actual evapotranspiration, interflow, percolation,
tively, a multisite calibration technique, in which one or more return flow, and deep groundwater losses (Neitsch et al. 2011).
variables and objective functions are considered at multiple sites The simulator uses a modification of the soil conservation ser-
(Niu et al. 2014; Piniewski and Okruszko 2011; Shope et al. 2014; vice-curve number (SCS-CN) method (USDA-SCS 1986), which
Zhang et al. 2008), can be used to appropriately represent the spa- determines the surface runoff on the basis of the area’s hydrologic
tial variability of a given basin. This method is indeed expected group, land use, and antecedent moisture content for each HRU.
to improve the performance of spatially distributed hydrologic The percolation through each soil layer is estimated using a storage
simulators by representing the spatial variability with different routing technique (Arnold et al. 1995). SWAT offers different meth-
parameter values. In this way, multisite calibration techniques ods to estimate the potential evapotranspiration (PET) from climate
can minimize the chance to optimize the model to physically unre- data (Neitsch et al. 2011). River routing can be performed by the
alistic parameter values. Thus, multisite calibration can poten- variable storage method (Williams 1969) or by the Muskingum
tially improve hydrologic partitioning of a spatially heterogeneous method (Chow 1959). In this study, a PET time series that was cal-
catchment because it may reduce parameter uncertainty through culated on the basis of the Penman-Monteith method, the SCS-CN
introducing more constraints (e.g., calibration data) during the method for surface runoff simulations, and the Muskingum routing
calibration process. In addition, for simulators like SWAT, which method for daily streamflow routing were used.
compute the hydrological processes at three spatial scales (basin,
subbasin, and hydrological response unit), a multisite calibration
technique can improve the representation of the different basin Study Area
hydrological processes (Migliaccio and Chaubey 2008). Hence, The Zenne River Basin is located in the central part of Belgium and
model calibration at different locations of a basin is crucial for is a part of the international Scheldt Basin. The river drains an area
the SWAT model, particularly for a spatially heterogeneous catch- of 1,162 km2 that traverses through the three administrative regions
ment. Although multisite calibration techniques are expected to of the country [Fig. 1(a)]. Upstream from Brussels, the river follows
improve hydrological process representation of spatially hetero- a natural meandering course, whereas in the Brussels-Capital re-
geneous catchment and reduce parameter uncertainty, one should gion, it has been vaulted over a distance of approximately 8 km.
not expect that it significantly affects the equifinality or parameters The river basin is crossed by the canal Brussels-Charleroi and the
nonuniqueness issue (Beven 2006; Cao et al. 2007; Migliaccio and sea canal Brussels-Scheldt that runs parallel to the river [Fig. 1(a)].
Chaubey 2007). Thus, a large number of alternative parameteriza- In the Walloon region, the canal is fed by the former tributaries of
tions still can produce similar and acceptable results even if multi- the river (Hain, Thines, and Samme). To prevent flooding in the
site calibration techniques are assumed to be the best calibration Brussels capital region, the canal also receives water from the river
strategy for heterogeneous catchments. through several overflow structures. Excess water in the canal can
So far, few comparisons of single and multisite calibration tech- be discharged back into the river downstream of Brussels. The geo-
niques for highly heterogeneous basins, such as the Zenne River logical map of the study area shows a clear difference in 2 regions
Basin (Belgium) and the complex hydrological simulator SWAT, [Fig. 3(a)] whereas the land cover and soil maps show similarity

© ASCE 05016036-2 J. Hydrol. Eng.

J. Hydrol. Eng., 05016036

Downloaded from ascelibrary.org by University of Hong Kong on 10/02/16. Copyright ASCE. For personal use only; all rights reserved.

Fig. 1. (a) Location of the Zenne River Basin in Belgium with the three regions; (b) the study area with its hydro-meteorological stations

between the 2 regions [Figs. 2(a and b)]. While the information • A 20 × 20 m land use map on the basis of data from the OC GIS
on the clear demarcation of geological formation will be used to Vlaanderen and the Ministère de la Région Wallonne-Direction
delineate the study area into two regions that facilitates the Générale de I’Aménagement du Territoire, du Logement et du
model calibration process, the land use, the soil type, and the topo- Patrimoine. The data stem from approximately 2000 and are still
graphic information will be used during model parameters representative for the period under study.
discretization. Daily precipitation data were obtained from the Royal Meteoro-
The study reported in this paper focuses on the upstream of logical Institute (RMI) of Belgium at the Ukkel station; from
the Zenne River Basin [Fig. 1(b)], which ends at the Lot gauging the Direction Générale Opérationnelle de la Mobilité et des Voies
station [Fig. 2(d)]. The upstream catchment has an area of approx- hydrauliques (DGVH) at Enghien, Seneffe, Soignies, and Wautier-
imately 747 km2 and a population density of 470 inhabitants per Braine stations; and from the Vlaamse Milieumaatschappij (VMM)
km2 . Its elevation ranges from 18 to 171 m above mean sea level at the St.Pieters Leeuw station [Fig. 1(b)]. Daily maximum and
[Fig. 2(c)]. The upstream catchment is dominated by agricultural minimum temperatures, wind speed, solar radiation, and relative
land use (56%), followed by pasture (22%) and mixed forest humidity were only available at the Ukkel station. For the model
(11%), and the remaining areas are covered by other land uses parameter optimization, daily streamflow data measured at Tubize
(urban, rangeland, and water bodies) [Fig. 2(a)]. The upstream top by DGVH and at Lot by VMM were used [Fig. 2(c)].
soil layer is predominately covered by loam soils (79%), whereas Lastly, from the time of analyses of this study, the authors be-
the other soils (anthropogenic, sandy loam, sand, loamy sand, and came aware that SWAT 2012 was available, but the newer version
clay) account for 21% of the area [Fig. 2(b)]. was not stable during the study period. In addition, compared to
SWAT 2009, there was not significant modification to the hydro-
logic component of SWAT 2012. Therefore, the use of SWAT 2009
Data version would not affect the outcome of this study because this
The ArcGIS version SWAT 2009 model was built on the basis of: work evaluated the influence of single and multisite calibration
• A 30 × 30 m digital elevation model (DEM; Brussels, Belgium) techniques on streamflow. Consequently, this study used the more
obtained from Vlaamse Land Maatschappij OC GIS-Vlaanderen stable SWAT 2009 version.
(VLM-OC-GIS) and the Advanced Spaceborne Thermal
Emission and Reflection Radiometer of Global DEM (ASTER
Model Setup
GDEM);
• A 20 × 20 m soil map from OC GIS Vlaanderen and the Carte The SWAT model was setup on the basis of the available geospatial
Numérique des Sols de Wallonie; and data (DEM, land use, and soil maps) and hydro-meteorological

© ASCE 05016036-3 J. Hydrol. Eng.

J. Hydrol. Eng., 05016036

Downloaded from ascelibrary.org by University of Hong Kong on 10/02/16. Copyright ASCE. For personal use only; all rights reserved.

