Environmental Modelling & Software 66 (2015) 57e68

Contents lists available at ScienceDirect

Environmental Modelling & Software

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

Assessment of the internal dynamics of the Australian Water Balance

Model under different calibration regimes
Li Li a, 1, Martin F. Lambert a, *, Holger R. Maier a, 2, Daniel Partington b, 3,
Craig T. Simmons b, 4
a
School of Civil, Environmental and Mining Engineering, The University of Adelaide, North Terrace, Adelaide, South Australia 5005, Australia
b
National Centre for Groundwater Research and Training and School of the Environment, Flinders University, GPO Box 2100, Adelaide,
South Australia 5001, Australia

a r t i c l e i n f o a b s t r a c t

Article history: Conceptual rainfall runoff models are used extensively in practice, as they provide a good balance be-
Received 1 August 2014 tween transparency and computational and data requirements. However, the degree to which they are
Received in revised form able to represent underlying physical processes is poorly understood. This is because the performance of
14 December 2014
such models is generally assessed based on their ability to match total streamow, rather than
Accepted 15 December 2014
Available online 13 January 2015
component processes. In this paper, the ability of the Australian Water Balance Model (AWBM) to
represent baseow and quickow is assessed for 66 synthetic catchments with different physical char-
acteristics and hydrological inputs under seven calibration regimes utilising a shufed complex evolution
Keywords:
Australian Water Balance Model
(SCE) algorithm. The observed total-, base- and quick-ow hydrographs for these catchments are
Calibration generated using HydroGeoSphere. The results indicate that while AWBM is generally able to match total
Baseow streamow well, the same does not apply to baseow and quickow, suggesting that these processes are
Regression models not represented well by AWBM.
HydroGeoSphere 2014 Elsevier Ltd. All rights reserved.
Virtual laboratory
Integrated surface water/groundwater
model

1. Introduction Black box modelling approaches, such as articial neural net-

Modelling approaches for estimating runoff from rainfall and
evapotranspiration (ET) can be traditionally classied into three
main groups: black box models, physical process based models and
conceptual rainfall runoff (CRR) models (Beven, 2005). While all of
these approaches have been shown to be able to predict total
streamow successfully, the degree to which they are able to
represent underlying streamow generating mechanisms is highly
variable (Chen and Adams, 2006; Ferket et al., 2010; Refsgaard and
Knudsen, 1996).
* Corresponding author. Tel.: 61 (0)8 8313 4320; fax: 61 (0)8 83034359.
E-mail addresses: l.li@alumni.adelaide.edu.au (L. Li), martin.lambert@adelaide.
edu.au (M.F. Lambert), holger.maier@adelaide.edu.au (H.R. Maier), daniel.
partington@inders.edu.au (D. Partington), craig.simmons@inders.edu.au
(C.T. Simmons).
1
Tel.: 61 0406745484; fax: 61 (0)8 83034359.
2 obtaining the data required (e.g. due to limitations of existing
Tel.: 61 (0)8 8313 4139; fax: 61 (0)8 83034359.
3
Tel.: 61 (0)8 8201 2648; fax: 61 8 8201 7906.
4
Tel.: 61 (0)8 8201 5509; fax: 61 8 8201 7906.
works (Abrahart et al., 2012; Maier et al., 2010; Wu et al., 2014), are
at one end of the spectrum, while physically based approaches are
at the other end. Black box models produce streamow outputs
solely as a function of their inputs and transfer characteristics,
without any knowledge or understanding of the underlying phys-
ical processes. However, they are generally computationally ef-
cient and can be developed using limited data. Physically based
approaches attempt to simulate the detailed mechanisms of the
component physical processes within the hydrologic cycle using
well-established physical laws, with numerical solutions of the
mathematical representation of these processes (Jayatilaka et al.,
1998). Such approaches include fully integrated surface water/
groundwater (SW/GW) models, such as InHM (VanderKwaak and
Loague, 2001), MODHMS (HydroGeoLogic, 2000), HydroGeo-
Sphere (HGS) (Therrien et al., 2009) and SHE (Abbott et al., 1986).
However, there are problems with the application of these models
in practice due to the difculties and expense associated with
instrumentation and intrinsic uncertainty in measurements), as
well as their high computational demands.

http://dx.doi.org/10.1016/j.envsoft.2014.12.015
1364-8152/ 2014 Elsevier Ltd. All rights reserved.
58 L. Li et al. / Environmental Modelling & Software 66 (2015) 57e68
CRR models represent a compromise between the high data and
computational requirements of physical process based models and
the lack of transparency of black-box models. They are more
computationally efcient and less data intensive than process
based models, as they do not attempt to represent all physical
processes explicitly. However, they are more transparent than
black-box models, as they represent the assumed underlying
physical processes in a conceptual manner, generally in the form of
a number of interconnected storages that are linked with empirical
mathematical equations to conceptualise the movement of water
into, between the storages of and out of a catchment. Many
different CRR models have been proposed in the literature, such as
the Australian Water Balance Model (AWBM) (Boughton, 1993,
2004), the Soil Moisture and Accounting Model (SMAR) (Tuteja
and Cunnane, 1999), SIMHYD (Chiew et al., 2002) and GR4J
(Oudin et al., 2005), for example. These models have been able to
capture enough of the dynamics of rainfall runoff simulation time
series to be useful in water resources assessments.
Among the different rainfallerunoff modelling approaches
mentioned above, CRR models are the most widely utilised in
practice, due to their relatively simple structure, small number of
parameters and production of generally acceptable results. A
common feature of CRR models is that some of their model pa-
rameters have limited physical interpretation (Delleur, 1982;
Troutman, 1985), due to the fact that many complex catchment
physical processes are lumped together. In addition, although some
of the CRR model parameters, representing the physical properties
of the catchment (e.g. catchment area, surface slope), are usually
measurable, there are some physical parameters, such as hydraulic
conductivity and porosity, which are measurable in theory, but
difcult to measure in practice. Therefore, such CRR model pa-
rameters are generally estimated by calibration, by comparing the
modelled total streamow time series with the corresponding
observed data until an acceptable t to the objective function, or an
acceptable trade-off between objective functions in cases where
multi-objective optimisation is used (e.g. Ahmadi et al., 2014; Gibbs
et al., 2012), has been obtained. Signicant research effort has been
directed towards obtaining a well-dened optimal parameter set,
including local-type direct search optimisation methods and
globally based optimisation methods (Duan et al., 1992). However,
because CRR models are usually calibrated using only observed
total streamow time series, while internally they calculate a
number of additional states and uxes, such as baseow and
quickow, there may be many combinations of parameter values
that give similar objective function values. This phenomenon is
called equinality (Beven, 1993), and is caused by problems such
as over-parameterisation of models, data limitations and structural
faults in the model. As a result, even though the structure of CRR
models is based on a conceptual representation of underlying
physical processes, how well these processes are represented by
calibrated models is generally unknown, as a good match to total
streamow does not necessarily mean that the component pro-
cesses are modelled accurately. For example, similar total stream-
ow time series can be obtained with very different combinations
of baseow and quickow, without consideration of the appropri-
ateness of their quantities and dynamics.
While there have been many studies comparing the perfor-
mance of CRR models using total streamow time series (Ferket
et al., 2010; Knapp et al., 1991; Post et al., 2007; Ranatunga et al.,
2008), very few attempts have been made to use baseow or
quickow estimates for CRR model internal dynamic performance
assessment, due to the difculty of accurately measuring baseow
or quickow in the eld (Dukic, 2006; McCallum et al., 2010).
Recently, Ferket et al. (2010) use baseow estimated from a
physically-based digital lter (Furey and Gupta, 2001) to validate
the internal dynamics of two CRR models (HBV and PDM) for a
subcatchment of the Dender catchment in Belgium. As part of the
study, two optimisation algorithms (SCE-UA and MWARPE) are
used to calibrate the models by matching total streamow to ob-
servations. They conclude that no clear picture emerges of which
model produces the best results of simulating total streamow, but
that the MWARPE calibration algorithm and the HBV model lead to
the best baseow estimates, giving the best internal model dy-
namics, at least when compared with the results obtained using the
Furey and Gupta lter (Furey and Gupta, 2001).
This study builds on the research by Ferket et al. (2010) by
assessing (i) how well the Australian Water Balance Model, which is
a commonly used CRR, is able to represent total-, base- and quick-
ow for 66 synthetic catchments with different catchment char-
acteristics and hydrological inputs and (ii) the impact of seven
different calibration regimes that take internal model dynamics
into account in different ways on the accuracy of total-, base- and
quick-ow hydrograph prediction. While the methodology is
illustrated for a particular case study, its generic nature means it
could easily be adapted and applied to other CRR models around
the world. The remainder of this paper is organised as follows. The
methodology is given in Section 2, followed by the results and
discussion in Section 3 and summary and conclusions in Section 4.
2. Methodology
The underlying premise of the proposed methodology for assessing the internal
dynamics of CRR models is that fully integrated SW/GW models can be used to
obtain reasonably accurate estimates of actual total streamow, quickow and
baseow (see Partington et al., 2012; Li et al., 2013, 2014), thereby providing a
benchmark against which the internal dynamics of CRR models can be assessed (e.g.
whether ow components that make up total streamow are predicted accurately).
This is a reasonable assumption, as fully integrated SW/GW models provide a
rigorous representation of the underlying physical processes of hydrologic systems
(Brookeld et al., 2009; Furman, 2008; Partington et al., 2012; Sulis et al., 2010;
Therrien and Sudicky, 1996). They typically represent 3D variably saturated sub-
surface ow with the Richards' equations, and 1D and 2D surface ow with the
diffusion wave approximation to the St. Venant equations. A unique feature is that
such models can simulate the partitioning of rainfall into different components,
including overland ow, streamow, evaporation, inltration and recharge, as well
as subsurface discharge to surface water features (e.g. lakes and streams), in a
physically realistic fashion (Therrien et al., 2009). All of the governing ow equations
implemented by fully integrated SW/GW models are solved simultaneously to
obtain total streamow, baseow and quickow, making them ideal candidates for
assessing the internal dynamics of CRR models.
While it is acknowledged that fully integrated SW/GW models are in themselves
an approximation of the actual processes in real catchments, they provide the best
means of quantifying the absolute volume of the ow components (e.g. baseow
and quickow) currently available (see Partington et al., 2012; Li et al., 2013, 2014).
In addition, they can be used to obtain estimates of different ow components for
catchments with different characteristics (see Partington et al., 2013). Therefore they
are able to provide the rst step towards being able to assess the internal dynamics
of CRR models under a range of physical conditions in a controlled manner.
The steps in the methodology adopted for assessing the internal dynamic per-
formance of AWBM are given in Fig. 1. As shown, initially synthetic total streamow
(qobs obs obs
T ), baseow (qB ) and quickow (qQ ) hydrographs are generated using a fully
integrated SW/GW model for a number of catchments with different physical
properties and hydrological inputs in order to ensure the results are as generic as
possible. Next, AWBMs are developed for the same catchments by using the same
hydrological inputs, but different calibration methods. Finally, the performance of
the AWBMs calibrated using the different methods is compared in terms of the
ability to predict total-, base- and quick-ow hydrographs accurately, thereby
providing a means of assessing the performance of the internal dynamics of AWBM
(e.g. whether ow components that make up total streamow are predicted accu-
rately). Details of each step in the methodology are given in subsequent sections. It
should be noted that while the AWBM is used as the CRR model in this study, the
same approach should also be used to test the internal dynamic performance of
other CRR models.
2.1. Catchment characteristics and hydrological inputs
The 66 synthetic catchments with different physical characteristics and hydro-
logical inputs developed by Li et al. (2014) are used. These catchments have drainage
areas ranging from 6 to 192 km2 and are loosely based on a benchmarked integrated
surfaceesubsurface hydrology problem, the V-catchment test case, as shown in
 L. Li et al. / Environmental Modelling & Software 66 (2015) 57e68 59

