Journal of Hydrology 525 (2015) 7286

Contents lists available at ScienceDirect

Journal of Hydrology
journal homepage: www.elsevier.com/locate/jhydrol

Comparing ow duration curve and rainfallrunoff modelling

for predicting daily runoff in ungauged catchments
Yongqiang Zhang a,, Jai Vaze a, Francis H.S. Chiew a, Ming Li b
a
CSIRO Land and Water Flagship, GPO BOX 1666, Canberra, ACT 2601, Australia
b
CSIRO Digital Productivity and Services Flagship, Floreat, Western Australia, Australia

a r t i c l e i n f o s u m m a r y

Article history: Predicting daily runoff time series in ungauged catchments is both important and challenging. For the last
Received 5 November 2014 few decades, the rainfallrunoff (RR) modelling approach has been the method of choice. There have been
Received in revised form 16 March 2015 very few studies reported in literature which attempt to use ow duration curve (FDC) to predict daily
Accepted 21 March 2015
runoff time series. This study comprehensively compares the two approaches using an extensive dataset
Available online 27 March 2015
This manuscript was handled by
(228 catchments) for a large region of south-eastern Australia and provides guidelines for choosing the
Konstantine P. Georgakakos, Editor-in-Chief, suitable method. For each approach we used the nearest neighbour method and two weightings a 5-
with the assistance of Hamid Moradkhani, donor simple mathematical average (SA) and a 5-donor inverse-distance weighting (5-IDW) to predict
Associate Editor daily runoff time series. The results show that 5-IDW was noticeably better than a single donor to predict
daily runoff time series, especially for the FDC approach. The RR modelling approach calibrated against
Keywords: daily runoff outperformed the FDC approach for predicting high ows. The FDC approach was better at
High ow predicting medium to low ows in traditional calibration against the NashSutcliffe-Efciency or Root
Low ow Mean Square Error, but when calibrated against a low ow objective function, both the FDC and rain-
Ungauged catchments fallrunoff models performed equally well in simulating the low ows. These results indicate that both
Flow duration curve methods can be further improved to simulate daily hydrographs describing the range of ow metrics in
Hydrological model
ungauged catchments. Further studies should be carried out for improving the accuracy of predicted FDC
South-eastern Australia
in ungauged catchments, including improving the FDC model structure and parameter tting.
Crown Copyright 2015 Published by Elsevier B.V. All rights reserved.

1. Introduction used to predict climate change impact on water availability

Predicting catchment runoff time series is still a challenging
task in surface water hydrology since many catchments around
the world are ungauged or poorly gauged (Sivapalan et al., 2003).
Therefore, the International Association of Hydrological Sciences
launched a decade-long initiative, Predictions in Ungauged Basins
(PUB) (http://www.iahs-pub.org/) in 2003, which has been the
subject of great interest by hydrologists around the world.
Predicting catchment runoff time series in ungauged catch-
ments has direct practical applications. It can be used for designing
characteristics of spillways and embankments, managing water
resource for irrigation, human and industrial water use, hydro-
power operation and environmental ow estimates (Parajka
et al., 2013). It can also be used for risk management, such as ood
and drought forecasting, water quality simulations and predic-
tions. When combined with climate change scenarios, it can be
Corresponding author at: CSIRO Land and Water Flagship, Clunies Ross Street,
Canberra 2601, Australia. Tel.: +61 2 6246 5761; fax: +61 2 6246 5800.
E-mail address: yongqiang.zhang@csiro.au (Y. Zhang).
(Vaze and Teng, 2011; Li et al., 2014b). Furthermore, it can provide
parameter set(s) for hydrological modelling to investigate land use
and land cover change impacts on hydrological processes.
Numerous approaches have been developed for predicting run-
off time series in ungauged catchments, the major one being the
use of hydrological modelling (Parajka et al., 2013). The rainfall
runoff (RR) modelling approach has been widely used for predict-
ing runoff times series in ungauged catchments in Europe
(Bardossy, 2007; Goswami et al., 2007; McIntyre et al., 2005;
Merz and Bloschl, 2004; Oudin et al., 2008; Parajka et al., 2006,
2007; Samaniego et al., 2010a, 2010b), the U.S.A. (Bai et al.,
2009; Kokkonen et al., 2003; Singh et al., 2011; Wagener and
McIntyre, 2005), Australia (Petheram et al., 2012; Post and
Jakeman, 1999; Reichl et al., 2009; Vaze et al., 2010a, 2011;
Zhang and Chiew, 2009), Canada (Samuel et al., 2011), Central
America (Caballero et al., 2013), South America (Allasia et al.,
2006), Africa (Kim and Kaluarachchi, 2008), and the Tibetan
Plateau (Li et al., 2014a). It has also been used to predict climate
change impact on runoff (Chiew and Siriwardena, 2005; Chiew
et al., 2009; Vaze et al., 2010b; Wang et al., 2011) and land use