Fig. 2. (a) Land use; (b) soil type; (c) DEM of the study area; (d) the catchment boundaries for calibration and delineated subbasins

data. Using the DEM [Fig. 2(c)], the upstream part of the catch- values signify that if the percent of land use, soil type, and slope
ment, which is the part of the Zenne catchment up to the Lot gaug- class are below the defined values, they will be excluded from HRU
ing station, was divided into 21 subbasins [Fig. 2(d)]. The following classification. Larger threshold values for soil and slope class were
considerations were taken into account: set in a way that cannot neglect the soil types and slope classes of
• Because SWAT allows only one (flow or pollution) point source the area. In addition, though the topographic variability of the area
per subbasin, the subbasin outlets were defined at each location is low, this study used two slope classes of 0–7 and >7%, instead of
of major wastewater treatment plants (WWTP) [Fig. 1(b)]. assuming a uniform (average) slope.
• The upstream area, which drains to the canal Brussels-Charleroi To prevent flooding in Brussels, the overflow at subbasin 6 was
in the southeast part of the catchment, was excluded from the designed for extremely high flow conditions. It is not supposed to
watershed modeling process [Fig. 2(d)]. Thus, the correspond- be operational during the low flow events. As SWAT is not able to
ing simulated streamflows from this area were not considered on take (the operation of) weirs into consideration, a simplified rep-
the downstream subbasins. resentation of the overflow was implemented within the model
• An additional outlet was manually defined at the overflow code (van Griensven 2002). For low flows, when the flow is smaller
between the Zenne River and the canal in Lembeek, which is than a constant threshold rate Qlim , the following equation holds:
located at the outlet of subbasin 6 [Fig. 2(d)].
If Qi ≤ Qlim ; Qo ¼ 0; else Qt ¼ Qi ð1Þ
The delineated subbasins were further subdivided into different
HRUs on the basis of the land use, soil, and slope class of the basin. where Qi = inflow rate at the upstream of the weir; Qo = overflow
The HRUs were defined on the basis of threshold values of 5% rate to the canal; and Qt = outflow rate at the downstream of
for land use, 15% for soil, and 35% for slope class. The threshold the weir.

© ASCE 05016036-4 J. Hydrol. Eng.

J. Hydrol. Eng., 05016036

Downloaded from ascelibrary.org by University of Hong Kong on 10/02/16. Copyright ASCE. For personal use only; all rights reserved.

Fig. 3. (a) Geological map of the Zenne Basin; (b) the annual average recharge map of the upstream Zenne Basin (adapted from Meyus et al. 2004,
with permission)

When the upstream inflow rate exceeds the threshold rate, the In addition, to understand the effect of the four calibration tech-
following holds: niques on streamflow hydrographs and for clear illustration, the
modeled area was categorized into three regions: Catchment I, rep-
If Qi > Qlim ; Qo ¼ ðQi − Qlim Þ × fo ; and resenting the upstream area up to Lot; Subcatchment II, which re-
Qt ¼ Qlim þ ðQi − Qlim Þ × ð1 − fo Þ ð2Þ fers to the area upstream of Tubize; and Subcatchment III, which is
Catchment I minus Subcatchment II [Fig. 2(d)]. While the SSC
with fo , the overflow fraction, above a threshold value Qlim . technique was applied to Catchment I, Subcatchments II and III
The parameters Qlim and f o were considered as calibration were used for the application of SC, SMSC, and SMSC-LSRP tech-
parameters. To verify the overflow performance of SWAT, the niques. For the SSC technique, the observations at Tubize, which is
simulated overflow results of SWAT at the outlet of subbasin 6 were inside Catchment I, were used only for validation purposes.
compared with the hydraulic model results of the Zenne River For the SSC calibration procedure, although the same (lumped)
(Shrestha 2013). groundwater related parameter values were assigned for the entire
area of Catchment I, parameters such as soil water holding capacity
and surface runoff curve number at soil moisture condition II that
Model Calibration are spatially variable depending on land use, soil, and slope types,
In this study, the following calibration techniques were applied: were adjusted using global multipliers or relative change to their
1. SSC: calibration at the main watershed outlet only; original values. In such way, the natural variability and hetero-
2. SC: step-by-step calibration, starting at the upstream station and geneity of the catchment was retained. Nevertheless, in the case
subsequently moving to downstream; of SC and SMSC techniques, different parameter values were used
3. SMSC: in which data of two stations were used in a single between Subcatchments II and III. Different values were consid-
calibration; and ered for groundwater-related parameters because of clear demarca-
4. Simultaneous, multisite calibration with lumped surface runoff tion in geological formation and groundwater recharge between
parameters (SMSC-LSRP): similar to SMSC, but lumped values Subcatchments II and III (Fig. 3). The SMSC-LSRP technique
were assigned for surface runoff related parameters. is similar to SMSC, but the surface runoff related parameters were
To run the aforementioned calibration techniques, the ob- assumed to vary uniformly (the same global multipliers) as a result
served daily flows at the Lot and the Tubize stations were used. of insignificant differences in land use between the two regions

© ASCE 05016036-5 J. Hydrol. Eng.

J. Hydrol. Eng., 05016036

 Varies with soil type. For these parameters, relative change was used during calibration process and the calibrated ranges were obtained by multiplying the optimized relative value with the original values on the
SMSC-LSRP
(Fig. 2). Furthermore, despite using the same parameters of SSC

0.14–0.30
2.4–9.7

59–83
0.001

34.98

13.73
14.39

18.00
0.04

0.01

0.15

0.79

0.80
(Table 1), the values of those parameters were separately defined
as a set for Subcatchments II and III in the necessary files
(e.g., changepar) of the SWAT project. The approach of SC and
SMSC techniques were generally similar, but they differed on

Lot (Subcatchment III)

0.14–0.30
how the parameter values were set and optimized. For the SC tech-

2.0–8.0
SMSC

47–67
0.001

19.41

87.80
17.19

18.00
0.03

0.00

0.19

0.94

0.80
nique, the parameters of Subcatchment II first, the parameters of
Subcatchment II were optimized and then fixed to their optimal
values. Then, the parameters of Subcatchment III were calibrated

0.15–0.32
(optimized). In contrast to SC technique, the SMSC technique al-

2.0–10.0

Note: SC = sequential calibration; SMSC = simultaneous, multisite calibration; SMSC = SMSC, but lumped surface runoff related parameters; SSC = single site calibration.
54–77
0.001

73.18
92.59

18.00
0.02

0.00
9.11
0.18

0.53

0.80
SC
lowed to simultaneously change and optimize all parameter sets of
both subcatchments in a single calibration. This also enabled the
SWAT basin scale parameters [e.g., surface runoff lag coefficient,
Downloaded from ascelibrary.org by University of Hong Kong on 10/02/16. Copyright ASCE. For personal use only; all rights reserved.

basis of land use, soil, and slope value. These parameters are spatially varied, whereas groundwater parameters are assumed to be constant within subcatchment.
(SURLAG)] to be calibrated at once, which was not possible in the

0.15–0.32
2.0–7.0

150.00
66–87

40.00

55.00

18.00
SSC
1.00

0.05

0.05

0.15

0.50

0.80
case of the SC technique. As the SWAT version used in this study
did not allow to setup and execute the SMSC technique in its
graphical user interface, some of the SWAT input and output text
files (e.g., changepar, fig.fig, and objmet) were manually manipu-

SMSC-LSRP

0.14–0.30
lated to achieve the multisite calibration technique.

2.4–9.7

63–84
0.997

98.46

18.00
0.04

0.01
6.31
0.20

6.07

0.79

0.80
The model parameters autocalibration was done using the
Shuffled Complex Evolution-University of Arizona (SCE-UA)
algorithm (Duan et al. 1994) as implemented in SWAT (van

Tubize (Subcatchment II)

Griensven et al. 2002, 2006). For this algorithm, the SSR was used

0.14–0.34
100.00
SMSC
to define the objective function, whereas the model performance

71–94
0.967

11.82

18.00
2.2–9
0.04

0.01
4.26
0.20

0.94

0.80
Table 1. Optimized SWAT Parameter Values on the Basis of the Four Applied Calibration Techniques for Zenne Catchment
was evaluated using the Nash-Sutcliffe efficiency (NSE) (Nash
and Sutcliffe 1970). For the SSC and SC techniques, a single ob-
jective function, SSR, was used on the basis of observation from

0.14–0.34
2.0–10.0
only one gauging station. However, in the case of SMSC and

71–94
0.992

96.60

18.00
0.05

0.02
4.40
0.20

1.32

1.33

0.80
SMSC-LSRP techniques, two SSR objective functions were used SC
simultaneously, which were formulated on the basis of the two
streamflow gauging stations, i.e., Tubize and Lot. The SSR at Tub-