Fig. 1. Schematic representation of steps in the proposed methodology.

Fig. 2 (Panday and Huyakorn, 2004). A detailed description of these catchments can 2.2. Fully integrated SW/GW model
be found in Li et al. (2014), thus only a brief overview is provided here. Different
As stated above, fully integrated SW/GW models are used for assessing the in-
physical catchment characteristics are represented using seven variables: catchment
ternal dynamic performance of AWBM under a range of calibration approaches. This
area (A), catchment hill slope (S1, which is perpendicular to the channel), catchment
is because they provide the best possible approximation to the physical processes of
channel slope (S2, which is parallel to the channel), catchment aspect ratio (AR), and
water ow within catchments and can therefore be used as an approximation to
soil type, which includes Ks and van Genuchten parameters a and b. The hydrological
such processes subject to a variety of physical characteristics and forcings. In this
inputs are represented using the ratio of daily rainfall to ET from ve Australian cities
study, Hydrogeosphere (HGS) is used as the fully integrated SW/GW model to model
(Li et al., 2014), which are obtained from the Australian Bureau of Meteorology
the 66 synthetic catchments' response to rainfall and ET inputs under different
National Climate Centre. Details of the different values of the physical catchment
catchment characteristics and hydrological inputs, which are assumed to represent
characteristics considered are given in Table 1 and the characteristics of the hy-
the true catchment physical processes. This is because HGS can be used to simulate
drological inputs used are given in Table 2. Li et al. (2014) used Latin Hypercube
hydrological processes within catchments in a physically based manner (Therrien
Sampling (LHS) to generate the 66 catchments with different physical catchment
et al., 2009). HGS has been applied successfully to various studies, such as the
characteristics and hydrological inputs, which is also used in this study by sampling
comparison of baseow estimation methods (Li et al., 2013; Partington et al., 2011,
from these catchment characteristics and hydrological inputs.
2012), SW/GW disconnection problems (Banks et al., 2011; Brunner et al., 2009),

Fig. 2. Schematic representation of tilted V-catchment ow problem (adopted from Panday and Huyakorn, 2004).
60 L. Li et al. / Environmental Modelling & Software 66 (2015) 57e68

Table 1
Catchment characteristics considered (adopted from Li et al., 2014).

Catchment characteristic Unit Explanation Values considered

Ks m/s Saturated hydraulic conductivity 2.44  1005, 3.99  1005, 1.12  1004, 2.11  1004, 9.70  1004
a e van Genuchten parameter 0.572, 3.370, 6.161, 8.955, 11.75
b e van Genuchten parameter 1.32, 1.556, 1.793, 2.029, 2.270
A km2 Catchment area 6, 48, 80, 120, 192
S1 e Hill slope (perpendicular to the channel) 0.005, 0.008, 0.012, 0.016, 0.02
S2 e Channel slope (along the channel) 0.0025, 0.004, 0.006, 0.008, 0.01
AR e Ratio of catchment width to length (x/y) 0.5, 0.75, 1.0, 1.25, 1.5