http://dx.doi.org/10.1016/j.jhydrol.2015.03.043
0022-1694/Crown Copyright 2015 Published by Elsevier B.V. All rights reserved.
 Y. Zhang et al. / Journal of Hydrology 525 (2015) 7286
change impact on runoff (Hundecha et al., 2008; Li et al., 2012,
2013; Tuteja et al., 2007; Zhou et al., 2013).
Various regionalisation methods have been used in hydrological
models to transfer calibrated parameters obtained from gauged to
ungauged catchments (Bloschl and Sivapalan, 1995). Most studies
have used entire sets of parameter values from a donor catchment
to model runoff in the target ungauged catchment, with choice of
the donor catchment based on similarities to the target ungauged
catchment and/or spatial proximity to it. Many studies have shown
that the geographically closest catchment (i.e. spatial proximity) to
the target ungauged catchment is often the best donor catchment
(Bardossy, 2007; McIntyre et al., 2005; Merz and Bloschl, 2004;
Oudin et al., 2008; Parajka et al., 2005; Zhang and Chiew, 2009).
Parameter regression method is used in some studies to transfer
parameters, in which the calibrated parameter sets are related to
catchment attributes to obtain empirical relationships that are
used to estimate model parameters in ungauged catchments
(Kokkonen et al., 2003; Merz and Bloschl, 2004; Oudin et al.,
2008; Parajka et al., 2005). Some comparison studies show that
the spatial proximity performs better than the parameter regres-
sion for regions with dense networks of gauging stations, such as
Austria and France (Oudin et al., 2008; Parajka et al., 2005), and
for the less dense stream-gauge network in Canada (Samuel
et al., 2011). The spatial proximity is chosen in this study to select
the catchment to transfer parameters to predict runoff time series,
which are compared to those obtained using ow duration curve
(FDC) approach.
The FDC approach is widely used for water resource assess-
ments (Vogel and Fennessey, 1995), such as hydropower design
schemes, reliability of water supply and water quality assessments.
However, there are only limited studies reported to use the FDC to
predict runoff time series in ungauged catchments, mainly
reported in North America (the U.S.A. and Canada) (Bloeschl and
Montanari, 2010; Shu and Ouarda, 2012). The FDC approach rst
establishes cumulative distribution functions in gauged catch-
ments, and then estimates the FDC (exceedance probabilities (EP)
of daily runoff) for an ungauged catchment. Several cumulative
distribution functions have been used to present FDC in last several
decades. Among them lognormal distribution was widely used
since this distribution applies several parameters whose values
can be estimated as functions of readily available catchment cli-
mate and physical properties (Cutore et al., 2007; Fenicia et al.,
2011; Fernandez et al., 2000; Goswami et al., 2007; Gupta et al.,
2009; Li et al., 2010). The nal step is to obtain the daily runoff
time series using the estimated FDC and EP in the ungauged (tar-
get) catchment. The EP on any given day is estimated according
to that observed in the neighbouring (donor) catchment (s) on that
day (Shu and Ouarda, 2012).
Numerous approaches have been developed for predicting FDCs
in ungauged catchments in last several decades (Blschl et al.,
2013; Booker and Snelder, 2012; Ganora et al., 2009; Hope and
Bart, 2012; Li et al., 2010; Mohamoud, 2008; Shu and Ouarda,
2012). These approaches can be grouped into spatial interpolation
methods (Castiglioni et al., 2009; Chokmani and Ouarda, 2004; Shu
and Ouarda, 2012) and index ow methods (Castellarin et al., 2004,
2007; Li et al., 2010; Rianna et al., 2011; Shao et al., 2009). The spa-
tial interpolation methods predict a FDC for the target ungauged
catchment using observations from its neighbouring gauged catch-
ments that are put on same or different weights. The index ow
methods estimate FDCs in ungauged catchments using relation-
ships between each parameter of a FDC model and the catchments
climatic and physical characteristics. The index ow methods
become more and more favourable since some FDC parameters
are physically meaningful, such as mean annual runoff
(Castellarin et al., 2007; Smakhtin, 1997) and standard deviation
of daily runoff (Li et al., 2010). The index ow methods are found
73
being superior to the spatial interpolation methods for predicting
FDC in south-eastern Australia (Li et al., 2010; Zhang et al.,
2014). The FDC model developed by Li et al. (2010) and the index
ow model developed by Zhang et al. (2014) are chosen in this
study to predict FDCs in ungauged catchments.
The EP required in the FDC approach for predicting runoff time
series is obtained using various regionalisation approaches (Booker
and Snelder, 2012; Ganora et al., 2009; Hope and Bart, 2012; Li
et al., 2010; Mohamoud, 2008; Shu and Ouarda, 2012). This is simi-
lar to those applied to the RR modelling, as foreshadowed.
Therefore, the regionalisation approach spatial proximity is
chosen in this study to select donor catchment (s) for predicting
the EP in ungauged catchments.
To predict runoff for an ungauged catchment, the FDC and RR
modelling approaches can be used not only with the nearest donor
but also with several neighbouring donors. Several studies have
shown that the Simple mathematical Average (SA), i.e. using the
same weight for each donor to predict runoff, reduced uncertainty
of RR modelling in ungauged catchments (McIntyre et al., 2005;
Oudin et al., 2008; Zhang and Chiew, 2009). However, recent stud-
ies have shown that SA is often not the best solution. Inverse
Distance Weighting (IDW) is found to perform better than SA in
predicting daily runoff time series (Shu and Ouarda, 2012). This
study investigates the performance of a single donor and multiple
donors that were weighted using SA and IDW, respectively.
Studies on runoff predictions in Australia have been intensively
carried out in the last several decades. Surprisingly, the hydrologi-
cal modelling is almost used in all studies, and no attempts have
been reported to use the FDC approach to predict runoff time ser-
ies. Is there any possibility of successfully applying the FDC
approach in this country? Is the FDC approach comparable to the
RR modelling for predicting high ows and/or medium to low
ows?
To answer these questions, this study comprehensively com-
pares the FDC and RR modelling for predicting runoff time series
using a large dataset (228 unregulated catchments) in south-east-
ern Australia. To achieve this objective, following models and mod-
elling experiments are selected, and following comparisons are
designed, including
(1) Selecting the three-parameter FDC model developed by Li
et al. (2010) and a parsimonious RR model, GR4J (Perrin
et al., 2003).
(2) Choosing donor catchments using three regionalisation
schemes: the nearest neighbour, 5-donor SA, and 5-donor
IDW.
(3) Designing two objective functions for calibrating the GR4J
model: one for predicting high daily ows, another for pre-
dicting medium to low daily ows.
(4) Compared the FDC and RR modelling approaches for predict-
ing both high daily ows, and medium to low daily ows,
respectively.
It is expected that the ndings from this study will provide
guidelines for checking suitability of the FDC approach in
Australia, and stimulate similar studies to be carried in other
regions where climatic conditions are different.
2. Study area and data
The study area is located in south-eastern Australia and
includes New South Wales, Victoria, part of South Australia, and
part of Queensland (Fig. 1), and covers the most populated and
important agricultural region of Australia, the Murray Darling
Basin. The area is also of interest because of its low runoff
74 Y. Zhang et al. / Journal of Hydrology 525 (2015) 7286

Fig. 1. Location of the 228 selected catchments in southeastern Australia.

coefcient, high aridity index, and high inter-annual runoff vari-
ability compared to Europe, China, the U.S., and Canada (Peel
et al., 2010; Zhang et al., 2010).
Daily runoff data (Q) were obtained from state water agencies
for 228 unregulated catchments, varying from 50 km2 to
2000 km2 (Vaze et al., 2010a, 2011), and have been checked for
errors to be usable in large scale hydrological modelling. The main
checks for errors include plotting time series and scatter plots of
daily and monthly rainfall and runoff (and runoff coefcient) to
identify inconsistency in the data and checking for recording errors
(spikes in data, same data value for a long period, etc.). The checks
led to the removal of less reliable or suspect data in parts of the
time series in about 5% the catchments. The study region covers
0.15 kPa and 12.2 mm/month, respectively (Jeffrey et al., 2001).
This indicates reasonably good data quality. The maximum and
minimum temperatures, incoming solar radiation, and vapour
pressure data were used to calculate potential evapotranspiration
(ETp) using the PriestleyTaylor model (Priestley and Taylor,
1972). The daily ETp and rainfall data were used as inputs to rain-
fallrunoff modelling. Catchment climatic and physical character-
istics for the 228 catchments are summarised in Table 1.
Table 1
Summary of catchment physical and climatic characteristics in the 228 catchments.