0.15–0.32
ize was used to calibrate the parameters of Subcatchment II,
2.0–7.0

150.00
66–87

40.00

55.00

18.00
SSC
1.00

0.05

0.05

0.15

0.50

0.80
whereas the SSR at Lot was used to calibrate the parameters of
Subcatchment III. In this way, the two objective functions were
weighted equally because both represented a single variable
0.02–0.2
0–5,000
(streamflow) at the two sites. By simultaneously calculating the
Range

35–98

0–100

0–500
0–10

0–10
N/A
N/A
0–1

0–1

0–1

0–1
two objective functions at the two sites, SMSC found a set of op-
timal parameter values, which is similar to the Pareto optimization
approach (Zhang et al. 2008). To run the SCE-UA algorithm, the
m3 s−1
Day−1

parameter values were constrained on the basis of their lower and

Unit

Day

Day
mm
mm
mm
—
—
—

upper bounds as obtained from literature (Arnold et al. 2011). —

Then, as suggested by Van Liew et al. (2005), a manual calibration
Minimum depth for groundwater revap occurrence

was carried out for all techniques to obtain well represented annual
Minimum depth for groundwater flow occurrence

Threshold value for overflow occurrence to canal

water-balance components. This was achieved by only fine tuning

water-balance related parameters. Such an approach substantially
Fraction of overflow above threshold value

reduced the time consuming manual calibration, and also allowed

Curve number at moisture condition IIb
Soil evaporation compensation factor

quantitative and qualitative comparisons (Arnold et al. 2012).

To carry out the calibration processes, the SWAT model simu-
Manning’s roughness coefficient
Description

lation time was split into three periods: (1) a warming up period
Groundwater revap coefficient

Soil water available capacityc

Surface runoff lag coefficient

(1985–1987) to initialize the state variables of the system, (2) a

Maximum canopy storagea

calibration period (1998–2008), and (3) a validation period

Base flow alpha factor

Varies with land use, soil, and slope.

(1988–1997). The eleven years (1998–2008) of streamflows that

Groundwater delay

had high, normal, and low flows conditions were selected for the
model calibration. Prior to calibration, a detailed SA was performed
for the calibration period by using the Latin hypercube one-factor-
at-a-time (LH-OAT) technique as implemented in SWAT (van
Varies with land use.

Griensven et al. 2006). Then the model was calibrated using the
most sensitive/important parameters. The applied LH-OAT tech-
nique only yields qualitative results (e.g., parameter rankings); thus,
GW_DELAY
GW_REVAP
ALPHA_BF

the LH-OAT method cannot provide quantitative information

REVAPMN
SOL_AWC
GWQMN
Parameter

SURLAG
CANMX

on the importance level and difference among parameters. Also,

CH_N2

ESCO

the method requires sufficiently large sampling intervals to achieve

CN2

Qlim
fo

converged parameter rankings (Nossent 2012).

b
a

© ASCE 05016036-6 J. Hydrol. Eng.

J. Hydrol. Eng., 05016036

 Model Performance Evaluation index (BLAI), and the primary channel Manning roughness coef-
ficient (CH_N2) parameters, which primarily influence surface run-
Model calibration and validation should include multiple statistical
evaluation criteria, considering that the single statistical metric only off and river flow processes. The sampling intervals also noticeably
evaluates a specific part of model performance (Arnold et al. 2012; affected the sensitivity rank of the groundwater related parameters,
Moriasi et al. 2007). The SWAT model performance was evaluated such as the groundwater delay (GW_DELAY) and the minimum
through graphical comparison and by concurrently using three stat- depth for groundwater revap occurrence (REVAPMN).
istical criteria for goodness-of-fit: the NSE (Nash and Sutcliffe The SA results at the catchment outlet showed that the most
1970), the percent bias (PBIAS) (Moriasi et al. 2007), and the root sensitive parameters were the runoff curve number at soil moisture
mean squared error (RMSE) to the observation standard deviation condition II (CN2) of the Subcatchment II, followed by the CN2 of
ratio (RSR) (Moriasi et al. 2007). the Subcatchment III (Fig. 4). The high sensitivity of the CN2 of the
Finally, to determine the general performance rating of SWAT, Subcatchment II could be expected because the western part of
the daily streamflow performance ratings were formulated on the the basin is generally characterized by a lower infiltration and, thus,
basis of the previously named three statistical metrics as suggested a lower recharge to the groundwater system (Fig. 3). As a conse-
Downloaded from ascelibrary.org by University of Hong Kong on 10/02/16. Copyright ASCE. For personal use only; all rights reserved.

by Moriasi et al. (2007). However, Moriasi et al. (2007) defined these quence, more surface runoff is expected. The BLAI parameter,
ratings for monthly time-step simulations and, in general, the stat- which could affect the surface runoff process and the evapotranspi-
istical metrics become less good for daily time-step simulations. ration, was identified as a parameter with a medium sensitivity for
Thus, the ratings could be flexible for daily time scale. Despite this Subcatchment II, but it was not seen any more for Subcatchment
fact, this study directly used the same ratings and threshold criteria of III. The base flow alpha factor (ALPHA_BF) and the minimum
Moriasi et al. (2007) for the daily streamflow simulations. depth for groundwater flow occurrence (GWQMN), which could
affect the shape of recession hydrograph and groundwater contri-
bution to streamflow, respectively, were found to be the third and
Results and Discussion fourth sensitive parameters for Subcatchment III. This was reason-
able because the groundwater contribution is higher in the down-
stream part of the catchment.
Sensitivity Analysis In general, CN2, CH_N2, ALPHA_BF, CH_K2, maximum
Because the SWAT model has a large number of parameters, the canopy storage (CANMX), GWQMN, recharge losses to the deep
LH-OAT SA method as incorporated in the model was used to aquifer (RCHRG_DP), soil water available capacity (SOL_AWC),
screen out the sensitive parameters. The SA was run using the si- soil evaporation compensation factor (ESCO), GW_DELAY,
multaneous multisite method that simultaneously used the observed SURLAG, and groundwater revap coefficient (GW_REVAP) were
daily streamflow data at Tubize and Lot stations. To well explore the most sensitive parameters for the catchment (see Table 1 for
the parameter hyperspace, the number of sampling intervals were parameters description). However, considering the physical reasons
increased from the SWAT default value of 10–200. Sampling in- related to climate, topography, and geology, the parameters CH_K2
tervals of 200 were found to be sufficient because a stable ranking and RCHRG_DP were set to their SWAT default values. Thus,
was achieved for these intervals. It was found that the number these parameters were not considered for further analysis. For ex-
of sampling intervals may significantly affect the sensitivity rank ample, as the catchment is located in a humid climate and classified
of the SWAT parameters (Fig. 4). The sampling intervals of the as a perennial river, channel transmission losses are not expected in
LH-OAT method primarily affected the rank of the SURLAG, such catchments. Hence, the SWAT default value of zero was used
the channel transmission losses (CH_K2), the potential leaf area for CH_K2. Similarly, the RCHRG_DP, which is considered as a

Fig. 4. Sensitivity rank of the SWAT parameters affected by the number of sampling intervals of the Latin hypercube, one-factor-at-a-time sensitivity
analysis (the first 21 parameters represent Subcatchment III, whereas the remaining parameters represent Subcatchment II)

© ASCE 05016036-7 J. Hydrol. Eng.