bank storage dynamic processes analysis (Doble et al., 2012) and the study of dual
permeability systems (Schwartz et al., 2010). Further description of the code and its
numerical formulation can be found in Therrien et al. (2009) and Brunner and
Simmons (2012).
As shown in Fig. 2, the catchment used in this study is symmetrical. As a result,
all simulations are conducted for only half of the catchment, as shown for one
example of the catchment congurations considered in Fig. 3, and the reported
uxes are correspondingly half of those expected when accounting for both sides of
the stream. Detailed information of the catchment model with the selected values of
the HGS parameters can be found in Li et al. (2014).
As mentioned previously, the HGS models are used to obtain the observed
streamow (qobs obs obs
T ), baseow (qB ) and quickow (qQ ) hydrographs for the 66
catchments. The baseow hydrograph (qobs B ) is extracted from the model using the
Hydraulic Mixing-Cell (HMC) method (Partington et al., 2011, 2013).
The HMC method developed by Partington et al. (2011) utilises the spatiotem-
poral information of ow generation mechanisms to obtain the component hydro-
graphs (i.e. baseow and quickow hydrographs). When combining the HMC
method with the fully integrated SW/GW model, each node in the surface domain of
the model is treated as a mixing-cell. The method uses the uid mass balance from
the fully integrated SW/GW model of each node at each model time step to calculate
the fraction of water in the cell that comes from different streamow generation
mechanisms (e.g. groundwater discharge to the stream). For each component of
streamow, the fraction is determined using the modied mixing rule (Campana
and Simpson, 1984). The equation for each fraction f for each streamow genera-
tion mechanism k at time N in cell i is given by Partington et al. (2011) as:
0 1
 Pn N
BV N1 Pm V N C  N1
j1 Vji N1 fjk
N B i j1 ij N1 C N1
fik B N  Cf
@ V N
Vi A ik
Vi N
i
and spatial variability of the abstractions, for modelling rainfallerunoff relation-
ships. It also adopts the structure of transferring a fraction of the generated runoff
direct to the baseow store at the same time as the residual is transferred to the
surface attenuation store. AWBM can be operated at either daily or hourly time
steps, although a daily time step is used in this study. At each time step, rainfall is
added to each of the surface stores and evapotranspiration is subtracted. If there is
any excess from any store, it becomes runoff and is divided between surface runoff
(quickow) and baseow, and total streamow is the sum of surface runoff and
baseow (Boughton, 2004). As the aim of the AWBM is to represent the underlying
physical processes, it is a reasonable assumption that it should be able to reproduce
the dynamics of the rainfallerunoff relationships generated synthetically using HGS.
The detailed structure of the AWBM is given in Fig. 4. As can be seen, the AWBM
uses rainfall and actual evapotranspiration as inputs and principally consists of a
conguration of three different surface storages. The depths of these storages
correspond to the parameters C1, C2 and C3. A fraction of the total area is associated
with each surface storage, as represented by the parameters A1, A2 and A3. Moisture
capacity variation over the catchment is described by the combination of the surface
storages and the related fractional areas. An important feature of AWBM is the
ability to account for baseow when predicting streamow by using a baseow
index (BFI), which is the ratio of the amount of baseow to the total amount of
streamow and determines the proportion of excess moisture at each time step that
is returned to the baseow. The daily baseow recession constant (KBase) and daily
routed surface runoff recession constant (KSurf) are used to describe the daily
discharge from baseow and surface runoff (quickow) storage. A summary of the
AWBM parameters and their ranges used is given in Table 3. A number of software
frameworks have been developed for hydrological modelling and are in active use,
such as the Rainfall Runoff Library (RRL) and hydromad (Hydrological Model
Assessment and Development) package (Andrews et al., 2011). In this study, AWBM
is implemented using the RRL developed by the Cooperative Research Centre on
Catchment Hydrology (www.toolkit.net.au/rrl).

N1 denotes fraction k at time N1

Where there are n sources and m sinks for cell i; fjk
in the neighbouring cell j; V denotes the volume with the superscript denoting time
state and subscript i denoting the cell, ij denoting volume into cell j from cell i over
the time step from N1 to N and ji denoting volume from neighbour j into i.
The HMC method extracts streamow generation mechanisms using only hy-
draulic information and is able to capture the storage effects and time lags within
catchments (see Partington et al., 2013), which provides a means of estimating
baseow from fully integrated SW/GW models. The quickow (qobs Q ) is taken as the
difference between the total streamow (qobs obs
T ) and baseow (qB ).
The dataset of the 66 catchment characteristics and hydrological inputs, as well
as the corresponding simulated total streamow and baseow hydrographs ob-
tained using the integrated SW/GW model, can be downloaded as supplementary
material.
2.3. Australian Water Balance Model (AWBM)
The AWBM is a saturation overland ow model developed by Boughton (1993,
2004) and is now one of the most widely used CRR models in Australia (Marshall
et al., 2004; Ranatunga et al., 2008). The AWBM is a typical lumped CRR model,
with interconnected storages and algorithms that mimic the underlying hydrolog-
ical processes used to describe the movement of water into and out of storages.
AWBM uses three surface stores, representing the impacts of antecedent wetness
2.4. Calibration of Australian Water Balance Model (AWBM)
As stated previously, the effectiveness of calibration methods that take internal
model dynamics into account in different ways is investigated in this study. The
calibration methods are summarised in Section 2.4.1. Details of the optimisation
method and error measure used in the calibration methods are given Sections 2.4.2
and 2.4.3, respectively.
2.4.1. Calibration methods
Values of all of the eight parameters listed in Table 3 are obtained as part of the
different calibration methods investigated. Five of these parameters (A1, A2, C1, C2
and C3) are related to the moisture capacity of the catchment, two (BFI and KBase)
affect the separation of rainfall excess into quickow/baseow storage and the
baseow component of total streamow, and KSurf affects the routed surface runoff,
which is referred to as quickow in this study. Details of the calibration methods are
given in Table 4 and Fig. 5.
In order to be able to assess the impact of calibration methods that take internal
model dynamics into account in different ways on the ability to predict total-, base-
and quick-ow hydrographs, a calibration approach that does not take internal
model dynamics into account explicitly is used to provide a basis of comparison
(Method 1).

Table 2
Hydrological inputs considered (adopted from Li et al., 2014).

City Gauge Average annual rainfall Rainfall data period for Average annual potential ET R/ET (average annual rainfall/average annual
no. (mm/a) sampling (mm/a) potential ET)

Adelaide 023011 510.16 1984e1994 1470.97 0.347

Melbourne 086071 525.33 1968e1978 911.33 0.576
Sydney 070062 1095.85 1982e1992 920.90 1.190
Brisbane 031011 2238.22 1997e2007 2077 1.078
Darwin 014015 1707.16 1999e2009 2056.37 0.811
 L. Li et al. / Environmental Modelling & Software 66 (2015) 57e68 61

Fig. 3. An example of the 3D catchment model for the case study (adopted from Li et al., 2014).

Fig. 4. Structure of AWBM.

Method 1:
In Method 1, all parameters are calibrated simultaneously so as to minimise the
selected error measure between the total streamow obtained using AWBM (qsim T )
and that obtained using HGS (qobs T ) using the selected optimisation method, as
shown in Fig. 5. The aim of this calibration method is to provide the best possible t
to total streamow, without any consideration of internal model dynamics.
As can be seen from Fig. 5, in order to take internal model dynamics into account
explicitly during calibration, six different methods (Methods 2e7) are used. In each
of these methods, a two-step process is adopted as follows:
of total streamow, respectively, and are therefore the two parameters that have
a direct effect on the baseow hydrograph produced by AWBM. Consequently,
by estimating values of these parameters rst, an attempt is made to obtain the
best possible match to the baseow hydrograph.
2. In the second step, BFI and KBase are xed at the values obtained in the rst step
and the remaining parameters are calibrated simultaneously so as to minimise
the selected error measure between the total streamow obtained using AWBM
(qsim obs
T ) and that obtained using HGS (qT ) using the selected optimisation
method.

1. In the rst step, values of two of the model parameters, BFI and KBase, are The methods used for obtaining values of BFI and KBase for Methods 2e7 from
estimated. BFI and KBase are selected for this step as they control the separation the above step 1 can be divided into three broad categories. Methods 2e4 belong to
of rainfall excess into quickow/baseow storage and the baseow component the rst category, in which BFI and KBase are calibrated to baseow extracted from
62 L. Li et al. / Environmental Modelling & Software 66 (2015) 57e68