a wide range of climate regimes, and mean annual rainfall for the Characteristics Notation Min 25th Median 75th Max
selected catchments ranges from 457 mm to 2099 mm. Data from Area A (km2) 51 160 333 633 2000
1975 to 2009 were used. Mean annual rainfall P (mm/ 457 708 854 1070 2100
Meteorological data came from the SILO Data Drill of the y)
Queensland Department of Natural Resources and Water (www. Mean annual potential ETp 1032 1126 1202 1288 1508
ET (mm/y)
nrw.gov.au/silo) (Jeffrey et al., 2001) and included daily time series
Standard deviation of P Std (P) 6 9 11 12 37
of maximum and minimum temperatures, incoming solar radia- (mm/y)
tion, vapour pressure, and precipitation (P) from 1975 to 2009 at Aridity index (ETp/P) AI () 0.76 1.22 1.55 1.89 2.98
0.05  0.05 (5  5 km) grid cells. The SILO Data Drill provides Mean catchment E (m) 57 307 519 814 1445
surfaces of daily rainfall and other climate data interpolated from elevation
Catchment slope in S () 0.42 2.75 4.55 7.78 13.85
approximately 4600 point measurements across Australia. The degree
observation network is densest in this study region, compared to Stream length SL (km) 27 121 246 475 1753
other Australian regions. The ordinary kriging method was used Median solum ST (mm) 0.44 0.86 0.96 1.20 2.00
to interpolate daily and monthly precipitation and a thin plate thickness
Plant-available water- PAWC 50.0 82.3 110.8 158.3 265.8
smoothing spline was used to interpolate other climate variables.
holding capacity in (mm)
Cross-validation for SILO maximum daily temperature, minimum the solum
daily temperature, vapour pressure, and precipitation data across Mean woody fwoody () 0.00 0.23 0.52 0.83 1.00
Australia shows the mean absolute errors of 1.0 C, 1.4 C, vegetation fraction
 Y. Zhang et al. / Journal of Hydrology 525 (2015) 7286
3. Methodology
3.1. Overview
This paper uses the two approaches FDC and RR modelling
to predict daily runoff time series in ungauged catchments. Each
of the 228 unregulated catchments was left out, as an ungauged
catchment, and its ve nearest catchments were regarded as
gauged catchments (or donors). All 228 catchments were
stepped through in this way, making sure that both approaches
are evaluated at each ungauged catchment. Each approach was
then evaluated using three regionalisation schemes: the nearest
neighbour (a single donor) and the two weightings 5-SA, and
5-IDW to predict daily runoff time series. The three regionalisa-
tion schemes were used to make sure that the two approaches
were thoroughly compared. Section 3.3.1 describes 5-SA and
5-IDW in detail.
Section 3.2 describes how regionalisation methods using a sin-
gle donor and multiple donors work. Section 3.3 shows details for
the two approaches for predicting daily runoff time series in
ungauged catchments. Finally, Section 3.4 shows criteria for eval-
uating the performance of the two approaches.
3.2. A single donor and multiple donors
In the single-donor FDC approach, the daily runoff time series
for an ungauged catchment is estimated using the daily observed
runoff time series from a nearby donor catchment. In the single-
donor RR modelling approach, however, the daily runoff time ser-
ies for an ungauged catchment is estimated using parameter set
calibrated at a nearby donor catchment together with climatic
forcing data input for the ungauged catchment.
In the multiple-donor FDC approaches, daily runoff time series
for an ungauged catchment is simulated using N closest donors.
The N sets of predicted daily runoffs were then averaged using
SA and IDW. In the multiple-donor RR modelling approach,
parameter values from each donor were used separately to simu-
late runoff in the ungauged catchment. The N sets of modelled
daily runoffs were then averaged using SA and IDW to estimate
the daily runoff for the ungauged catchment.
Our preliminary modelling results showed that for multiple-
donor weightings SA and IDW performed consistently better
as the donor number increased from one to ve. However, when
it increased from ve to seven or more, the predictions from the
weightings were very similar (data not shown here). This nding
is similar to those of hydrological modelling studies carried out
in Europe, Australia, and the U.S.A. (McIntyre et al., 2005; Oudin
et al., 2008; Patil and Stieglitz, 2012). Therefore, ve nearest donors
Fig. 2. Flow duration curves (FDCs) simulated using the three-parameter FDC model. Left panel: for mean of log-transformed daily runoff log Q of 5.0 (top), 3.0 (middle),
and 1.0 (bottom), where zero ow ratio, Rzero, is 0.8 and standard deviation of log Q, std(log Q), is 2.0. Right panel: same values as middle curve at left except that std(log Q) is
1.0, 1.5, and 2.0.
75
were used in the two approaches for predicting runoff, and the
multiple-donor modelling results shown hereafter are all from ve
donors.
3.3. The two approaches for predicting daily runoff time series
3.3.1. Flow duration curve
As mentioned in Section 1, the FDC approach assumes that an
ungauged catchment and its neighbour have an equivalent EP of
daily runoff for a given day. To implement this approach, three
steps were used: (1) predicting FDC for ungauged catchments;
(2) quantifying the EP of daily runoff for a given day; and (3)
estimating daily runoff time series for the ungauged catchment
using the two weightings, SA and IDW.
We used the three-parameter lognormal distribution (F1) devel-
oped by Li et al. (2010) to represent the FDC with the following
cumulative distribution function:
( n o
1  sU log Qr l if x>0
F 1 Q ; l; r; s 1
s if x 0;
where U is the cumulative distribution function of a standard nor-
mal distribution; l and r are the mean and standard deviation of
logarithmic daily runoff (Q), respectively; s is the proportion of zero
runoff.
In step one, the three parameters in the FDC model were quan-
tied using observed daily runoff data for each gauged catchment.
The parameter s was estimated as the zero ow ratio dened as the
ratio of cease ow days to total observation days. The other two
parameters l and r were estimated using the mean (l ^ ) and
standard deviation (r ^ ) of daily runoff at a logarithmic scale as
follows
1X n
l^ logQ i ; 2
n i1
r
1 Xn
r^ flogQ i  l ^ g2 ; 3
n i1
The parameters l and r relatively describe the central tendency
and variation of non-zero runoff respectively. Fig. 2 shows how
the FDC changes with different values of l and r, respectively.
To obtain the three parameters for an ungauged catchment, the
index ow model (Zhang et al., 2014) was used to establish regres-
sions between a parameter and the 11 catchments climatic and
physical characteristics using data from the 228 catchments
(Table 1). The stepwise regression was used to select prediction
variables for each of the three parameters. Since this study used
76 Y. Zhang et al. / Journal of Hydrology 525 (2015) 7286
Table 2
Summary of performance the Eqs. (4)(6) for predicting three parameters. RMSE is the
root mean square error; NSE is the NashSutcliffe Efciency; r is the Pearson
correlation.
Parameter RMSE NSE
Proportion of zero runoff (s, Eq. (4)) 0.12 0.58
Mean of logarithmic daily runoff (l, Eq. (5)) 0.61 0.73
Standard deviation of logarithmic daily runoff 0.30 0.25
(r, Eq. (6))
large number of samples, the relationships established from the
228 catchments can be regarded as same as those established from
227 of them, with one of them left out for cross-validation. The
established regressions were then used to predict the three
parameters for each catchment. The regressions for predicting s,
l, and r for each ungauged catchment were respectively:
s minmax0:38 0:33AI  0:0083S 0:137f woody ; 0; 1
l 0:96  2:34AI; and
r 0:59 0:51AI 0:035stdP;
where AI is the aridity index (the ratio of mean annual potential
evapotranspiration to mean annual precipitation), S is the catch-
ment slope (090), fwoody is the area ratio of deep-rooted vegetation
to total vegetation (01), and P is the daily precipitation (mm/d).
Table 2 summarises the performance of Eqs. (4)(6). The index ow
model performs well for predicting s and l, but performs relatively
poorly for predicting r. For each parameter, the NashSutcliffe
Efciency (NSE) (Nash and Sutcliffe, 1970) obtained by using the
228 catchments is almost same as that obtained using half the
catchments randomly selected (Zhang et al., 2014). This indicates
that the index ow model performs consistently when the tted
samples are large enough.
Once the three parameters are available, a percentile runoff (Qx)
in the FDC can be obtained from the three signatures s, l, and r
using the lognormal distribution as follows:
Q x expfr/1 1  x=1  s lgI0 6 x 6 1  s;
where /() is the standard normal cumulative distribution function.
In step two, the EP of daily runoff for a given day for the
ungauged catchment (pu) was estimated using the following
equation:
pu pj ;
where pj is the exceedance probability of daily runoff on a given day
for the jth gauged catchment
In step three, the daily runoff time series for the ungauged
catchment was determined using the predicted FDC, and the EP
estimated from a neighbouring catchment (Eq. (8)). When N
donors were used to estimate the EP, N sets of daily runoff time
series were generated. They were then averaged using SA and
IDW to obtain one set of runoff time series for the ungauged catch-
ment using the following equation:
,
X
N X
N
Qu wj Q j wj ;
j1 j1
where Qu is the daily runoff in depth (mm/d) for an ungauged catch-
ment for a given day, j is the jth donor, N is the total number of
donor, wj is the weight for the donor j, and Qj is the daily runoff
for the donor j for the given day.
The SA scheme uses the same weight for the donors, i.e. Wj = 1.
This means that daily runoff for an ungauged catchment is esti-
mated using the average of daily runoffs of its donors.
The IDW scheme estimates Qu by averaging runoff observed
from nearby donors, but uses different weights that are deter-
mined by distances between the ungauged catchment and its
donors. The closer a donor catchment is to the ungauged catch-
r ment, the more weight it is given for IDW interpolation. This
0.77 scheme estimates wj using the following equation:
0.86
0.50 1
wj ; 10
dX j ; X u 2
where Xj is the latitude and longitude of centroid of the jth donor, Xu
is the latitude and longitude of centroid of the ungauged catchment,
d is the Euclidian distance between Xj and Xu, and the number 2 is
the power parameter. The higher it is, the higher weight is put on a
closer catchment. This study used 2, a default value recommended
in the ARCGIS software.
3.3.2. RR modelling
4 The four-parameter rainfallrunoff model, GR4J, was used to
simulate daily runoff time series in ungauged catchments for the
5 period of 19752009. This model has been used under a wide
range of hydro-climatic conditions (Coron et al., 2012; Lerat
6 et al., 2012; Li et al., 2014b; Oudin et al., 2008; Perrin et al.,
2003; Zhang et al., 2011), including applications across south-east-
ern Australia (Vaze et al., 2010a). A detailed description on GR4J
can be found in Perrin et al. (2003). The GR4J model was rst cali-
brated against observed daily runoff time series in each of the 228
gauged catchments. As mentioned earlier, the Jackknife left-one-
out cross-validation method was used to estimate the daily runoff
time series for each ungauged catchment where the GR4J model
was driven using the calibrated parameters obtained from its
nearby gauged catchments together with climatic forcing data.
As explained before, this study not only applied the single-donor
(i.e. nearest neighbour) method, but also 5-SA and 5-IDW (Eqs.
(9) and (10)).
A global optimisation method, the genetic algorithm, was used
to optimise four parameters in the GR4J model for each of the 228
gauged catchments. Model calibration period is from 1975 to 2009,
same as the prediction period. In the genetic algorithm, a pop-
7 ulation size of 100 and the maxim generation of 100 were set up
for searching the optimum point. The genetic algorithm normally
converges to the optimum point at about 50 generations of
searching.
Two model calibration experiments were designed to compre-
hensively evaluate GR4J performance. The rst experiment is to
8 test RR modelling performance in simulating high ows.
Therefore, the objective function (F1) for a single catchment
depended on the value of 1  NSE and a logarithmic function of
water balance error (Viney et al., 2009), and is given as:
2:5
F 1 1  NSE 5jln1 WBEj ; 11
where NSE is dened as:
PM  2
i1 Q obs;i  Q sim;i
NSE 1  2 ; 12
PM 
i1 Q obs;i  Q obs
9 and WBE is expressed as:
PM PM
i1 Q sim;i  i1 Q obs;i
WBE PM ; 13
i1 Q obs;i
in which WBE is the water balance error, Qsim and Qobs are the simu-
lated and observed daily runoff, respectively, Q obs is the arithmetic
mean of the observed daily runoff, i is the ith day, and M is the total
number of days sampled. NSE measures the agreement between the
modelled and observed daily values, with NSE = 1.0 indicating
 Y. Zhang et al. / Journal of Hydrology 525 (2015) 7286 77