J. Hydrol. Eng., 05016036

 loss for the catchment, was not expected for the case study. Con- (SR), and the base flow (BF), were compared with observations. On
sequently, the SWAT default value, which was 5% of total aquifer the basis of the base flow-filtering program (Arnold and Allen
recharge, was presumed to be considered as negligible. 1999), it was estimated that the base flow, which is the sum of the
groundwater flow and the lateral flow, accounts for 67 and 76% of
the total streamflow at Tubize and Lot, respectively. Because of
Calibrated Parameters missing observed streamflow data at Tubize (June–September
The SSC values of GW_DELAY and GWQMN might be too high of 1988–1997), the water balance of validation period at this station
considering the soil and geological characteristics of the catchment was restricted to October–May.
(Table 1). Their values were manually increased to reduce the The SSC technique led to an overestimation of the water yield at
groundwater flow to the river. Conversely, the SC led to different Lot by 10% for the calibration period and 21% for the validation
parameter values (Table 1) as compared with the SSC technique. period. This was most likely because of an overestimation of the
However, the primary drawback of the SC technique is that the surface runoff by the model (Table 2). When the model was inter-
catchment scale parameters, such as the SURLAG, got consider- nally validated at Tubize, the simulated base flow was 76% of the
Downloaded from ascelibrary.org by University of Hong Kong on 10/02/16. Copyright ASCE. For personal use only; all rights reserved.

ably different values. Table 1 shows that the value of the GW_ total flow. This value was higher than the filtered base flow, which
DELAY of Subcatchment III was approximately 5-fold of the value was 67% of the total flow. In contrast, the surface runoff was under-
of Subcatchment II. This could be partly explained by the soil char- estimated by about 8%. Although the model can be categorized as
acteristics or geological formation because the western part of the very good at Lot with respect to the water yield and the base flow,
catchment is characterized by shallow soil depth, low infiltration, the model results at Tubize were unsatisfactory according to the
and groundwater recharge (Fig. 3). Also, the ALPHA_BF of Sub- criteria of Moriasi et al. (2007). This highlights that on the basis of
catchment III was nearly zero, whereas it was close to the upper the information (observation) at the basin outlet, the model was not
bound value for Subcatchment II. The CN2 values of Subcatchment able to represent the heterogeneity of the hydrological processes
II were higher than Subcatchment III. The low value for GW_ within the catchment.
DELAY and the high value for ALPHA_BF for Subcatchment II As opposed to the SSC technique, the SC technique adequately
show that the southwestern part of the catchment rapidly responds represented the water balance components at both stations (Table 2).
to recharge. In addition, the PBIAS indicated a very good model performance
The difference in the optimized parameter values indicated that for WYLD, BF, and SR during the calibration period (Moriasi et al.
the relationship between streamflow, topography, land use, soil, 2007). When compared with the SSC technique, the results of SC
and precipitation was different for the two parts of the basin. This technique suggested that the representation of the hydrological
could be explained partially by the heterogeneity in the geological processes of Subcatchment II was significantly improved as a result
formation of the basin (Meyus et al. 2004; Peeters 2010), as shown of the calibration strategy and the use of more information (obser-
in Fig. 3(a). In addition, the recharge map of the region, as reported vations at two stations instead of one).
by Meyus et al. (2004) and Zomlot et al. (2015), showed that the Considering the SMSC technique, the simulated base flow rep-
western part of the Zenne Basin is essentially acting as a discharge resented 70% of the flow at Tubize and 77% at Lot, indicating a
zone (recharge close to 0), whereas the eastern part of the basin very good agreement with the filtered base flow values (Table 2).
shows a high recharge zone [Fig. 3(b)]. Furthermore, the base flow fractions indicated that the groundwater
The results of the simultaneous, SMSC technique were quite flow contribution was larger in the downstream part of the catch-
similar to the ones of the SC technique (Table 1). Finally, after real- ment, which is in line with the optimized parameters. At Lot, the
izing the difference in geological formation and hydrological observed annual average water yield was 34% (of the precipitation)
processes, the model was further calibrated only by splitting the during calibration and 32% during the validation period. The cor-
groundwater-related parameters, but keeping lumped (same) values responding simulated values were 34 and 37%, respectively. The
for the surface runoff-related parameters (e.g., CN2, SOL_AWC, model systematically overestimated the total water yield for the val-
ESCO, CH_N2). The parameters that depend on land use, soil, and idation period (Table 2). The simulated and observed SR showed
slope types were adjusted on the basis of relative (percent) change to a good agreement, with low PBIAS for both the calibration and
their original values. Hence, the term lumped only refers to the same validation periods (Table 2). In contrast to the calibration period,
relative (magnitude) change over the catchment. As compared to it was found that the model slightly overestimated the surface run-
SMSC, the SMSC-LSRP technique gave similar values except for off during the validation (Table 2). The latter may be explained by
CN2, GW_DELAY, and GWQMN (Table 1). In comparison with the use of relatively wet years (310 mm mean annual flow) for the
the SC and SMSC techniques, although the CN2 values of Sub- calibration period, whereas the validation period used seemingly
catchment II were somehow decreased, the values were noticeably dry years (258 mm mean annual flow). Compared with the SC and
increased for Subcatchment III. This clearly resulted in overestima- SMSC techniques, similar findings were discerned for the SMSC-
tion of peak flows at Lot, whereas a significant underestimation was LSRP technique in the downstream part of the catchment. How-
clearly noticed at Tubize (see daily streamflow section). ever, the latter technique substantially underestimated the surface
In general, these findings highlight that the eastern part of the runoff at Tubize, but base flow was considerably overestimated
basin (Subcatchment III) should get low CN2 value because of the (Table 2). Lastly, during the calibration period, the evapotranspira-
high infiltration and the recharge processes that can result in low tion (ET) of Subcatchment II was somehow lower than the ET of
surface runoff processes. Consequently, splitting the surface runoff Subcatchment III. Because the climatic condition was similar, this
related parameters is essential for the basin. The latter is indeed was most likely because of land use type because more forest cover
needed because there could be interactions between the ground- was observed in the eastern part of the catchment [Fig. 2(a)].
water and surface runoff processes. Overall, the findings suggested that multisite calibration im-
proved the representation of water balance components. At the same
time, the discrepancy between observations and simulations was
Annual Water Balance
noticeably decreased for the SC and SMSC techniques (Table 2).
The simulated annual average water balance components of the This indicated that more information about the observations is
catchment, such as the total water yield (WYLD), the surface runoff needed to well represent and capture the hydrological processes

© ASCE 05016036-8 J. Hydrol. Eng.

J. Hydrol. Eng., 05016036

 Table 2. Observed and Simulated Water Balance Components (in Millimeters)
Method Location Period Type Precipitation WYLD BF SR ET
Single site Lot Calibration Observed 901 310 234 76 470
Simulated 341 242 99
PBIAS (%) 10 4 31
Validation Observed 804 258 199 60 432
Simulated 284 198 86
PBIAS (%) 21 11 52
Sequential Tubize Calibration Observed 887 317 211 106 451
Simulated 320 223 96
PBIAS (%) 1 6 −9
Validation Observed 538 207 137 70 178
Simulated 248 175 74
Downloaded from ascelibrary.org by University of Hong Kong on 10/02/16. Copyright ASCE. For personal use only; all rights reserved.

PBIAS (%) 20 28 5
Lot Calibration Observed 901 310 234 76 467
Simulated 325 241 86
PBIAS (%) 5 3 13
Validation Observed 804 258 199 60 428
Simulated 293 221 74
PBIAS (%) 14 11 25
Simultaneous multisite (SMS) Tubize Calibration Observed 887 317 211 106 452
Simulated 320 219 101
PBIAS (%) 1 4 −4
Validation Observed 538 207 137 70 179
Simulated 250 173 78
PBIAS (%) 21 27 10
Lot Calibration Observed 901 310 234 76 467
Simulated 307 239 68
PBIAS (%) −1 2 −11
Validation Observed 804 258 199 60 429
Simulated 301 238 63
PBIAS (%) 17 20 6
SMS-lumped runoff parameters Tubize Calibration Observed 887 317 211 106 455
Simulated 311 252 59
PBIAS (%) −2 19 −44
Validation Observed 538 207 137 70 178
Simulated 257 215 42
PBIAS (%) 24 57 −40
Lot Calibration Observed 901 310 234 76 468
Simulated 321 248 74
PBIAS (%) 4 6 −2
Validation Observed 804 258 199 60 428
Simulated 310 242 68
PBIAS (%) 20 22 15
Note: BF = base flow; ET = evapotranspiration; PBIAS = percent bias; SR = surface runoff; WYLD = total water yield; the bold signifies when the PBIAS
values are above or below 25% in one of the water balance components.