Table 3 Method 5:
AWBM parameter description and ranges. In Method 5, values of BFI and KBase are obtained directly (i.e. without cali-
bration) by developing regression models predicting KBase and BFI as a function of
Parameter Description Parameter range the catchment characteristics (Table 1) and hydrologic inputs (Table 2) investigated,
A1a Partial area 0e1 similar to the regression relationships for predicting lter performance and optimal
A2a Partial area 0e1 lter parameters developed by Li et al. (2014) used in Methods 6 and 7. Values of BFI
BFI Baseow index 0e1 are calculated as the ratio of the volume of baseow to the volume of total
C1 Surface storage capacity 0e50 streamow obtained from HGS and KBase is obtained by performing recession
C2 Surface storage capacity 0e200 analysis on the total streamow hydrographs obtained from HGS. The resulting
C3 Surface storage capacity 0e500 regression equations and scatter plots are shown in Fig. 6. As can be seen, BFI can be
KBase Daily baseow recession constant 0e1 predicted very well and KBase can be predicted reasonably well. These results also
KSurf Daily surface ow recession constant 0e1 indicate that BFI and KBase are closely related to catchment characteristics and
hydrological inputs.
a
A1 A2 must be less than or equal to 1.0. Method 6:
In Method 6, a modied version of the procedure for improving baseow esti-
mation using RDFs suggested by Li et al. (2014) is used in order to improve the
baseow estimates obtained using the LH lter in Method 5. First, the regression
total HGS streamow using recursive digital lters (RDFs). In order to do this, KSurf, relationship developed by Li et al. (2014) is used to determine whether the LH lter
which determines the magnitude of the quickow component within total can provide satisfactory estimates of baseow for a particular catchment based on
streamow (Fig. 4), is set as 1.0, so that the total streamow produced by AWBM is catchment properties and hydrological inputs. If the performance of the LH lter is
comprised solely of baseow. Method 5 belongs to the second category, in which BFI predicted to be acceptable, Method 3 is used to obtain estimates of BFI and KBase.
and KBase are estimated using regression relationships with catchment character- However, if this is not the case, Method 5 is used.
istics and hydrological inputs. Methods 6 and 7 belong to the third category, in Method 7:
which BFI and KBase are either estimated by calibration to baseow extracted using Method 7 is identical to Method 6, except that Method 4 (Eckhart lter) is used
RDFs, as in Methods 2e4, or by regression, as in Method 5, based on the predicted instead of Method 3 (LH Filter).
accuracy of the RDF. Details of the different methods for estimating BFI and KBase in In order to be able to assess the absolute performance of the different calibration
the rst step of the two-step process described above are given below. methods in terms of their ability to predict base- and quick-ow hydrographs
Method 2: accurately, and to be able to assess the degree to which the structure of AWBM is
In Method 2, BFI and KBase are calibrated simultaneously so as to minimise the able to represent base- and quick-ow processes, two benchmarks are developed. As
selected error measure between the baseow obtained using AWBM (qsim part of the benchmark development process, the parameters that affect baseow
B ) and that
obtained using the Lyne and Hollick (LH) RDF (Nathan and McMahon, 1990), with and quickow are calibrated to the observed base- and quick-ow hydrographs
obs LH0:925 produced by HGS, which are also used to assess the performance of the different
the lter parameter set to its default value of 0.925 (qB ), using the selected
optimisation method. The LH lter is used to obtain the baseow hydrograph as it is calibration methods. A two-step calibration process similar to that used for Methods
one of the most widely used RDFs and has been found to give good results in a 2e4 is used in order to obtain the benchmark base- and quick-ow hydrographs. The
number of case studies (Arnold and Allen, 1999; Li et al., 2014; Murphy et al., 2009). only difference is in the way the parameters affecting baseow (i.e. BFI and KBase)
It should be noted that the LH lter with a lter parameter value of 0.925 is adopted are estimated in step 1 of the procedure for obtaining the baseow benchmark
by Nathan and McMahon (1990) based on comparison of the results of this lter (Benchmark 1) and the way the parameters affecting quickow (BFI and KSurf) are
with manual methods, and this value has been widely adopted by a number of re- estimated in step 1 of the procedure for obtaining the quickow benchmark
searchers since. When people have actually tried to measure baseow (although in (Benchmark 2), as detailed below:
only a few catchments), a value of the LH lter parameter of 0.98 is found to produce Benchmark 1 (Baseow):
a better t (CSIRO and SKM, 2010). Although LH lter parameter values in the range BFI and KBase are calibrated simultaneously so as to minimise the selected error
of 0.9e0.98 are found to be reasonable for real catchments, 0.925 is the most measure between the baseow obtained using AWBM (qsim B ) and that obtained with
commonly accepted value in most published studies. During the calibration process HGS (qBobs HGS ), using the selected optimisation method.
using Method 2, the remaining parameters are set to the optimal values obtained Benchmark 2 (Quickow):
using Method 1. BFI and KSurf are calibrated simultaneously so as to minimise the selected error
Method 3: measure between the quickow obtained using AWBM (qsim Q ) and that obtained
Method 3 is identical to Method 2, except that the LH lter parameter is esti- using HGS (qqobs HGS ), using the selected optimisation method. In order to do this,
mated using the regression relationship developed by Li et al. (2014), which provides KBase, which determines the magnitude of the baseow component within total
an estimate of the optimal lter parameter as a function of catchment characteristics streamow (Fig. 4), is set as 1.0, so that the total streamow produced by AWBM is
and hydrological inputs, rather than using the default value of 0.925. Consequently, comprised solely of quickow.
obs LHopt
the observed baseow is given by (qB ).
Method 4: 2.4.2. Optimisation method
Method 4 is identical to Method 3, except that the Eckhart lter (Eckhardt, 2005) Model calibration is conducted using the shufed complex evolution (SCE-UA)
obs Eckopt
is used to obtain the hydrograph of observed baseow (qB ), rather than algorithm, because it has been proven to be both accurate and efcient in previous
the LH lter. The Eckhart lter is used as it is mathematically identical to the studies (Hapuarachchi et al., 2001). In addition, SCE-UA is widely recognised as
Boughton lter (Boughton, 1993), which has been recommended for the purposes of being one of the best search procedures for use in CRR modelling applications (Ajami
estimating BFI for AWBM by Boughton (2004), but has a lter parameter (BFImax) et al., 2004; Franchini et al., 1998). The SCE-UA algorithm is based on the strengths of
that can be estimated more easily from catchment characteristics and hydrological several existing search procedures, including genetic algorithms (GAs) (Goldberg,
inputs than the corresponding lter parameter in the Boughton lter (Li et al., 2014). 1989) and the Nelder & Mead Simplex downhill search scheme (Nelder and Mead,

Table 4
Summary of calibration methods.

Calibration Details
method

1 Calibrate all parameters simultaneously to total streamow.

2 First calibrate BFI and KBase to baseow obtained using LH lter with default lter parameter (0.925) and then calibrate remaining parameters to total
streamow.
3 First calibrate BFI and KBase to baseow obtained using LH lter with optimal lter parameters and then calibrate remaining parameters to total
streamow.
4 First calibrate BFI and KBase to baseow obtained using Eckhart lter with optimal lter parameters and then calibrate remaining parameters to total
streamow.
5 Estimate BFI and KBase based on regression relationships with catchment characteristics and hydrological inputs and then calibrate remaining
parameters to total streamow.
6 Use the regression relationship developed by Li et al. (2014) to determine whether the LH lter can provide satisfactory estimates of baseow. If so, use
Method 3. If not, use calibration Method 5.
7 Identical to Method 6, except that Method 4 is used instead of Method 3.
 L. Li et al. / Environmental Modelling & Software 66 (2015) 57e68 63

Fig. 5. Summary of AWBM calibration methods.

1965), but also introduces the concept of complex shufing (Duan et al., 1992). A
detailed description of this method can be found in Duan et al. (1992).
In this study, SCE-UA is implemented using the Rainfall Runoff Library (www.
toolkit.net.au/rrl). The number of complexes is set equal to the number of calibra-
tion parameters to reduce the chance of premature termination of the search al-
gorithm, as suggested by Kuczera (1997). All of the other parameters are set to the
recommended values in Duan et al. (1994). In order to check whether parameter
equinality is a potential problem and to ensure near globally optimal solutions are
obtained, each calibration run is repeated ten times. Another important aspect in the
application of SCE-UA is that a parameter space needs to be dened, in which the
algorithm searches for the optimal parameter combination. The ranges of all of the
eight AWBM parameters used are based on the suggestions in the Rainfall Runoff
Library (www.toolkit.net.au/rrl) and are listed in Table 3.
The length of streamow data available for AWBM calibration is 10 years (Li
et al., 2014). AWBM runs on a daily time step, and therefore, is calibrated against
daily streamow. The calibration period requires a warm up period, which enables
the models to account for residual water in the catchment by partially lling their
storages, which primes the models for the calibration period. Initial testing dem-
onstrates that a 1 year warm up period is sufcient for model calibration. Chrono-
logically, the warm up period directly precedes the calibration period.
2.4.3. Error measure
The choice of an appropriate error measure is very important (e.g. see Bennett
et al., 2013). In this study, the NasheSutcliffe coefcient (Ef) (Nash and Sutcliffe,
1970) is chosen for this purpose, as it is one of the most highly used performance
measures in hydrology. Ef values are calculated by comparing the difference