r
perfect agreement between the modelled and observed daily runoff 1 XM
at a given catchment. WBE measures the water balance error RMSE Q obs;i  Q sim;i 2 : 17
M ii1
between modelled and observed mean annual values, with
WBE = 0 indicating a perfect agreement. The criterion of r is used as a measure of correlation between mod-
The second experiment is to test RR modelling in simulating elled and observed daily runoff and RMSE (mm/d) is used as a mea-
low and medium ows. The objective function (F2) for a single sure of the differences between the two. With smaller RMSE and
catchment is to maximise NSE of log-transformed daily runoff larger r values, the better the modelling results are.
(NSElog):
4. Results
F 1 1  NSElog ; 14
PM 2 The accuracy of predicting daily runoff time series in cross-val-
i1 logQ obs;i e  logQ sim;i e idation mode is summarised in boxplots of NSE (Fig. 3), boxplots of
NSElog 1  PM 2
; 15
i1 logQ obs;i e  logQ obs NSElog (Fig. 4), and boxplots of the absolute WBE results (Fig. 5). The
accuracy of predicting percentile daily runoff is summarised in
where e is a minimum value. The logarithmic transformation does scatterplots for high runoff percentiles (Fig. 6) and medium to
not allow zero ow. To avoid this issue, the minimum value e, low runoff percentiles (Fig. 7), respectively.
0.001 mm/d, is added to both simulated and observed daily runoff
before transformation.
4.1. Comparing a single donor to ve donors
3.4. Evaluation criteria
Figs. 3 and 4 compare the FDC and RR modelling approaches
using a single donor (i.e. the nearest neighbour) to those using
The accuracy of the two approaches for predicting daily runoff
the ve-donor weightings, 5-SA and 5-IDW, in terms of NSE
time series was evaluated in their cross-validation mode for each
(Fig. 3) and NSElog (Fig. 4). For both approaches FDC and RR mod-
of the 228 catchments. Three criteria used are NSE of daily runoff
elling (GR4J) use of ve neighbouring donors noticeably
(Eq. (12)), absolute WBE (Eq. (13)), and NSE of log-transformed
improved overall predictions compared to use of the single nearest
daily runoff (Eq. (14)).
neighbour. This indicates that use of model output averaging can
We also evaluated the two approaches for predicting percentile
overall reduce model uncertainty. A comparison between two
runoff for all 228 catchments. The Pearson correlation (r) and root
model output averaging schemes 5-SA and 5-IDW shows that
mean square error (RMSE) were used to check correlation and
using the inverse-distance weighted interpolation scheme further
errors between observations and predictions across the 228 catch-
improved the performance of the two approaches compared to
ments. The criteria of r and RMSE are expressed as:
using the simple mathematical interpolation scheme. This indi-
0 12
PN cates that to get better modelling results, a closer donor could be
B i1 Q sim;i  Q sim Q obs;i  Q sim C given a higher weight since hydrological signatures such as runoff
r @q A ; 16
PN 2 PN 2 coefcient and FDC are more similar in the closer catchments than
i1 Q sim;i  Q sim i1 Q obs;i  Q sim
in the ones further away.

Fig. 3. Summary of the performance, in terms of NashSutcliffe Efciency (NSE) of daily runoff, of the various methods of calculating daily runoff obtained from 228
catchments. In each boxplot, the bottom, middle, and top of the box are the 25th, 50th, and 75th percentiles, and the bottom and top whiskers are the 10th and 90th
percentiles. FDC is the ow duration curve; the numbers in brackets are the number of donors used; log in GR4J brackets means obtained using calibration parameter sets
against logarithm-transformed daily runoff; SA is simple average; IDW is inverse distance weighting. Abbreviations and boxplot key apply to Figs. 4 and 5.
78 Y. Zhang et al. / Journal of Hydrology 525 (2015) 7286

Fig. 4. Summary of the performance of various methods in terms of NashSutcliffe efciency (NSE) of logarithm-transformed daily runoff obtained from the 228 catchments.
Key as per Fig. 3.

Fig. 5. Summary of the performance of various methods in terms of the absolute water balance error (WBE) obtained from the 228 catchments. Key as per Fig. 3.

Fig. 5 compares the absolute WBE of the FDC and RR modelling
approaches using a single donor to that using the two ve-donor
weightings. For GR4J modelling, use of ve donors noticeably
reduced model biases compared to use of the nearest neighbour,
but the absolute WBE for two-weighting schemes (5-SA and 5-
IDW) are similar. For the FDC approach, using ve donors got simi-
lar biases to using a nearest neighbour.
The results in Sections 4.1 indicate that 5-IDW is the best
among the three donor schemes tested in this study. Therefore,
in the following Sections 4.2 and 4.3, the results obtained from
5-IDW were used to evaluate the performance of the FDC and RR
modelling approaches.
4.2. Comparing the FDC and RR modelling approaches for predicting
runoff time series
The two approaches, FDC and RR modelling, were compared to
each other for predicting high ows (Figs. 3 and 5) and medium to
 Y. Zhang et al. / Journal of Hydrology 525 (2015) 7286 79

Fig. 6. Scatterplots between observed (Qo) and simulated (Qs) daily runoff at 99th percentile (left), 95th percentile (middle) and 90th percentile (right) for the 228
catchments. Qs at the top three panels are from 5-IDW FDC modelling; Qs at the middle three panels are obtained from 5-IDW GR4J modelling (using parameter sets against
daily runoff); Qs at the bottom three panels are from 5-IDW GR4J modelling (using parameters sets against logarithm-transformed daily runoff). Panel keys are applied to
Fig. 7.