variability of the basin. Also, the SSC method showed that despite streamflow simulations. Similarly, two years of daily streamflow
noticeable underestimation or overestimation in one component of hydrographs were presented only for the period of 1996–1997
the total WYLD, the discrepancy of WYLD was low because of the during the validation period.
compensation effects of its components. Such characterizations Although the daily dynamics of the streamflows at Tubize were
may have serious implications on sediment and nutrients transport reproduced well by the SSC technique, the model systematically
modeling, including the time of travel; therefore, each component underestimated the peak flows for both the calibration and valida-
of annual flow needs to be evaluated separately at the time of model tion periods [Figs. 5(c and d)]. Also, the daily statistical metrics
calibration. (Table 3) revealed that the performance of the model was satisfac-
tory (Moriasi et al. 2007). However, during the validation period
(1988–1997), the NSE was lowered to 0.39, indicating an unsat-
Daily Streamflow isfactory model performance. In contrast to Tubize, the peak
For a clear illustration of daily streamflow temporal evolution, only streamflows at Lot were overestimated, despite that the CN2 and
two years of daily streamflow hydrographs were shown during the SURLAG values were lowered [Figs. 6(c and d)]. Additionally,
calibration, although the model was calibrated for the period of the high values of GWQMN and GW_DELAY were not able to
1998–2008. The selected two years were 2001–2002 because reduce the simulated peak flows of Lot (Table 1). Furthermore,
they clearly demonstrated the good and poor fit part of stream- the ALPHA_BF was set to its maximum value to increase the
flow hydrographs. Thus, this period could provide more informa- recession curve, but this was not able to resolve the problem. More
tion about the effect of all the applied calibration techniques on importantly, in contrast to the results of Tubize, the simulated

© ASCE 05016036-9 J. Hydrol. Eng.

J. Hydrol. Eng., 05016036

Downloaded from ascelibrary.org by University of Hong Kong on 10/02/16. Copyright ASCE. For personal use only; all rights reserved.

Fig. 5. (a and b) Areal average daily precipitation for the selected calibration (2001–2002) and validation (1996–1997) periods at Tubize and stream-
flows for (c and d) single site, (e and f) sequential, (g and h) simultaneous multisite, and (I and j) simultaneous multisite with lumped surface runoff
parameters techniques

peak flows at Lot were higher than observations during winter sea- When the SC technique was applied, most of the daily stream-
son, whereas the low flows of summer season were consistently flows dynamics and variability at Tubize were captured well by
underestimated. Because of such contrasting characterizations, the model, except for some peak flow events of 2001 and 2002
matching the peak flows at Lot was at the expense of further [Fig. 5(e)]. The model was not able to simulate peak flows greater
underestimating the low flows and vice versa. Because the find- than 30 m3 s−1 for those events, even though high CN2 values were
ings at Tubize and Lot stations showed contrasting situations, used for Subcatchment II (Table 1). The summary of the goodness-
it highlighted that the relationship between the hydrological of-fit statistics for the SC technique at Tubize showed a very good
processes and inputs for Subcatchments II and III might not be model performance during the calibration period (Table 3), indicat-
similar. ing significant improvement compared with the results of SSC

© ASCE 05016036-10 J. Hydrol. Eng.

J. Hydrol. Eng., 05016036

 Table 3. Goodness-of-Fit Statistics for the Daily Streamflow at Tubize and Lot Using the Four Calibration Techniques
Method Location Period NSE RSR PBIAS (%)
Single site (SSC) Tubize Validation 1 (1998–2008) 0.62 0.61 16
Validation 2 (1988–1997) 0.39 0.78 33
Lot Calibration (1998–2008) 0.51 0.70 10
Validation (1988–1997) 0.52 0.75 21
Sequential (SC) Tubize Calibration (1998–2008) 0.80 0.45 9
Validation (1988–1997) 0.42 0.82 27
Lot Calibration (1998–2008) 0.81 0.44 3
Validation (1988–1997) 0.82 0.50 19
Simultaneous multisite (SMSC) Tubize Calibration (1998–2008) 0.80 0.46 9
Validation (1988–1997) 0.38 0.79 24
Downloaded from ascelibrary.org by University of Hong Kong on 10/02/16. Copyright ASCE. For personal use only; all rights reserved.

Lot Calibration (1998–2008) 0.81 0.46 −1

Validation (1988–1997) 0.83 0.46 16
Simultaneous multisite-lumped surface Tubize Calibration (1998–2008) 0.71 0.54 6
runoff parameters (SMSC-LSRP) Validation (1988–1997) 0.26 0.86 23
Lot Calibration (1998–2008) 0.77 0.48 4
Validation (1988–1997) 0.77 0.55 20
Note: NSE = Nash-Sutcliffe efficiency; PBIAS = percent bias; RSR = root mean squared error (RMSE) to observation standard deviation ratio.

technique. For instance, the NSE value of Lot, which was 0.51 for Basin (Leta et al. 2015). In addition, researchers suggested that
the SSC technique, increased to 0.81 for SC (approximately 55% errors in gauging stations can contribute to less matching of hydro-
increment). This showed that incorporating more observations can graph and, thus, model performance (Leta et al. 2015; Rossi et al.
capture the heterogeneity of the catchment and improve the perfor- 2009; Shrestha et al. 2013a).
mance of the model. When the SMSC technique was used, SWAT Considering the SMSC-LSRP technique, the simulated base
adequately tracked the trends of the streamflow hydrograph at flow component was highly overestimated at Tubize, whereas the
Tubize [Figs. 5(g and h)]. However, as was the case for the SC surface runoff was clearly underestimated (Table 2). Although
technique, the model showed a tendency to underestimate some the performance of the SMSC-LSRP technique was better than the
peak flows during calibration. Conversely, model overestimation SSC technique, the peak flows of the downstream part (Lot) was
was clearly noticed during the validation period [Figs. 5(g and h)]. overestimated remarkably [Fig. 6(i)]. Conversely, the peak flows
While the low to medium flows were accurately represented, the of Subcatchment II at Tubize were underestimated substantially
model was not able to generate peak flows that exceeded 30 m3 s−1 , [Fig. 5(i)]. This clearly indicated that the surface runoff related
especially for the peak events of 2001–2002, although high CN2 parameters, such as CN2, need to be increased for Subcatchment
values were used (Table 1). The reason why the model was under- II, signifying the importance of splitting surface runoff-related
estimating the peak flow events of this period was not clear. How- parameter values from Subcatchment III. This also meant that simi-
ever, although there was overflow to the canal at Lembeek, on the larities in land use and soil characteristics do not necessarily mean
same days, only half of the peak flows of Tubize were recorded at that similar values should be taken for the surface runoff parame-
Lot. This could point to problems with the stage-discharge relation- ters, even in the case of neighboring catchments.
ship, particularly for peak flows at Tubize, since the gauging station Overall, the different calibration techniques applied in this
is located in a flood-prone area. study not only impacted the water balance components, but also
Compared with Tubize, the SMSC technique well represented potentially affected the shape of streamflow hydrographs (Fig. 7).
the peak flows of Lot. The low flows of Lot were also simulated Additionally, the daily streamflow simulation results stressed the im-
well for some years, except for the period of 2001–2002. For the portance of collecting more detailed and spatially well-distributed
latter period, the model had a tendency to overestimate the low hydrological data to perform a multisite calibration. This also would
flows primarily during the summer season [Fig. 6(g)]. A possible help to capture the catchment spatial variability that primarily gov-
explanation might be that the section of this station was not stable erns the basin’s hydrological processes. More importantly, the
because of sedimentation and bed scouring, or dredging activities simulation results of the different calibration techniques may have
that occurred near the monitoring site. This might play a significant serious implications for water resources management decisions.
role in stage-discharge relationship. When calibrating the model, On the basis of the simulated annual water balance, daily
if the groundwater flow was further decreased to match the period streamflows, estimated parameter values, and the calculated statis-
of 2001–2002, the simulated groundwater flow for the other tical metrics, the SC and SMSC techniques outperformed the SSC
calibration period became worse. Thus, it was suspected that the and SMSC-LSRP techniques. Even though the performances of the
measured flows of 2001–2002 showed errors because the flows SC and SMSC methods were similar, the SMSC technique should
during this period were matched well at Tubize [Fig. 5(g)]. Overall, be selected over the SC technique for the following reasons. First,
the performance of the SMSC technique for the daily streamflow it handles the entire spatial variability of the catchment by si-
simulation at both stations was very good, except for the validation multaneously assigning spatially different parameter values and
period at Tubize, which also was the case for the other techniques searching for their optimal values. SMSC also would allow use of
(Table 3). However, if the years 1994–1997 were considered, the multiple objective functions calculated at multiple stations simul-
NSE was noticeably increased to 0.61, showing a good model per- taneously, whereas the calibration data information is communi-
formance according to Moriasi et al. (2007). Hence, it was sus- cated among stations at once. In this way, the SMSC technique
pected that there were measurement errors for this period at Tubize, can search and find a set of optimal parameter values that can
which also was reported in the previous studies of the upper Zenne provide best fits at multiple stations. Second, the calibration results