Fig. 6. Nonlinear regression models for the prediction of BFI and KBase.
64 L. Li et al. / Environmental Modelling & Software 66 (2015) 57e68

between the observed (e.g. outputs from HGS simulations) and simulated time of the KBase and KSurf parameters based on the calibration pro-
series for each time step for total streamow, baseow and quickow from AWBM. cedure (i.e. Method 1) and only accept values that seem reasonable.
The Ef value ranges from  (very bad model performance) to 1.0 (perfect model
performance), with 0.0 indicating a model that can make predication of the same
Overall, the results show that total streamow is predicted well
quality as the mean of the observations. An exact criterion of ranges of Ef values using all calibration methods investigated, with poor model
indicting model performance does not exist in the literature, although various performance for fewer than 10% of the catchments, except when
empirical guidelines are given when implementing Ef as the metric. Chiew and Method 5 (Regression) is used, in which case 15% of catchments
McMahon (1993) and Ladson (2008) suggest that Ef values above about 0.7e0.8
result in poor model performance (Fig. 7(a)). For most of the
indicate acceptable model performance, while Moriasi et al. (2007) suggest, in
general, the performance of the model can be judged as satisfactory when Ef values catchments (54%e68%) good performance is obtained. However,
are greater than 0.5. In this study, Ef values between 0.5 and 1.0 correspond to good the internal model dynamics are not represented as well, especially
model performance; Ef values between 0.0 and 0.5 show acceptable model per- for baseow, where poor performance is obtained for more than
formance and poor model performance is represented by negative values of Ef. half (58%e74% (ignoring Method 1 with parameter set 2)) of the
However, Gupta and Kling (2011) indicate that high values of Ef can give poor model
performance. Consequently, although more subjective than the use of statistical
catchments and good performance for fewer than 20% of the
measures of goodness-of-t, plots of simulated and observed hydrographs are also catchments (Fig. 7(b)). The ability of AWBM to represent quickow
inspected following optimisation. is slightly better, with good performance for 32%e41% of the
catchments when the calibration methods that consider internal
2.5. Evaluation of model performance
As mentioned previously, the performance of AWBM models calibrated using
the different methods outlined in Section 2.4.1 is compared in terms of overall model
predictive performance (i.e. how well the total streamow generated using AWBM
(qsim obs
T ) matches the corresponding streamow generated using HGS (qT )) and the
accuracy of the resulting internal model dynamics (i.e. how well the AWBM
generated baseow (qsim sim
B ) and quickow (qQ ) hydrographs match the corre-
sponding hydrographs obtained using HGS) (i.e. (qobs B and qobs
Q )) (Fig. 1). The per-
formance of models developed using the different calibration methods is assessed
using Ef and by visual inspection.

3. Results and discussion

The performance of the AWBMs calibrated with the seven

different methods investigated, as well as that of the two bench-
marks, is summarised in Fig. 7. As can be seen, the results are
presented in terms of the percentage of models developed for the
66 catchments resulting in good (Ef  0.5), acceptable
(0  Ef < 0.5) and poor (Ef < 0) performance, for each of the total-,
base- and quick-ow hydrographs. It can also be seen that two sets
of results are presented for Method 1. This is because there are two
distinct sets of model parameters that result in similar model
performance in terms of total streamow during the 10 calibration
trials conducted for some of the 66 catchments. This is because
when Method 1 is used, there is no control on internal model dy-
namics and the only objective is to nd a set of model parameters
that provides the best match to the observed total streamow
hydrograph, as discussed previously. However, as shown in Fig. 4,
the AWBM total streamow is the sum of routed surface runoff and
baseow, both of which are modelled in an identical fashion, each
with a single parameter (KBase for baseow and KSurf for routed
surface runoff). Consequently, similar total streamow can be ob-
tained by exchanging parameter values for KBase and KSurf,
resulting in model equinality, as observed in the calibration re-
sults for Method 1.
An example of this is given in Fig. 8. As can be seen in Fig. 8(a)
and (b), good overall model performance is obtained for both
parameter sets in terms of matching total streamow. However, the
internal dynamics are much better when parameter set 1 is used, as
indicated by signicantly better matches to the base- and quick-
ow hydrographs. It is clearly evident that the baseow and
quickow patterns are reversed for the models with the different
parameter sets (e.g. the pattern of baseow obtained with param-
eter set 1 is very similar to the pattern of quickow obtained with
parameter set 2). In Fig. 7, all of the results with parameters that
result in better internal model dynamics (e.g. as in Fig. 8(a)) are
represented by Set 1, whereas the results with parameters that
result in poorer internal model dynamics (e.g. as in Fig. 8(b)) are
represented by Set 2. However, in the practical application of
AWBM, the parameters of Set 2 can be avoided. This is because
experienced modellers would usually subjectively assess the values Fig. 7. Performance of AWBMs for the different calibration methods investigated.
 L. Li et al. / Environmental Modelling & Software 66 (2015) 57e68 65

Fig. 8. Example of total-, base- and quick-ow hydrographs obtained for two distinct parameter sets obtained using Method 1 for one of the 66 catchments investigated; (a) gives
the total-, base- and quick-ow hydrographs simulated using AWBM with the good set of parameters from calibration Method 1; (b) shows the total-, base- and quick-ow
hydrographs simulated using AWBM with the bad set of parameters from calibration Method 1.