low ows (Fig. 4), respectively. The FDC approach was poorer than
the GR4J model calibrated against daily ow, indicated by the NSE
of daily runoff being about 0.10 lower at 50th75th percentiles and
about 0.15 lower at 25th percentile (Fig. 3), and the absolute WBE
being about 0.1 higher at 25th75th percentiles (Fig. 5). The FDC
approach, however, was slightly better than GR4J model calibrated
against log-transformed daily runoff, indicated by the NSE of daily
runoff being about 0.02 higher at the 50th percentile (Fig. 3).
The FDC approach was much better for low ows than the GR4J
model calibrated against daily runoff, indicated by NSElog being
about 0.15 higher at the 50th percentile; it was similar to the
GR4J model calibrated against log-transformed daily runoff, indi-
cated by NSElog being 0.010.04 higher at more than 75th per-
centiles but being 0.010.02 lower at less than 50th percentiles.
4.3. Comparing the FDC and RR modelling approaches for predicting
percentile daily runoff
Fig. 6 shows that GR4J modelling against daily runoff was best
in predicting daily runoff at 99th, 95th and 90th percentiles;
GR4J modelling against log-transformed daily runoff was inter-
mediate; the FDC approach performed worst. The difference is big-
gest at 99th percentile (i.e. extreme high ows/ood conditions),
where the FDC approach got an RMSE of 4.64 mm/d, compared to
an RMSE of 2.46 mm/d obtained from GR4J modelling against daily
runoff time series and an RMSE of 3.42 mm/d obtained from GR4J
modelling against log-transformed daily runoff time series. For
lower percentiles, the difference is reduced. At 90th percentile,
the RMSE value obtained from the FDC approach is only 0.10
0.13 mm/d higher than those obtained from GR4J modelling and
the r value obtained from the FDC approach are only 0.020.03
smaller than those obtained from GR4J modelling.
Fig. 7 shows that GR4J modelling performed marginally better
than the FDC approach at 50th percentile since both approaches
obtained similar RMSE and r value. But at 10th percentile (low ow
conditions), the FDC approach performed better than GR4J mod-
elling against daily runoff, and was similar to GR4J modelling
against log-transformed daily runoff. For prediction of cease to
ow ratio, the FDC approach performed much better than GR4J
modelling against daily runoff and log-transformed daily runoff.
GR4J modelling against daily runoff basically cannot simulate the
cease to ow threshold (0.001 mm/d) for all 228 catchments while
GR4J modelling against log-transformed daily runoff cannot simu-
late cease to ow threshold for majority of the catchments.
To further understand the difference between the FDC and RR
modelling approaches, observed hydrograph is compared to the
simulated ones for a random selected catchment (Wullwye Creek
at Woolway). It clearly shows that the hydrograph simulated by
the GR4J model (calibrated against daily runoff) matches the
observed one best during high ow periods (Fig. 8). However, it
overestimates low ows. Compared to that, the FDC approach ts
the medium to low ows best. In this particular case, the GR4J
80 Y. Zhang et al. / Journal of Hydrology 525 (2015) 7286

Fig. 7. Scatterplots between observed (Qo) and simulated (Qs) daily runoff at 50th percentile (left) and 10th percentile (middle) and scatterplots between observed and
simulated cease ow ratio (right) for the 228 catchments. Key as per Fig. 6.

Fig. 8. Hydrographs of the observed daily runoff (black) and simulated daily runoff from the FDC modelling (blue), simulated from the GR4J modelling (red for experiment 1,
Eq. (11), cyan for experiment 2, Eq. (14)) for catchment Wullwye Creek at Woolway. (For interpretation of the references to colour in this gure legend, the reader is referred
to the web version of this article.)
 Y. Zhang et al. / Journal of Hydrology 525 (2015) 7286 81

model calibrated against log-transformed daily runoff shows the Table 3

worst prediction on high ows, and the second best prediction Summary the probability (%) that 5-IDW is better than (NSE difference > 0.02), similar
to (0.02 6 NSE difference 6 0.02), or poorer (NSE difference < 0.02) than each of
on medium to low ows. ve nearest donors.

Nearest Superiority NSE NSE (log, NSE NSE (log,

5. Discussion (FDC) NSE) (GR4J) GR4J)
1st Better 58.8 78.1 55.7 59.6
5.1. Why model output averaging is better Similar 25.0 12.7 25.9 16.2
Worse 16.2 9.2 18.4 24.1
It is widely believed that the model output averaging reduces 2nd Better 72.8 82.5 57.9 54.8
model uncertainty and signicantly improves model performance Similar 15.8 7.5 14.5 11.8
(McIntyre et al., 2005; Merz and Bloschl, 2004; Merz et al., 2009, Worse 11.4 10.1 27.6 33.3
2011; Nester et al., 2011, 2012). To investigate this further, we 3rd Better 85.5 89.5 68.0 64.5
compared the values NSE and NSElog obtained from 5-IDW to those Similar 9.6 7.0 15.8 10.1
obtained from each of ve neighbouring donors (Fig. 9). This gure Worse 4.8 3.5 16.2 25.4
clearly shows that, for the majority of ungauged catchments, each 4th Better 89.5 93.4 60.1 62.7
individual donor performed more poorly than 5-IDW for both Similar 7.0 3.1 15.4 11.0
Worse 3.5 3.5 24.6 26.3
approaches. This result indicates that use of 5-IDW shows an
averaging effect that does help to reduce modelling uncertainty. 5th Better 95.2 94.3 70.2 63.2
Similar 2.6 1.3 7.9 8.3
Furthermore, as expected, the NSE difference between 5-IDW and
Worse 2.2 4.4 21.9 28.5
a more distant donor was more signicant than the difference
between 5-IDW and a nearer donor (Fig. 9 and Table 3). This result
suggests that use of a nearest neighbour is generally better than
the use of a more distant donor. Use of IDW to put more weight whereas the 5-IDW gave an NSE value of 0.46 (Fig. 10). In this case,
on a closer catchment and less weight on a more distant catchment the best donor slightly underestimates low ows, but signicantly
is generally better than using the closest neighbour alone. underestimates high ows. In contrast, other four donors under-
To further understand the averaging effect from 5-IDW, simu- estimate low ows also, but overestimate high ows. By averaging
lated and observed hydrographs were compared to each other for a the outputs from the ve nearest donors, the 5-IDW signicantly
catchment where the NSE value from the best donor was 0.17 improves estimates of high ows. This example demonstrates that

FDC approach FDC approach GR4J modelling GR4J modelling

Fig. 9. Comparing NSE and NSElog obtained from the 5-IDW to those obtained from each of the 5 nearest donors. The ve ordinates from top to bottom are: numbers 1st to
5th refer to the 1st-nearest to 5th-nearest donor. The left two rows are for the FDC approach; the right two rows are for the GR4J modelling approach.
82 Y. Zhang et al. / Journal of Hydrology 525 (2015) 7286
Fig. 10. Hydrographs of the observed daily runoff (black) and GR4J simulated daily runoff (best donor out of ve, red, and 5-IDW, blue) for catchment 226218. Shadowed area
covers ve sets of simulated results, and the lower and upper boundaries are the minimum and maximum simulations. (For interpretation of the references to colour in this
gure legend, the reader is referred to the web version of this article.)
the weighted average output from multiple donors improve esti-
mate of high ows.
The improvement in performance when using 5-IDW is more
noticeable for the FDC approach than the GR4J modelling, suggest-
ing that the average effect is more helpful for the FDC approach
(Table 3). This probably reects the nature of the two approaches.
The FDC approach predicts runoff time series for the ungauged
catchment only using regionalised FDC and EP. The information
among the donors is independent. Therefore, failure of one donor
at a particular date could be offset by success of other donors on
that day, resulting in a strong averaging effect. The RR modelling
approach uses climatic forcings (P and PET) together with region-
alised parameters to predict runoff time series. The forcing data
among donors are not independent, which means that climatic
forcing data among those donors normally have similar data qual-
ity. Therefore, the chance that one fails but others are successful on
a particular date for the RR modelling approach is smaller than that
for the FDC approach, resulting in the weaker averaging effect.
The current model output averaging results are comparable to
published PUB results. Parajka et al. (2013) summarised 33 runoff
predictions studies that use nearest neighbour, kriging, and inverse
distance weighting interpolating methods. The median NSE from
the 33 studies are 0.58, which is slightly poorer that the median
NSE of 0.605 obtained by GR4J (5-IDW) and are better than the
median NSE of 0.485 obtained by FDC (5-IDW). It is noted that a
summary on NSElog from the published PUB studies is not available
and a direct NSElog comparison between the current and previous
studies cannot be made.
5.2. Can the FDC approach predict high ows better?
As described in Sections 4.2 and 4.3, the FDC approach per-
formed noticeably poorer than rainfallrunoff modelling in pre-
dicting high ow (or oods) in ungauged catchments. This study
used a widely used distribution the lognormal distribution to
present FDC. Except that, other probability distributions such as
Gumbel, Extreme value, or Gamma (including Pearson type III)
can be used to model hydrological extremes (i.e. high ow) in
FDC (Booker and Snelder, 2012; Kovacs et al., 2012; Li et al.,
2009). These alternative probability distributions have great
potential to represent high ows better but are conceptually more
complicated. In particular, more parameters are generally used in
these distributions and the estimation of these distributional
parameters can be interesting but very challenging in the context
of the ungauged catchments. We tested the Gamma distribution,
as an example, to explore the possibility of using other dis-
tributions and understand the difculties of improving the predic-
tion of high ows. The three-parameter Gamma distribution is
chosen in this pilot study because the parameters in the ungauged
catchments can be easily inferred (see the description of parameter
estimation below). A comprehensive comparison of other dis-
tributions to present FDC for better high ow prediction is a natu-
ral extension of this work and will be considered in the future.
The cumulative distribution function of the three-parameter
Gamma distribution is expressed as:
(  
1s
Ck c k; Qh if x>0
F 2 Q ; k; h; s 18
s if x 0;
where C and c are the gamma function and the lower incomplete
Gamma function, respectively; k and h are the shape and scale
parameters, respectively; s is the proportion of zero runoff.
For the Gamma distribution, we estimate the parameters so
that the tted distribution matches the mean and variance of daily
runoff (at an original scale) as follows:
^ V 1 Q 2 ;
k 19
^h VQ 1 ; 20
1
Pn 1
Pn  2
where Q n i1 Q i and V n i1 Q i  Q are the mean and
variance of observed non-zero runoff. The mean daily runoff, Q , in
ungauged catchments was estimated from a widely used Budyko-
framework, the one-parameter Fu model, (Fu, 1981; Teng et al.,
 Y. Zhang et al. / Journal of Hydrology 525 (2015) 7286 83