© ASCE 05016036-11 J. Hydrol. Eng.

J. Hydrol. Eng., 05016036

Downloaded from ascelibrary.org by University of Hong Kong on 10/02/16. Copyright ASCE. For personal use only; all rights reserved.

Fig. 6. (a and b) Areal average daily precipitation for the selected calibration (2001–2002) and validation (1996–1997) periods at Lot and streamflows
for (c and d) single site, (e and f) sequential, (g and h) simultaneous multisite, and (i and j) simultaneous multisite with lumped surface runoff
parameters techniques.

of the two stations are interrelated in the sense that Subcatchment II to the calibrated parameter values of Subcatchment III. It is likely
feeds the Subcatchment III (Catchment I). This implies that the cal- that the calibrated parameter values of Subcatchment III could com-
ibration at Tubize can obviously influence the results at Lot. For pensate for such effect (fixed parameter values) because the down-
example, if one site calibration, such as SC technique, is used, this stream information (observations) is not communicated during the
technique does not consider how well the model simulates at the Subcatchment II calibration process. For such situations, simulta-
downstream, and therefore lacks information communication about neous, multisite calibration is more appropriate to reduce model
the downstream responses. As a consequence, fixing the parameters calibration bias and to introduce more constraints (calibration data)
of Subcatchment II to their optimal values before calibrating the on the calibration process (Migliaccio and Chaubey 2007). Third,
parameters of Subcatchment III would introduce additional bias the SMSC method gives the opportunity to calibrate all the SWAT

© ASCE 05016036-12 J. Hydrol. Eng.

J. Hydrol. Eng., 05016036

Downloaded from ascelibrary.org by University of Hong Kong on 10/02/16. Copyright ASCE. For personal use only; all rights reserved.

Fig. 7. Observed and simulated daily streamflow for the different calibration strategies at (a and c) Tubize and (b and d) Lot for the selected wet years
(2001–2002) and dry years (2003–2004)

parameters structured at three spatial scales (basin, subbasin, Ultimately, the overflow calibration results of the SMSC tech-
and HRU) at once, such that the basin scale parameters, such as nique were compared with a hydraulic model [the storm water
SURLAG, can be appropriately optimized at once. Otherwise, the management model (SWMM)] obtained from Shrestha (2013).
SC technique that had been used by other researchers for assessing The comparison showed a good agreement between both models,
the effect of anthropogenic engineered structures (e.g., roads, cul- despite a simplified representation of overflow that was used in the
verts, and sediment retention ponds) on hydrologic partitioning SWAT model (Leta 2013). The reader is referred to Leta (2013) for
could be used (Shope et al. 2014). the detailed results.
Finally, as an example of uncertainty quantifications that are in-
volved in model performance, when the different uncertainty sour-
Model Performance Rating
ces (e.g., parameter, rainfall, and streamflow) were considered, the
peak flows of 2001–2002, which were systematically underesti- The general performance of the SWAT model for daily streamflows
mated by SWAT at Tubize, were captured at 95% prediction uncer- simulation was rated on the basis of NSE, RSR, and PBIAS
tainty (Fig. 8). Further, the performance of the model was improved statistical metrics as suggested by Moriasi et al. (2007). The four
remarkably as the percent of observations bracketed at 95% predic- qualitative performance rating (very good, good, satisfactory, and
tion uncertainty interval (Abbaspour et al. 2007) was 97% with unsatisfactory) of Moriasi et al. (2007), and the corresponding
an average uncertainty band width (Yang et al. 2008) of 1.19. This threshold values were used for the previously named three stat-
suggested that accounting for the different sources of uncertainty istical metrics. However, to get a general performance rating of
has a paramount importance during model performance evaluation the SMSC technique, a performance weight of 4 for very good,
and also improves representation of observations. For the detailed 3 for good, 2 for satisfactory, and 1 for unsatisfactory were as-
decomposition of the different sources of uncertainty and treat- signed to each statistical metrics. Then, the average value of the
ments, refer to Leta et al. (2015). metrics’ weight was used to obtain the general performance rating.

© ASCE 05016036-13 J. Hydrol. Eng.

J. Hydrol. Eng., 05016036

Downloaded from ascelibrary.org by University of Hong Kong on 10/02/16. Copyright ASCE. For personal use only; all rights reserved.

Fig. 8. Total predictive uncertainty on daily streamflow of (a) 2000–2001 and (b) 2002–2003 at Tubize station

Table 4. General Performance Rating for the SWAT Model of the Zenne (2) SC, starting at the upstream station of Tubize and subsequently
River on the Basis of the SMSC moving to the downstream station at Lot; (3) SMSC, in which data
Station Period Time span NSE RSR PBIAS Rating of both stations were used in a single calibration; and (4) SMSC-
LSRP, similar to the SMSC, but lumped values were used for sur-
Tubize Calibration 1998–2008 0.80 0.46 −9 Very good
Validation 1988–1997 0.38 0.79 24 Unsatisfactory face runoff related parameters.
Lot Calibration 1998–2008 0.81 0.46 −1 Very good The different calibration methods indeed led to substantially dif-
Validation 1988–1997 0.83 0.46 16 Very good ferent model outputs in terms of water balance components, param-
eter values, and model performances. The SSC technique did not
Note: NSE = Nash-Sutcliffe efficiency; PBIAS = percent bias; RSR = root
represent well the water balance components and the shape of the
mean square error (RMSE) to observation standard deviation ratio.
streamflow hydrograph, whereas the SC and SMSC techniques sig-
nificantly improved these representations. The results indicated that
The results of the analysis are shown in Table 4 and shows that the a multisite calibration could significantly improve the representa-
model performance was generally rated as very good, except for the tion of observations in this heterogeneous catchment. For example,
validation period at Tubize (1988–1997), in which measurement the SMSC technique increased the Nash-Sutcliffe efficiency to
errors were suspected. about 0.8, as compared with the value of 0.5 for the SSC technique
at Lot. In addition, although the SC and SMSC techniques provided
Summary and Conclusions similar results, the SMSC should be selected over SC for hetero-
geneous catchment as it simultaneously manages the entire catch-
A SWAT model was built for the upper Zenne River Basin up to ment spatial variability by allocating different parameter values
Lot. For the calibration of the model, the upstream daily streamflow and allows data information communications among multiple sta-
data were available at the Lot and Tubize stations. The different tions in one calibration. At the same time, the SMSC technique gives
geology in the catchment means that the basin can be considered an opportunity to simultaneously calibrate all the SWAT parameters
as quite heterogeneous with regard to the hydrologic processes. that are structured at three spatial scales (basin, subbasin, and HRU).
A detailed multisite sensitivity analysis was performed to iden- Findings generally highlighted the importance of incorporating
tify the most important parameters for a subsequent calibration of more observations to capture the heterogeneity of the hydrological
the sensitive parameters. It was found that the top five sensitive processes. By doing so, this study was indeed able to relate the
parameters for the basin were the CN2, the CH_N2, the GWQMN, spatial variability of the SWAT parameter values to the catchment’s
the ALPHA_BF, and the RCHRG_DP. geological formation and the related hydrological processes.
In this study, four different model calibration techniques for Conversely, the study’s findings suggested the risks associated to
heterogeneous catchment were applied: (1) SSC, which only used an improvident regionalization of parameters, even for a so-called
the observed time series of the flows at the watershed outlet at Lot; physically-based model such as SWAT.