model dynamics explicitly (i.e. Methods 2e7) are used and poor
performance for fewer than 30% of the catchments (12%e29%)
(Fig. 7(c)). It should be noted that when Method 1 is used, good
performance is only obtained for 12% of catchments, but accept-
able performance is achieved for the vast majority of the
remaining catchments (71%), with poor performance for only 17% of
the catchments.
Method 1 is the best-performing calibration approach in terms
of matching total streamow (Fig. 7(a)). This is not surprising, as
the method calibrates all of the model parameters simultaneously
in order to obtain the best match to total streamow. However, this
is at the expense of internal model dynamics. As discussed above,
there is a potential problem with equinality as a result of the
structure of AWBM. As shown in Fig. 7, even if the results for Set 1
are considered, the method results in the second highest percent-
age of catchments with poor model performance for baseow
estimation and the lowest percentage of catchments with good
model performance for quickow estimation. Nevertheless, the
performance of Method 1 is comparable with that of the other
methods in terms of baseow estimation, as baseow is estimated
poorly, irrespective of the method used. In relation to quickow
estimation, there is only a small percentage of catchments with
poor model performance, although Method 1 results in a signif-
icantly smaller number of catchments with good performance
compared with the other calibration methods. Consequently, the
overall performance of the method is best in terms of total
streamow prediction and reasonable compared with that of the
other methods in terms of baseow and quickow prediction.
However, the equinality problem requires careful attention,
although in practice, experienced modellers could avoid an inap-
propriate parameter set (i.e. Set 2) by accepting the parameter set
that seems reasonable, as discussed above. This is because very
poor internal model dynamics can be obtained (e.g. poor per-
formance for 100% of the catchments in terms of baseow predic-
tion and for 50% of the catchments in terms of quickow
prediction) if the inappropriate parameter set (Method 1 (Set 2)) is
used.
The methods that calibrate the BFI and KBase to the baseow
extracted using RDFs (i.e. Methods 2-4) result in the best match to
baseow. This indicates that the RDFs are able to produce reason-
ably accurate estimates of baseow, which is evidenced by the fact
that the performance of Methods 2e4 is only slightly worse than
that of Benchmark 1 (i.e. where BFI and KBase are calibrated to the
baseow produced by HGS, which is the same baseow used for
66 L. Li et al. / Environmental Modelling & Software 66 (2015) 57e68
performance assessment). It also indicates that there is some
benet in terms of improving internal model dynamics by con-
straining BFI and KBase to ensure that baseow is matched as well
as possible.
As stated previously, the benchmarks are used to test how much
the changes in performance are due to the calibration method or
the model structure. The relatively poor performance for Bench-
mark 1 (i.e. poor performance for 57% of the catchments) tends to
suggest that the improvement that is possible by adopting this
approach is rather limited, due to problems imposed by the con-
straints of the model structure in AWBM preventing simulation of
the whole range of baseow dynamics resulting from the complex
underlying physical processes associated with baseow generation
and depletion.
Of the three methods considered, Method 2, which uses the LH
lter with its default lter parameter value of 0.925, performs
worst, while the performance of the other two methods (i.e. LH and
Eckhart lters with optimal lter parameters) is very similar. The
Eckhart lter performs slightly better, with a slightly larger per-
centage of catchments with good performance and a slightly
smaller percentage with poor model performance. This supports
the suggestion by Boughton (2004) that the Boughton lter, which
is mathematically identical to the Eckhart lter, as mentioned
previously, is well suited to use with AWBM.
While use of Methods 2e4 is able to produce the best results in
terms of baseow estimation, there are some trade-offs in terms of
the ability to match total streamow and quickow. Even though
use of Methods 2e4 results in a signicant increase in the per-
centage of catchments with good performance in terms of
quickow prediction compared with Method 1, there is also a slight
increase in the percentage of catchments with bad performance.
In relation to total streamow, use of Methods 2e4 also results in a
slight reduction in performance compared with Method 1. Overall,
the LH method with optimal lter parameter values produces the
best trade-offs in performance between matching total-, base- and
quick-ow among Methods 2e4. However, a disadvantage of this
method compared with Method 2 is that information on catchment
properties and hydrological inputs is needed in order to apply the
regression equations used to obtain optimal lter parameter values,
making it more difcult to apply.
The regression method (Method 5) results in the best perfor-
mance in terms of matching quickow, performing only slightly
worse than the quickow benchmark (Benchmark 2). The method
results in the smallest percentage of catchments with poor per-
formance and second highest percentage of catchments with
good performance. However, it is also the worst-performing
method in terms of matching total- and quick-ow. An advantage
of the method is that it does not require estimates of baseow
hydrographs (only total streamow hydrographs are needed), but a
disadvantage is that information on catchment properties and hy-
drological inputs is needed in order to apply the regression equa-
tions used to obtain values of BFI and KBase as part of the
calibration process.
The two methods combining aspects of the lter (Methods 2-4)
and regression (Method 5) methods (Methods 6-7) provide a good
compromise in terms of performing reasonably well on all three
hydrographs (i.e. total-, base- and quick-ow). While they are not
the best-performing methods for any of the three hydrograph
components, their performance is more consistent than that of any
of the other methods and not far from that of the best-performing
method in each case. Of the two methods, the method using the LH
lter (Method 6) performs better than the method using the Eck-
hart lter (Method 7) overall. A disadvantage of these methods is
that they are more complex to apply than the other methods, as
they require baseow extraction using a RDF, as well as data on
catchment characteristics and hydrological inputs in order to apply
the regression equations for predicting lter performance, obtain-
ing the optimal lter parameters and obtaining direct estimates of
BFI and KBase in the case where predicted lter performance is
poor.
Overall, the results quantitatively demonstrate that while total
streamow can be predicted very well over a wide range of
catchment characteristics and hydrological inputs using AWBM, the
component hydrographs are not modelled very well, particularly
baseow. This raises questions about the way the processes asso-
ciated with these streamow components are conceptualised in
AWBM. The pertinence of such questions is dependent on the
purpose of the model though. If the model purpose is only pre-
dicting total streamow over short time periods (days to weeks)
within the bounds (minimum and maximum ow) for which it has
been calibrated, then internal model dynamics are likely a non-
issue and it is just a curve tting exercise. However, in such in-
stances, black-box models, such as articial neural networks, might
perform better. If the model purpose requires the model to be right
for the right reasons, then process representation is important.
This highly difcult challenge is strived for where the model use is,
for example: aiding catchment behaviour understanding and/or
simulating ows outside of the calibration range.
Of the methods investigated, Methods 3 and 4, as part of which
BFI and KBase are calibrated to the baseow obtained using the LH
and Eckhart lters with optimal lter parameters, provide the best
performance in terms of baseow hydrograph prediction. The
regression method (Method 5) provides the best match to the
quickow hydrographs and the hybrid method using the LH lter
with optimal lter parameters and the regression approach
(Method 6) provides the best overall trade-offs in terms of
matching all three hydrographs.
4. Summary and conclusions
In this paper, the impact of seven different calibration methods
on the ability of AWBM to predict total-, base- and quick-ow
hydrographs is assessed for 66 synthetic catchments with
different physical properties and hydrological inputs. The results
indicate that total streamow can generally be predicted to an
acceptable level for more than 90% of the 66 synthetic catchments.
In contrast, baseow is predicted poorly, with acceptable perfor-
mance levels ranging from 26% to 41% for the different calibration
methods investigated. However, prediction of quickow is much
better, with acceptable performance levels ranging from 71% to
88%. This disparity in performance between baseow and quick-
ow prediction is despite the fact that the hydrographs for the 66
synthetic catchments consist of a wide range of BFI values
(0.006 < BFI < 0.997, Median BFI 0.53), suggesting that the way
AWBM represents internal physical processes could be improved by