2012; Zhang et al., 2004; Zhang and Chiew, 2012); the variance of
non-zero runoff, V, was estimated from the linear relationships
between V and variance of daily rainfall (selecting the days with
rainfall more than 2 mm/d after trail-and-error test); the proportion
of zero runoff, s, was estimated from Eq. (4).
Use of the gamma distribution performed more poorly than use
of the lognormal distribution for predicting high ow at percentiles
of 99th, 95th and 90th (Figs. 11 and 6). This may be caused by the
ineffective parameter estimation and inappropriate distribution
representation. The parameter estimation given by Eqs. (19) and
(20) is based on the method of moments that ensures the mean
and variance of daily runoff is preserved. This moment-based
estimation basically focuses on the central tendency of the dis-
tribution but is not necessarily effective to make inference relating
to extremes. Alternatively, the likelihood-based estimation is more
appropriate for extreme studies and may be considered in the
future study. The Gamma distribution may not be able to approxi-
mate the FDC of daily runoff very well, which is a mixture of both
low and high values. A mixture of two distributions (e.g. lognormal
and Gamma) could represent the low and high ows separately
and could be a better approximation of the FDC of daily runoff.
More analysis is required to investigate this in full details and this
is outside the scope of this study.
A recent study has shown that for a given period of time, RR
model calibration against FDC performs better than calibration
against an observed runoff time series for model simulation in
another independent period of time (Westerberg et al., 2011).
This nding stimulates interest in calibrating the RR model against
predicted FDC and thereby predicting runoff time series in
ungauged catchments. As discussed in Section 3.3.2, FDC for an
ungauged catchment can be well estimated using the index ow
model. We further calibrated GR4J against the predicted FDC in
each ungauged catchment. This model calibration still used Eq.
(11) as the objective function, but used the NSE of the ow dura-
tion curve, rather than the NSE of daily runoff. For optimising the
GR4J model, a total of 999 ows were picked from the non-ex-
ceedance probabilities of 0.0010.999 at a step of 0.001. Finally,
for each ungauged catchment, the calibrated parameter values,

Fig. 11. Scatterplots between observed (Qo) and Gamma-distribution simulated (Qs) daily runoff at 99th percentile (left), 95th percentile (middle) and 90th percentile (right)
for the 228 catchments for the 228 catchments. Key as per Fig. 6.