© ASCE 05016036-14 J. Hydrol. Eng.

J. Hydrol. Eng., 05016036

 In general, it is expected that the study’s findings provide some Duan, Q., Sorooshian, S., and Gupta, V. K. (1994). “Optimal use of the
insights to the SWAT model users in particular, and the distributed SCE-UA global optimization method for calibrating watershed
hydrologic models users in general, to determine how a multisite models.” J. Hydrol., 158(3–4), 265–284.
calibration technique outperforms the single-site calibration tech- Gassman, P. W., Reyes, M. R., Green, C. H., and Arnold, J. G. (2007). “The
nique, particularly in heterogeneous catchments. soil and water assessment tool: Historical development, applications,
and future research directions.” Trans. ASABE, 50(4), 1211–1250.
Hogue, T. S., Sorooshian, S., Gupta, H., Holz, A., and Braatz, D. (2000). “A
multistep automatic calibration scheme for river forecasting models.”
Acknowledgments J. Hydrometeorol., 1(6), 524–542.
Leta, O. T. (2013). “Catchment processes modeling, including the assess-
The authors would like to thank INNOVIRIS (Environment Im- ment of different sources of uncertainty, using the SWAT model: The
pulse Programme of the Brussels-Capital Region) for supporting river Zenne basin (Belgium) case study.” Ph.D. dissertation, Free Univ.
the GESZ research project. The authors also are indebted to all of Brussels (VUB), Brussels, Belgium, 245.
the institutes and administrations mentioned in this paper for pro- Leta, O. T., et al. (2015). “Assessment of the different sources of uncertainty
Downloaded from ascelibrary.org by University of Hong Kong on 10/02/16. Copyright ASCE. For personal use only; all rights reserved.

viding the required data. in a SWAT model of the River Senne (Belgium).” Environ. Modell.
Software, 68(0), 129–146.
Leta, O. T., Shrestha, N. K., de Fraine, B., van Griensven, A., and Bauwens,
W. (2014). “Integrated water quality modelling of the river Zenne
References
(Belgium) using OpenMI.” Advances in hydroinformatics: Springer hy-
Abbaspour, K. C., et al. (2007). “Modelling hydrology and water quality drogeology, P. Gourbesville, J. Cunge, and G. Caignaert, eds., Springer,
in the pre-alpine/alpine Thur watershed using SWAT.” J. Hydrol., Singapore, 259–274.
333(2–4), 413–430. Li, X., Weller, D. E., and Jordan, T. E. (2010). “Watershed model cal-
Ajami, N. K., Gupta, H., Wagener, T., and Sorooshian, S. (2004). ibration using multi-objective optimization and multi-site averaging.”
“Calibration of a semi-distributed hydrologic model for streamflow J. Hydrol., 380(3–4), 277–288.
estimation along a river system.” J. Hydrol., 298(1–4), 112–135. Li, Z., Liu, W.-Z., Zhang, X.-C., and Zheng, F.-L. (2009). “Impacts of land
Andrews, F. T., Croke, B. F. W., and Jakeman, A. J. (2011). “An open use change and climate variability on hydrology in an agricultural catch-
software environment for hydrological model assessment and develop- ment on the Loess Plateau of China.” J. Hydrol., 377(1–2), 35–42.
ment.” Environ. Modell. Software, 26(10), 1171–1185. Meyus, Y., Woldeamlak, S. T., Batelaan, O., and De Smedt, F. (2004).
ArcGIS [Computer software]. Esri, Redlands, CA. “Opbouw van een Vlaams groundwatervoedingsmodel: Deelrapport 2,
Arnold, J. G., et al. (2011). “Soil and water assessment tool. input/output totaal Vlaams groundwater model (VGM)-karteergebied en vlaande-
file documentation, version 2009.” Agrilife Blackland Research Center, ren.” Vrije Universiteit Brussel, Brussels, Belgium.
Temple, TX. Migliaccio, K., and Chaubey, I. (2008). “Spatial distributions and stochastic
Arnold, J. G., et al. (2012). “SWAT: Model use, calibration and validation.” parameter influences on SWAT flow and sediment predictions.” J.
Trans. ASABE, 55(4), 1491–1508. Hydrol. Eng., 10.1061/(ASCE)1084-0699(2008)13:4(258), 258–269.
Arnold, J. G., and Allen, P. M. (1999). “Automated methods for estimating Migliaccio, K. W., and Chaubey, I. (2007). “Comment on Cao W, Bowden
baseflow and ground water recharge from streamflow records.” J. Am. BW, Davie T, Fenemor A. 2006. ‘Multi-variable and multi-site calibra-
Water Resour. Assoc., 35(2), 411–424. tion and validation of SWAT in a large mountainous catchment with
Arnold, J. G., Allen, P. M., Volk, M., Williams, J. R., and Bosch, D. D. high spatial variability.’ Hydrological Processes 20(5), 1057–1073.”
(2010). “Assessment of different representations of spatial variability Hydrol. Processess, 21(23), 3226–3228.
on SWAT model performance.” Trans. ASABE, 53(5), 1433–1443. Moriasi, D. N., et al. (2007). “Model evaluation guidelines for systematic
Arnold, J. G., Srinivasan, R., Mukundan, R. S., and Williams, J. R. (1998). quantification of accuracy in watershed simulations.” Trans. ASABE,
“Large area hydrologic modeling and assessment. Part I: Model devel- 50(3), 885–900.
opment.” J. Am. Water Resour. Assoc., 34(1), 73–89. Mukundan, R., et al. (2013). “Suspended sediment source areas and future
Arnold, J. G., Williams, J. R., and Maidment, D. R. (1995). “Continuous- climate impact on soil erosion and sediment yield in a New York City
time water and sediment-routing model for large basins.” J. Hydraul. water supply watershed, USA.” Geomorphology, 183, 110–119.
Eng., 10.1061/(ASCE)0733-9429(1995)121:2(171), 171–183. Narasimhan, B., Srinivasan, R., Bednarz, S. T., Ernst, M. R., and Allen,
Bae, D.-H., Jung, I.-W., and Lettenmaier, D. P. (2011). “Hydrologic uncer-
P. M. (2010). “A comprehensive modeling approach for reservoir water
tainties in climate change from IPCC AR4 GCM simulations of the
quality assessment and management due to point and nonpoint source
Chungju Basin, Korea.” J. Hydrol., 401(1–2), 90–105.
pollution.” Trans. ASABE, 53(5), 1605–1617.
Bannwarth, M. A., et al. (2015). “Simulation of stream flow components in
Nash, J. E., and Sutcliffe, J. V. (1970). “River flow forecasting through
a mountainous catchment in northern Thailand with SWAT, using
conceptual models. Part I—A discussion of principles.” J. Hydrol.,
the ANSELM calibration approach.” Hydrol. Processess, 29(6)
10(3), 282–290.
1340–1352.
Beven, K. (2006). “A manifesto for the equifinality thesis.” J. Hydrol., Neitsch, S. L., Arnold, J. G., Kiniry, J. R., and Williams, J. R. (2011). “Soil
320(1–2), 18–36. adn water assessment tool: Theoretical documentation, Version 2009.”
Cao, W., Davie, T., Fenemor, A., and Bowden, W. B. (2007). “Reply to Grassland, Soil and Water Research Laboratory, Agricultural Rsearch
comment on Cao W, Bowden BW, Davie T, and Fenemor A. Service Blackland Research Center-Texas AgriLife Research, Texas
(2006). ‘Multi-variable and multi-site calibration and validation of Water Resources Institute, College Station, TX.
SWAT in a large mountainous catchment with high spatial variability’. Niu, J., Shen, C., Li, S.-G., and Phanikumar, M. S. (2014). “Quantifying
Hydrological Processes 20(5), 1057–1073.” Hydrol. Processess, storage changes in regional Great Lakes watersheds using a coupled
21(23), 3229–3230. subsurface-land surface process model and GRACE, MODIS products.”
Chow, V. T. (1959). Open channel hydraulics, McGraw-Hill, New York. Water Resour. Res., 50(9), 7359–7377.
Cibin, R., Sudheer, K. P., and Chaubey, I. (2010). “Sensitivity and identi- Nossent, J. (2012). “Sensitivity and uncertainty analysis in view of
fiability of stream flow generation parameters of the SWAT model.” the parameter estimation of a SWAT model of the River Kleine
Hydrol. Processess, 24(9), 1133–1148. Nete, Belgium.” Ph.D. dissertation, Vrije Universiteit Brussel (VUB),
Debele, B., Srinivasan, R., and Parlange, J. Y. (2008). “Coupling upland Brussels, Belgium, 462.
watershed and downstream waterbody hydrodynamic and water qual- Nossent, J., and Bauwens, W. (2012). “Multi-variable sensitivity and
ity models (SWAT and CE-QUAL-W2) for better water resources identifiability analysis for a complex environmental model in view
management in complex river basins.” Environ. Model. Assess., 13(1), of integrated water quantity and water quality modeling.” Water Sci.
135–153. Technol, 65(3), 539–49.