using additional information about the catchment, as shown in
calibration Methods 2e7.
Use of the calibration methods that take internal model dy-
namics into account explicitly (Methods 2-7) results in improved
prediction of the component hydrographs. This improvement is
particularly signicant in terms of producing good estimates of
quickow, with the percentage of catchments for which good
performance is obtained increasing from ~10% to ~40% for most
methods. However, as AWBM is unable to provide good predictions
of baseow in general, as discussed above, the improvements ob-
tained by using Methods 2-7 are only small (generally < 5%).
This study offers an analysis of seven calibration methods under
which model internal consistency is assessed. Which calibration
method should be used is based which component of the ow is of
most interest. Generally, total streamow prediction is the most
 L. Li et al. / Environmental Modelling & Software 66 (2015) 57e68
important, and therefore, method 1 (i.e. simultaneously calibrating
all parameters to total streamow) is likely to be the most appro-
priate. However, in order to obtain reasonable internal model dy-
namics, the problem of equinality needs to be addressed, although
experienced modellers usually have an idea of which parameter
sets are reasonable. For situations where baseow is of most in-
terest, such as low ow forecasting, identication of source areas
and recharge estimation, the methods that calibrate BFI and KBase
to baseow extracted using recursive digital lters (RDFs)
(Methods 2-4) should be used. Of these methods, the method that
uses the Eckhardt lter with the optimal lter parameters obtained
using the regression equations developed by Li et al. (2014)
(Methods 4) performs best. If quickow prediction is the primary
objective (i.e. ood prediction), the regression method (Method 5)
should be used. When all three hydrographs are important for
water management, the hybrid method using the Lyne and Hollick
(LH) lter with optimal lter parameter and the regression method
(Method 6) is likely to give the best performance.
It should be noted that most of the methods that perform best
require information on catchment characteristics and hydrological
inputs in order to apply the regression equations for obtaining
optimal lter parameters (Methods 3, 4 and 6), predicting BFI and
KBase directly (Methods 5 and 6) and estimating the level of per-
formance of the LH lter with optimal lter parameters (Method 6),
which makes them difcult to apply if the required information is
not readily available. However, use of Method 2 (LH lter with
default lter parameter of 0.925) can be used in order to achieve
reasonable improvements in internal model dynamics in such
cases.
It is also important to highlight a number of limitations of this
study that provide avenues for future studies. Firstly, it is worth
pursuing other calibration methods that aim to take the internal
model dynamics into account explicitly. For example, changes
could be made to the objective function to ensure known features
of the total streamow hydrograph are captured, or different
transformations could be applied to the data (e.g. use a BoxeCox
transformation to shift the focus from high ows to low ows
(Romanowicz, 2010)). In addition, uncertainty could be considered
during the AWBM calibration process (e.g. Ahmadi et al., 2014;
Fonseca et al., 2014; Zuidema, 2011). In addition, the fact that
synthetic data are used in this study limits its complexity and re-
alism. However, there is a trade-off between realism and the ability
to assess model dynamics accurately. For real catchments, the
actual base- and quick-ow hydrographs are generally unknown
and have to be estimated using a variety of indirect methods (e.g.
tracers, RDFs). However, the approach adopted in this study enables
the base- and quick-ow hydrographs to be known with certainty,
enabling an accurate assessment of the internal dynamics of AWBM
to be obtained in a simplied setting. Lastly, while the computa-
tional experiments are conducted for a range of catchment char-
acteristics and hydrological inputs, these values are also
represented in a simplied manner (e.g. uniform rainfall, homo-
geneous soils, uniform slopes etc.) and the effect of additional
complexity (see e.g. Frei and Fleckenstein, 2014) on the results
should be investigated in future studies. In addition, it would be
interesting to compare the results obtained using the approach
presented in this study with alternative approaches to assessing the
internal dynamics of conceptual rainfallerunoff models, such as
sensitivity analysis (e.g. Shin et al., 2013).
Finally, while the focus of this paper is on the assessment of the
internal model dynamics of the AWBM, the generic approach pre-
sented is equally applicable to other CRR models. As mentioned in
the Introduction, the methodology presented in this paper provides
a rst step towards the more rigorous assessment of the internal
dynamics of CRR models, providing insight as to whether they
67
primarily provide a curve-tting service, similar to that of black-
box models, or whether they are able to provide a reasonable
representation of the underlying processes, as suggested by their
structure. If the latter applies, there can be greater condence in the
results obtained from CRR models, particularly when they are
applied in settings outside those used for model calibration.
Consequently, the methodology presented in this paper should be
applied to other CRR models in order to assess their internal
dynamics.
Acknowledgements
This research was supported by the Chinese Scholarship Council
(CSC) and the University of Adelaide. The authors also gratefully
thank the three anonymous reviewers for their detailed comments,
which have greatly improved the quality of the paper.
Appendix A. Supplementary data
Supplementary data related to this article can be found at http://
dx.doi.org/10.1016/j.envsoft.2014.12.015.
References
Abbott, M.B., Bathurst, J.C., Cunge, J.A., O'Connell, P.E., Rasmussen, J., 1986. An
introduction to the European Hydrological System d Systeme Hydrologique
Europeen, SHE, 1: history and philosophy of a physically-based, distributed
modelling system. J. Hydrol. 87 (1e2), 45e59.
Abrahart, R.J., Anctil, F., Coulibaly, P., Dawson, C.W., Mount, N.J., See, L.K.,
Shamseldin, A.Y., Solomatine, D.P., Toth, E., Wilby, R.L., 2012. Two decades of
anarchy? Emerging themes and outstanding challenges for neural network
river forecasting. Prog. Phys. Geogr. 36 (4), 480e513.
Ahmadi, M., Arabi, M., Ascough Ii, J.C., Fontane, D.G., Engel, B.A., 2014. Toward
improved calibration of watershed models: multisite multiobjective measures
of information. Environ. Model. Softw. 59 (0), 135e145.
Ajami, N.K., Gupta, V.K., Wagener, T., Sorooshian, S., 2004. Calibration of a semi-
distributed hydrologic model for streamow estimation along a river system.
J. Hydrol. 298, 112e135.
Andrews, F.T., Croke, B.F.W., Jakeman, A.J., 2011. An open software environment for
hydrological model assessment and development. Environ. Model. Softw. 26
(10), 1171e1185.
Arnold, J.G., Allen, P.M., 1999. Automated methods for estimating baseow and
ground water recharge from streamow records. J. Am. Water Resour. Assoc. 35
(2), 411e424.
Banks, E.W., Brunner, P., Simmons, C.T., 2011. Vegetation controls on variably
saturated processes between surface water and groundwater and their impact
on the state of connection. Water Resour. Res. 47 (11), W11517.
Bennett, N.D., Croke, B.F.W., Guariso, G., Guillaume, J.H.A., Hamilton, S.H.,
Jakeman, A.J., Marsili-Libelli, S., Newham, L.T.H., Norton, J.P., Perrin, C.,
Pierce, S.A., Robson, B., Seppelt, R., Voinov, A.A., Fath, B.D., Andreassian, V., 2013.
Characterising performance of environmental models. Environ. Model. Softw.
40 (0), 1e20.
Beven, K., 1993. Prophecy, reality and uncertainty in distributed hydrological
modelling. Adv. Water Resour. 16 (1), 41e51.
Beven, K.J., 2005. Rainfall-runoff modeling: introduction. In: Anderson, M.G. (Ed.),
Encyclopedia of Hydrological Sciences. Wiley and Sons, pp. 1e12.
Boughton, W.C., 1993. A hydrograph-based model for estimating the water yield of
ungauged catchments. Hydrology and Water Resources Symposium. Institution
of Engineers Australia, Newcastle, NSW, pp. 317e324.
Boughton, W.C., 2004. The Australian water balance model. Environ. Model. Softw.
19 (10), 943e956.
Brookeld, A.E., Sudicky, E.A., Park, Y.-J., Conant Jr., B., 2009. Thermal transport
modelling in a fully integrated surface/subsurface framework. Hydrol. Process.
23 (15), 2150e2164.
Brunner, P., Cook, P.G., Simmons, C.T., 2009. Hydrogeologic controls on disconnec-
tion between surface water and groundwater. Water Resour. Res. 45 (1),
W01422.
Brunner, P., Simmons, C.T., 2012. HydroGeoSphere: a fully integrated, physically
based hydrological model. Ground Water 50 (2), 170e176.
Campana, M.E., Simpson, E.S., 1984. Groundwater residence times and recharge
rates using a discrete-state compartment model and C-14 data. J. Hydrol. 72,
171e185.
Chen, J., Adams, B.J., 2006. Integration of articial neural networks with conceptual
models in rainfall-runoff modeling. J. Hydrol. 318 (1e4), 232e249.
Chiew, F.H.S., McMahon, T.A., 1993. Assessing the adequancy of catchment
streamow yield estimates. Aust. J. Soil Res. 31, 665e680.
68 L. Li et al. / Environmental Modelling & Software 66 (2015) 57e68
Chiew, F.H.S., Peel, M.C., Western, A.W., 2002. Application and testing of the simple
rainfall-runoff model SIMHYD. In: Singh, V.P., Frevert, D. (Eds.), Mathematical