Fig. 12. Comparing the FDC modelling performance among the three modes: the 5-IDW predicted mode (yellow, same as the FDC (5-IDW) shown in Figs. 35 respectively),
the mode using the observed ow duration curve (FDC) plus 5-IDW predicted exceedance probability (EP) (i.e. checking uncertainty of the index ow model), and the model
using the observed EC plus predicted FDC (i.e. checking uncertainty of EP transfer). (For interpretation of the references to colour in this gure legend, the reader is referred to
the web version of this article.)
84 Y. Zhang et al. / Journal of Hydrology 525 (2015) 7286
together with climatic forcings, were used to predict daily runoff
time series. The modelling results show that calibrating GR4J
against the predicted FDC performed much worse than GR4J cali-
bration against daily runoff time series (details not shown here).
This indicates that calibrating the RR model against the predicted
FDC is not good enough for predicting runoff time series in
ungauged catchments.
5.3. Uncertainty of the FDC approach
As foreshadowed in Section 3.3.1, the FDC approach uses the
three steps to predict daily runoff time series in ungauged catch-
ments. Uncertainty of the rst step comes from the three-parame-
ter index ow model, including the uncertainty of its model
structure and parameter tting (Eqs. 47). The uncertainty of step
2 comes from predicting EP on each day, i.e. transferring EPs from
gauged to ungauged catchments. Step 3 is a combination of steps
12. Fig. 12 compares the performance of the FDC approach in
the three modes: (I) prediction (including all the three steps), (II)
use of the observed FDC plus step 2 (i.e. checking the ow index
model uncertainty), and (III) use of the observed EP plus step 1
(i.e. checking the EP transfer accuracy). Fig. 12 clearly shows that
use of the ow index model exhibits more uncertainty than the
EP transfer for daily runoff predictions. This is because degradation
from mode III to mode I is more signicant than that from model II
to mode I. This nding provides a guideline on how to improve the
FDC approach for Australian catchments in future. More attention
should focus on improving the accuracy of predicting FDC in
ungauged catchments, including improving the FDC model struc-
ture and parameter tting.
5.4. Strength and weakness of the FDC approach
As shown in Sections 4.2 and 4.3, the major strength of the log-
normal FDC approach is predicting low ow and zero ow ratio.
This is particularly important in arid and semi arid regions where
rivers are often ephemeral. This approach can also be used for river
ecology studies and environmental ow managements (McIntyre
et al., 2005; Merz et al., 2009).
The daily observed runoff data is the only daily time series data
used in the FDC approach. Other data used are aggregated, such as
aridity index, forest ratio and standard deviation of precipitation.
Therefore, the FDC approach can replace the RR modelling
approach in ungauged catchments where there are considerable
gaps in climatic forcing data or rainfall gauges are very sparse.
Furthermore, the EP transfer using observed daily runoff time
series in step 2 (see Section 3.3.1) are less uncertain than the
FDC model and the aggregated data used for its parameter tting
(Fig. 12). This result suggests that the quality of daily runoff time
series data used in the FDC approach can be not as high as those
used for RR modelling, but they show a small impact on predicting
daily runoff. This is particularly important for some regions where
high quality daily runoff data are not available and it is possible to
use the FDC approach to more reliably predict daily runoff in
ungauged catchments.
The results from this study suggest that the FDC approach is
poorer than the RR modelling approach in predicting high ows
or oods in ungauged catchments. This is concerning since lots of
studies use various probability distributions estimating and design-
ing characteristics of spillways and embankments in ungauged and
poorly gauged catchments (Parajka et al., 2005; Post, 2009;
Samaniego et al., 2010b). The hydrological modelling approach
has been successfully used for continuous simulation of ood in
Switzerland (Reichl et al., 2009) and in UK (Samaniego et al.,
2010a). More studies are required to compare probability dis-
tribution approaches and hydrological modelling approaches for
better ood forecasting and design in ungauged catchments
(Samuel et al., 2011). It is also noted that the application of the
FDC approach depends on the availability of runoff in gauged catch-
ments. This approach will fail on days with runoff gaps in its donors.
6. Conclusions
This paper explored two approaches use of ow duration
curves and rainfallrunoff modelling to estimate daily stream-
ow time series for ungauged catchments based on extensive data-
sets for a large region of south-eastern Australia. This region covers
a wide climatic regime varying from dry to wet catchments. For
both approaches, using multiple donors noticeably improved pre-
dictions of daily runoff time series, especially for the FDC approach.
When calibrated against NSE or RMSE, the RR modelling approach
is considerably better than the FDC approach in simulating med-
ium and high ows, but the FDC approach is better at simulating
the low and zero ows. However, when calibrated against a low
ow objective function, both the FDC and RR modelling approaches
perform equally well at simulating the medium to low ows. These
results indicate that both methods can be further improved to
simulate daily hydrographs describing the range of ow metrics
in ungauged catchments. Further studies should focus on improv-
ing the accuracy of predicting FDC in ungauged catchments,
including improving the FDC model structure and parameter t-
ting. We encourage that similar studies are carried out for more
humid regions and catchments where climatic regimes are largely
different from those in south-eastern Australia.
Acknowledgements
This study is supported by the CSIRO Water for a Healthy
Country Flagship runoff estimation strategic project (support No.
R-02727-01) and the AWRA project in the WIRADA alliance
between CSIRO and the Australian Bureau of Meteorology. The
authors thank Juraj Parajka and another anonymous reviewer
and the editor for their useful comments and suggestions.
References
Allasia, D.G., da Silva, B.C., Collischonn, W., Morelli Tucci, C.E., 2006. Large basin
simulation experience in South America. In: Sivapalan, M. et al. (Eds.),
Predictions in Ungauged Basins: Promise and Progress. Iahs Publication, pp.
360370.
Bai, Y., Wagener, T., Reed, P., 2009. A top-down framework for watershed model
evaluation and selection under uncertainty. Environ. Model. Softw. 24 (8), 901
916.
Bardossy, A., 2007. Calibration of hydrological model parameters for ungauged
catchments. Hydrol. Earth Syst. Sci. 11 (2), 703710.
Bloeschl, G., Montanari, A., 2010. Climate change impacts-throwing the dice?
Hydrol. Process. 24 (3), 374381.
Bloschl, G., Sivapalan, M., 1995. Scale issues in hydrological modeling a review.
Hydrol. Process. 9 (34), 251290.
Blschl, G., Sivapalan, M., Wagener, T., 2013. Runoff Prediction in Ungauged Basins:
Synthesis across Processes, Places and Scales. Cambridge University Press, 136
162 pp.
Booker, D.J., Snelder, T.H., 2012. Comparing methods for estimating ow duration
curves at ungauged sites. J. Hydrol. 434, 7894.
Caballero, L.A., Easton, Z.M., Richards, B.K., Steenhuis, T.S., 2013. Evaluating the bio-
hydrological impact of a cloud forest in Central America using a semi-
distributed water balance model. J. Hydrol. Hydromech. 61 (1), 9-U96.
Castellarin, A., Galeati, G., Brandimarte, L., Montanari, A., Brath, A., 2004. Regional
ow-duration curves: reliability for ungauged basins. Adv. Water Resour. 27
(10), 953965.
Castellarin, A., Camorani, G., Brath, A., 2007. Predicting annual and long-term ow-
duration curves in ungauged basins. Adv. Water Resour. 30 (4), 937953.
Castiglioni, S., Castellarin, A., Montanari, A., 2009. Prediction of low-ow indices in
ungauged basins through physiographical space-based interpolation. J. Hydrol.
378 (34), 272280.
Chiew, F.H.S., Siriwardena, L., 2005. Estimation of SIMHYD parameter values for
application in ungauged catchments. In: Zerger, A., Argent, R.M. (Eds.),
MODSIM-2005. International Congress on Modelling and Simulation.
Modelling and Simulation Society of Australia and New Zealand, pp. 2883
2889.
 Y. Zhang et al. / Journal of Hydrology 525 (2015) 7286
Chiew, F.H.S., Teng, J., Vaze, J., Post, D.A., Perraud, J.M., Kirono, D.G.C., Viney, N.R.,
2009. Estimating climate change impact on runoff across southeast Australia:
method, results, and implications of the modeling method. Water Resour. Res.
45, W10414. http://dx.doi.org/10.1029/2008wr007338.
Chokmani, K., Ouarda, T.B.M.J., 2004. Physiographical space-based kriging for
regional ood frequency estimation at ungauged sites. Water Resour. Res. 40
(12), W12514. http://dx.doi.org/10.1029/2003wr002983.
Coron, L., Andreassian, V., Perrin, C., Lerat, J., Vaze, J., Bourqui, M., Hendrickx, F.,
2012. Crash testing hydrological models in contrasted climate conditions: an
experiment on 216 Australian catchments. Water Resour. Res. 48, W05552.
http://dx.doi.org/10.1029/2011wr011721.
Cutore, P., Cristaudo, G., Campisano, A., Modica, C., Cancelliere, A., Rossi, G., 2007.