© ASCE 05016036-15 J. Hydrol. Eng.

J. Hydrol. Eng., 05016036

 Pappenberger, F., Beven, K. J., Ratto, M., and Matgen, P. (2008). “Multi- Zenne (Belgium).” Ph.D. thesis, Vrije Universiteit Brussel (VUB),
method global sensitivity analysis of flood inundation models.” Brussels, Belgium, 265.
Adv. Water Resour., 31(1), 1–14. Shrestha, N. K., Leta, O. T., De Fraine, B., van Griensven, A., and
Peeters, L. (2010). “Groundwater and geochemical modelling of the un- Bauwens, W. (2013b). “OpenMI-based integrated sediment transport
confined Brussels aquifer, Belgium.” Ph.D. dissertation, KU Leuven, modelling of the river Zenne, Belgium.” Environ. Modell. Software,
Leuven, Belgium. 47, 193–206.
Piniewski, M., and Okruszko, T. (2011). “Multi-site calibration and vali- Singh, J., Knapp, H. V., Arnold, J. G., and Demissie, M. (2005). “Hydro-
dation of the hydrological component of SWAT in a large lowland logical modeling of the Iroquois river watershed using HSPF and
catchment.” Modelling of hydrological processes in the narew catch- SWAT.” J. Am. Water Resour. Assoc., 41(2), 343–360.
ment, D. Świątek and T. Okruszko, eds., Springer, Berlin, 15–41.
USDA-SCS (U.S. Department of Agriculture-Soil Conservation
Ravalico, J. K., Dandy, G. C., and Maier, H. R. (2010). “Management op-
Service). (1986). “Urban hydrology for small watersheds.” USDA,
tion rank equivalence (MORE)—A new method of sensitivity analysis
Washington, DC.
for decision-making.” Environ. Modell. Software, 25(2), 171–181.
Rossi, C. G., et al. (2009). “Hydrological evaluation of the lower Mekong van Griensven, A. (2002). “Developments towards integrated water qual-
river basin with the soil and water assessment tool model.” Int. Agric. ity modelling for river basins.” Ph.D. dissertation, Vrije Universiteit
Downloaded from ascelibrary.org by University of Hong Kong on 10/02/16. Copyright ASCE. For personal use only; all rights reserved.

Eng. J., 18(1–2), 1–13. Brussel (VUB), Brussels, Belgium, 280.

Santhi, C., et al. (2001). “Validation of the SWAT model on a large river van Griensven, A., et al. (2006). “A global sensitivity analysis tool for the
basin with point and nonpoint sources.” J. Am. Water Resour. Assoc., parameters of multi-variable catchment models.” J. Hydrol., 324(1–4),
37(5), 1169–1188. 10–23.
Santhi, C., Kannan, N., Arnold, J. G., and Di Luzio, M. (2008). “Spatial van Griensven, A., Francos, A., and Bauwens, W. (2002). “Sensitivity
calibration and temporal validation of flow for regional scale hydrologic analysis and auto-calibration of an integral dynamic model for river
modeling.” J. Am. Water Resour. Assoc., 44(4), 829–846. water quality.” Water Sci. Technol., 45(9), 325–332.
Santhi, C., Srinivasan, R., Arnold, J. G., and Williams, J. R. (2006). “A Van Liew, M. W., Arnold, J. G., and Bosch, D. D. (2005). “Problems and
modeling approach to evaluate the impacts of water quality manage- potential of autocalibrating a hydrologic model.” Trans. ASABE, 48(3),
ment plans implemented in a watershed in Texas.” Environ. Modell. 1025–1040.
Software, 21(8), 1141–1157. Vansteenkiste, T., et al. (2013). “Climate change impact on river flows
Schilling, K. E., Jha, M. K., Zhang, Y.-K., Gassman, P. W., and Wolter, and catchment hydrology: A comparison of two spatially distributed
C. F. (2008). “Impact of land use and land cover change on the water models.” Hydrol. Processess, 27(25) 3649–3662.
balance of a large agricultural watershed: Historical effects and future Williams, J. R. (1969). Flood routing with variable travel time or vari-
directions.” Water Resour. Res., 44(7), W00A09. able storage coefficients. Trans. Am. Soc. Agric. Biol. Eng., 12(1),
Shi, P., et al. (2013). “Application of a SWAT model for hydrological
100–103.
modeling in the Xixian Watershed, China.” J. Hydrol. Eng., 10.1061/
Yang, J., Reichert, P., Abbaspour, K. C., Xia, J., and Yang, H. (2008).
(ASCE)HE.1943-5584.0000578, 1522–1529.
“Comparing uncertainty analysis techniques for a SWAT application
Shope, C. L., et al. (2014). “Using the SWAT model to improve process
descriptions and define hydrologic partitioning in South Korea.” to the Chaohe basin in China.” J. Hydrol., 358(1–2), 1–23.
Hydrol. Earth Syst. Sci., 18(2), 539–557. Zhang, X., Srinivasan, R., and Van Liew, M. (2008). “Multi-site calibration
Shrestha, B., et al. (2013a). “Impact of climate change on sediment yield in of the SWAT model for hydrologic modeling.” Trans. ASABE, 51(6),
the Mekong River basin: A case study of the Nam Ou basin, Lao PDR.” 2039–2049.
Hydrol. Earth Syst. Sci., 17(1), 1–20. Zomlot, Z., Verbeiren, B., Huysmans, M., and Batelaan, O. (2015). “Spatial
Shrestha, N. K. (2013). “Development of OpenMI compliant simulators distribution of groundwater recharge and base flow: Assessment of
for integrated water quality management, with application on the river controlling factors.” J. Hydrol. Reg. Stud., 4(Part B), 349–368.

© ASCE 05016036-16 J. Hydrol. Eng.

J. Hydrol. Eng., 05016036