Models of Small Watershed Hydrology and Applications. Water Resources
Publications, Littleton, Colorado, pp. 335e367.
CSIRO, SKM, 2010. Baseow Assessment for the Murray-Darling Basin. CSIRO: Water
for a Healthy Country National Research Flagship, 78 pp.
Delleur, J.W., 1982. Introduction to urban hydrology and stormwater management.
Urban Stormwater Hydrology. AGU, Washington, DC, pp. 1e34.
Doble, R., Brunner, P., McCallum, J., Cook, P.G., 2012. An analysis of river bank slope
and unsaturated ow effects on bank storage. Ground Water 50 (1), 77e86.
Duan, Q., Sorooshian, S., Gupta, V., 1992. Effective and efcient global optimization
for conceptual rainfall-runoff models. Water Resour. Res. 28 (4), 1015e1031.
Duan, Q.Y., Sorooshian, S., Gupta, V.K., 1994. Optimal use of the SCE-UA global
optimization method for calibrating watershed models. J. Hydrol. 158, 265e284.
Dukic, V., 2006. Modeling of base ow of the Basin of Kloubara River in Serbia.
J. Hydrol. 327 (1e2), 1e12.
Eckhardt, K., 2005. How to construct recursive digital lters for baseow separation.
Hydrol. Process. 19 (2), 507e515.
Ferket, B.V.A., Samain, B., Pauwels, V.R.N., 2010. Internal validation of conceptual
rainfallerunoff models using baseow separation. J. Hydrol. 381, 158e173.
Fonseca, A., Ames, D.P., Yang, P., Botelho, C., Boaventura, R., Vilar, V., 2014. Water-
shed model parameter estimation and uncertainty in data-limited environ-
ments. Environ. Model. Softw. 51 (0), 84e93.
Franchini, M., Galeati, G., Berra, S., 1998. Global optimization techniques for the
calibration of conceptual rainfall-runoff models. Hydrol. Sci. 43 (3), 443e458.
Frei, S., Fleckenstein, J.H., 2014. Representing effects of micro-topography on runoff
generation and sub-surface ow patterns by using supercial rill/depression
storage height variations. Environ. Model. Softw. 52 (0), 5e18.
Furey, P.R., Gupta, V.K., 2001. A physically based lter for separating base ow from
streamow time series. Water Resour. Res. 37 (11), 2709e2722.
Furman, A., 2008. Modelling coupled surface-subsurface ow processes: a review.
Vadose Zone J. 7 (2), 741e756.
Gibbs, M.S., Maier, H.R., Dandy, G.C., 2012. A generic framework for regression
regionalization in ungauged catchments. Environ. Model. Softw. 27e28 (0),
1e14.
Goldberg, D.E., 1989. Genetic Algorithms in Search, Optimization, and Machine
Learning. Addison-Wesley Longman Publishing co., Inc., Boston, MA, USA.
Gupta, H.V., Kling, H., 2011. On typical range, sensitivity, and normalization of Mean
Squared Error and Nash-Sutcliffe Efciency type metrics. Water Resour. Res. 47
(10), W10601.
Hapuarachchi, H., Li, Z., Wang, S., 2001. Application of SCE-UA method for cali-
brating the Xinanjiang watershed model. J. Lake Sci. 12 (4), 304e314.
HydroGeoLogic, 2000. MODHMS: a Comprehensive MODFLOW-based Hydrologic
Modelling System, Code Documentation and User's Guide. HydroGeoLogic Inc.,
Herndon, VA.
Jayatilaka, C.J., Storm, B., Mudgway, L.B., 1998. Simulation of water ow on irrigation
bay scale with MIKE-SHE. J. Hydrol. 208, 108e130.
Knapp, H.V., Durgunoglu, A., Ortel, T.W., 1991. A Review of Rainfall-runoff Modeling
for Stormwater Management. U.S. Geologic Survey, Illinois District, p. 96.
Kuczera, G., 1997. Efcient subspace probabilistic parameter optimization for
catchment models. Water Resour. Res. 33 (1), 177e185.
Ladson, A., 2008. Hydrology: an Australian Introduction. Oxford University Press,
Australia & New Zealand.
Li, L., Maier, H.R., Lambert, M.F., Simmons, C.T., Partington, D., 2013. Framework for
assessing and improving the performance of recursive digital lters for base-
ow estimation with application to the Lyne and Hollick lter. Environ. Model.
Softw. 41 (0), 163e175.
Li, L., Marier, H.R., Partington, D., Lambert, M.F., Simmons, C.T., 2014. Performance
assessment and improvement of recursive digital baseow lters for catch-
ments with different physical characteristics and hydrological inputs. Environ.
Model. Softw. 54, 39e52.
Maier, H.R., Jain, A., Dandy, G.C., Sudheer, K.P., 2010. Methods used for the devel-
opment of neural networks for the prediction of water resource variables in
river system: current status and future directions. Environ. Model. Softw. 25 (8),
891e909.
Marshall, L., Nott, D., Sharma, A., 2004. A comparative study of Markov chain Monte
Carlo methods for conceptual rainfall-runoff modeling. Water Resour. Res. 40
(2), W02501.
McCallum, J.L., Cook, P.G., Brunner, P., Berhane, D., 2010. Solute dynamics during
bank storage ows and implications for chemical base ow separation. Water
Resour. Res. 46 (7), W07541.
Moriasi, D.N., Arnold, J.G., Liew, M.W.V., Bingner, R.L., Harmel, R.D., Veith, T.L., 2007.
Model evaluation guidelines for systematic quantication of accuracy in
watershed simulations. Am. Soc. Agric. Biol. Eng. 50 (3), 885e900.
Murphy, R., Graszkiewicz, Z., Hill, P., Neal, B., Nathan, R.J., Ladson, T., 2009. Project 7:
Baseow for Catchment Simulation (Phase 1- Selection of Baseow Separation
Approach). Australian Rainfall and Runoff Technical Committee, Australia.
Nash, J.E., Sutcliffe, J.V., 1970. River ow forecasting through conceptual models:
part 1. A discussion of principles. J. Hydrol. 10 (3), 282e290.
Nathan, R.J., McMahon, T.A., 1990. Evaluation of automated techniques for base ow
and recession analysis. Water Resour. Res. 26 (7), 1465e1473.
Nelder, J., Mead, R., 1965. The downhill simplex method. Comput. J. 7, 308e313.
Oudin, L., Hervieu, F., Michel, C., Perrin, C., Andreassian, V., Anctil, F., Loumagne, C.,
2005. Which potential evapotranspiration input for a lumped rainfallerunoff
model part 2dtowards a simple and efcient potential evapotranspiration
model for rainfallerunoff modelling. J. Hydrol. 303 (1e4), 290e306.
Panday, S., Huyakorn, P.S., 2004. A fully coupled physically-based spatially-distrib-
uted model for evaluating surface/subsurface ow. Adv. Water Resour. 27 (4),
361e382.
Partington, D., Brunner, P., Frei, S., Simmons, C.T., Werner, A.D., Therrien, R.,
Maier, H.R., Dandy, G.C., Fleckenstein, J.H., 2013. Interpreting streamow gen-
eration mechanisms from integrated surface-subsurface ow models of a ri-
parian wetland and catchment. Water Resour. Res. 49 (9), 5501e5519.
Partington, D., Brunner, P., Simmons, C.T., Therrien, R., Werner, A.D., Dandy, G.C.,
Maier, H.R., 2011. A hydraulic mixing-cell method to quantify the groundwater
component of streamow within spatially distributed fully integrated surface
water-groundwater ow models. Environ. Model. Softw. 26 (7), 886e898.
Partington, D., Brunner, P., Simmons, C.T., Werner, A.D., Therrien, R., Maier, H.R.,
Dandy, G.C., 2012. Evaluation of outputs from automated baseow separation
methods against simulated baseow from a physically based, surface water-
groundwater ow model. J. Hydrol. 458e459, 28e39.
Post, D., Vaze, J., Viney, N., Chiew, F., 2007. Regionalising the Hydrologic Response of
Ungauged Catchments Using the SIMHYD, IHACRES, and Sacramento Rainfall-
runoff Models, MODSIM 2007 International Congress on Modelling and Simu-
lation. Modelling and Simulation Society of Australian and New Zealand,
Christchurch, Christchurch, New Zealand, pp. 2534e2540.
Ranatunga, K., Nation, E.R., Barratt, D.G., 2008. Review of soil water models and
their applications in Australia. Environ. Model. Softw. 23 (9), 1182e1206.
Refsgaard, J.C., Knudsen, J., 1996. Operational validation and intercomparison of
different types of hydrological models. Water Resour. Res. 32 (7), 2189e2202.
Romanowicz, R., 2010. An application of a log-transformed low-ow (LTLF) model
to baseow separation. Hydrol. Sci. J. 55 (6), 952e964.
Schwartz, F.W., Sudicky, E.A., Mclaren, R.G., Park, Y.-J., Huber, M., Apted, M., 2010.
Ambiguous hydaulic head and 14C activities in transient regional ow. Ground
Water 48 (3), 366e379.
Shin, M.-J., Guillaume, J.H.A., Croke, B.F.W., Jakeman, A.J., 2013. Addressing ten
questions about conceptual rainfallerunoff models with global sensitivity an-
alyses in R. J. Hydrol. 503 (0), 135e152.
Sulis, M., Meyerhoff, S.B., Paniconi, C., Maxwell, R.M., Putti, M., Kollet, S.J., 2010.
A comparison of two physics-based numerical models for simulating surface
water-groundwater interactions. Adv. Water Resour. 33 (4), 456e467.
Therrien, R., McLaren, R.G., Sudicky, E.A., Panday, S.M., 2009. HydroGeoSphere: a
Three-dimensional Numerical Model Describing Fully-integrated Subsurface
and Surface Flow and Solute Transport. Waterloo.
Therrien, R., Sudicky, E.A., 1996. Three-dimensional analysis of variably-saturated
ow and solute transport in discretely-fractured porous media. J. Contam.
Hydrol. 23 (1e2), 1e44.
Troutman, B.M., 1985. Errors and parameter estimation in precipitation-runoff
modeling: 1. Theory. Water Resour. Res. 21 (8), 1195e1213.
Tuteja, N.K., Cunnane, C., 1999. A quasi physical snowmelt run-off modelling system
for small catchments. Hydrol. Process. 13 (12/13), 1961e1975.
VanderKwaak, J.E., Loague, K., 2001. Hydrologic-response simulations for the R-5
catchment with a comprehensive physics-based model. Water Resour. Res. 37
(4), 999e1013.
Wu, W., Dandy, G.C., Maier, H.R., 2014. Protocol for developing ANN models and its
application to the assessment of the quality of the ANN model development
process in drinking water quality modelling. Environ. Model. Softw. 54 (0),
108e127.
Zuidema, S., 2011. Identifying Groundwater Contributions to Baseow in a
Temperate Headwater Catchment. Environmental Earth Science, Eastern Con-
necticut State University. Eastern Connecticut State University: Eisenhower,
p. 24.