Regional models for the estimation of streamow series in ungauged basins.
Water Resour. Manage. 21 (5), 789800.
Fenicia, F., Kavetski, D., Savenije, H.H.G., 2011. Elements of a exible approach for
conceptual hydrological modeling: 1. Motivation and theoretical development.
Water Resour. Res. 47, W11510. http://dx.doi.org/10.1029/2010wr010174.
Fernandez, W., Vogel, R.M., Sankarasubramanian, A., 2000. Regional calibration of a
watershed model. Hydrol. Sci. J. J. Sci. Hydrol. 45 (5), 689707.
Fu, B.P., 1981. On the calculation of the evaporation from land surface. Sci. Atmos.
Sin. 5, 2331 (in Chinese).
Ganora, D., Claps, P., Laio, F., Viglione, A., 2009. An approach to estimate
nonparametric ow duration curves in ungauged basins. Water Resour. Res.
45, W10418. http://dx.doi.org/10.1029/2008wr007472.
Goswami, M., OConnor, K.M., Bhattarai, K.P., 2007. Development of regionalisation
procedures using a multi-model approach for ow simulation in an ungauged
catchment. J. Hydrol. 333 (24), 517531.
Gupta, H.V., Kling, H., Yilmaz, K.K., Martinez, G.F., 2009. Decomposition of the mean
squared error and NSE performance criteria: implications for improving
hydrological modelling. J. Hydrol. 377 (12), 8091.
Hope, A., Bart, R., 2012. Evaluation of a regionalization approach for daily ow
duration curves in central and southern California Watersheds. J. Am. Water
Resour. Assoc. 48 (1), 123133.
Hundecha, Y., Ouarda, T.B.M.J., Bardossy, A., 2008. Regional estimation of
parameters of a rainfallrunoff model at ungauged watersheds using the
spatial structures of the parameters within a canonical physiographic- Post, D.A., Jakeman, A.J., 1999. Predicting the daily streamow of ungauged
climatic space. Water Resour. Res. 44 (1), W01427. http://dx.doi.org/10.1029/
2006wr005439.
Jeffrey, S.J., Carter, J.O., Moodie, K.B., Beswick, A.R., 2001. Using spatial interpolation
to construct a comprehensive archive of Australian climate data. Environ.
Model. Softw. 16 (4), 309330.
Kim, U., Kaluarachchi, J.J., 2008. Application of parameter estimation and
regionalization methodologies to ungauged basins of the Upper Blue Nile
River Basin, Ethiopia. J. Hydrol. 362 (12), 3956.
Kokkonen, T.S., Jakeman, A.J., Young, P.C., Koivusalo, H.J., 2003. Predicting daily
ows in ungauged catchments: model regionalization from catchment
descriptors at the Coweeta Hydrologic Laboratory, North Carolina. Hydrol.
Process. 17 (11), 22192238.
Kovacs, A., Honti, M., Zessner, M., Eder, A., Clement, A., Bloeschl, G., 2012.
Identication of phosphorus emission hotspots in agricultural catchments.
Sci. Total Environ. 433, 7488.
Lerat, J., Andreassian, V., Perrin, C., Vaze, J., Perraud, J.M., Ribstein, P., Loumagne, C.,
2012. Do internal ow measurements improve the calibration of rainfallrunoff
models? Water Resour. Res. 48, W02511. http://dx.doi.org/10.1029/
2010wr010179.
Li, H., Zhang, Y., Chiew, F.H.S., Xu, S., 2009. Predicting runoff in ungauged
catchments by using Xinanjiang model with MODIS leaf area index. J. Hydrol.
370 (14), 155162.
Li, M., Shao, Q.X., Zhang, L., Chiew, F.H.S., 2010. A new regionalization approach and
its application to predict ow duration curve in ungauged basins. J. Hydrol. 389
(12), 137145.
Li, H.Y., Zhang, Y.Q., Vaze, J., Wang, B.D., 2012. Separating effects of vegetation
change and climate variability using hydrological modelling and sensitivity-
based approaches. J. Hydrol. 420, 403418.
Li, F., Zhang, Y., Xu, Z., Teng, J., Liu, C., Liu, W., Mpelasoka, F., 2013. The impact of
climate change on runoff in the southeastern Tibetan Plateau. J. Hydrol. 505,
188201.
Li, F., Zhang, Y., Xu, Z., Liu, C., Zhou, Y., Liu, W., 2014a. Runoff predictions in
ungauged catchments in southeast Tibetan Plateau. J. Hydrol. 511, 2838.
Li, F.P., Xu, Z.X., Liu, W.F., Zhang, Y.Q., 2014b. The impact of climate change on runoff
in the Yarlung Tsangpo River basin in the Tibetan Plateau. Stoch. Env. Res. Risk
Assess. 28 (3), 517526.
McIntyre, N., Lee, H., Wheater, H., Young, A., Wagener, T., 2005. Ensemble
predictions of runoff in ungauged catchments. Water Resour. Res. 41 (12),
W12434. http://dx.doi.org/10.1029/2005wr004289.
Merz, R., Bloschl, G., 2004. Regionalisation of catchment model parameters. J.
Hydrol. 287 (14), 95123.
Merz, R., Parajka, J., Bloeschl, G., 2009. Scale effects in conceptual hydrological
modeling. Water Resour. Res. 45, W09405. http://dx.doi.org/10.1029/
2009wr007872.
Merz, R., Parajka, J., Bloeschl, G., 2011. Time stability of catchment model
parameters: implications for climate impact analyses. Water Resour. Res. 47.
Mohamoud, Y.M., 2008. Prediction of daily ow duration curves and streamow for
ungauged catchments using regional ow duration curves. Hydrol. Sci. J. J. Sci.
Hydrol. 53 (4), 706724.
85
Nash, J.E., Sutcliffe, J.V., 1970. River forecasting using conceptual models, 1. A
discussion of principles. J. Hydrol. 10, 280290.
Nester, T., Kirnbauer, R., Gutknecht, D., Bloeschl, G., 2011. Climate and catchment
controls on the performance of regional ood simulations. J. Hydrol. 402 (34),
340356.
Nester, T., Kirnbauer, R., Parajka, J., Bloeschl, G., 2012. Evaluating the snow
component of a ood forecasting model. Hydrol. Res. 43 (6), 762779.
Oudin, L., Andrassian, V., Perrin, C., Michel, C., Le Moine, N., 2008. Spatial
proximity, physical similarity, regression and ungauged catchments: a
comparison of regionalization approaches based on 913 French catchments.
Water Resour. Res. 44 (3), W03413. http://dx.doi.org/10.1029/2007wr006240.
Parajka, J., Merz, R., Bloschl, G., 2005. A comparison of regionalisation methods for
catchment model parameters. Hydrol. Earth Syst. Sci. 9 (3), 157171.
Parajka, J., Naeimi, V., Bloeschl, G., Wagner, W., Merz, R., Scipal, K., 2006.
Assimilating scatterometer soil moisture data into conceptual hydrologic
models at the regional scale. Hydrol. Earth Syst. Sci. 10 (3), 353368.
Parajka, J., Bloschl, G., Merz, R., 2007. Regional calibration of catchment models:
potential for ungauged catchments. Water Resour. Res. 43 (6).
Parajka, J., Viglione, A., Rogger, M., Salinas, J.L., Sivapalan, M., Bloeschl, G., 2013.
Comparative assessment of predictions in ungauged basins part 1: runoff-
hydrograph studies. Hydrol. Earth Syst. Sci. 17 (5), 17831795.
Patil, S., Stieglitz, M., 2012. Controls on hydrologic similarity: role of nearby gauged
catchments for prediction at an ungauged catchment. Hydrol. Earth Syst. Sci. 16
(2), 551562.
Peel, M.C., McMahon, T.A., Finlayson, B.L., 2010. Vegetation impact on mean annual
evapotranspiration at a global catchment scale. Water Resour. Res. 46, W09508.
http://dx.doi.org/10.1029/2009WR008233.
Perrin, C., Michel, C., Andreassian, V., 2003. Improvement of a parsimonious model
for streamow simulation. J. Hydrol. 279 (14), 275289.
Petheram, C., Rustomji, P., Chiew, F.H.S., Vleeshouwer, J., 2012. Rainfallrunoff
modelling in northern Australia: a guide to modelling strategies in the tropics. J.
Hydrol. 462, 2841.
Post, D.A., 2009. Regionalizing rainfallrunoff model parameters to predict the daily
streamow of ungauged catchments in the dry tropics. Hydrol. Res. 40 (5), 433
444.
catchments in SE Australia by regionalising the parameters of a lumped
conceptual rainfallrunoff model. Ecol. Model. 123 (23), 91104.
Priestley, C.H.B., Taylor, R.J., 1972. On the assessment of surface heat ux and
evaporation using large-scale parameters. Mon. Weather Rev. 100, 8192.
Reichl, J.P.C., Western, A.W., McIntyre, N.R., Chiew, F.H.S., 2009. Optimization of a
similarity measure for estimating ungauged streamow. Water Resour. Res. 45,
W10423. http://dx.doi.org/10.1029/2008wr007248.
Rianna, M., Russo, F., Napolitano, F., 2011. Stochastic index model for intermittent
regimes: from preliminary analysis to regionalisation. Nat. Hazards Earth Syst.
Sci. 11 (4), 11891203.
Samaniego, L., Bardossy, A., Kumar, R., 2010a. Streamow prediction in ungauged
catchments using copula-based dissimilarity measures. Water Resour. Res. 46,
W02506. http://dx.doi.org/10.1029/2008wr007695.
Samaniego, L., Kumar, R., Attinger, S., 2010b. Multiscale parameter regionalization
of a grid-based hydrologic model at the mesoscale. Water Resour. Res. 46,
W05523. http://dx.doi.org/10.1029/2008wr007327.
Samuel, J., Coulibaly, P., Metcalfe, R.A., 2011. Estimation of continuous streamow
in Ontario ungauged basins: comparison of regionalization methods. J. Hydrol.
Eng. 16 (5), 447459.
Shao, Q., Zhang, L., Chen, Y.D., Singh, V.P., 2009. A new method for modelling ow
duration curves and predicting streamow regimes under altered land-use
conditions. Hydrol. Sci. J. J. Sci. Hydrol. 54 (3), 606622.
Shu, C., Ouarda, T., 2012. Improved methods for daily streamow estimates at
ungauged sites. Water Resour. Res. 48, W02523. http://dx.doi.org/10.1029/
2011wr011501.
Singh, R., Wagener, T., van Werkhoven, K., Mann, M.E., Crane, R., 2011. A trading-
space-for-time approach to probabilistic continuous streamow predictions in
a changing climate accounting for changing watershed behavior. Hydrol.
Earth Syst. Sci. 15 (11), 35913603.
Sivapalan, M., Takeuchi, K., Franks, S.W., Gupta, V.K., Karambiri, H., Lakshmi, V.,
Liang, X., McDonnell, J.J., Mendiondo, E.M., OConnell, P.E., Oki, T., Pomeroy, J.W.,
Schertzer, D., Uhlenbrook, S., Zehe, E., 2003. IAHS decade on Predictions in
Ungauged Basins (PUB), 20032012: shaping an exciting future for the
hydrological sciences. Hydrol. Sci. J. J. Sci. Hydrol. 48 (6), 857880.
Smakhtin, V.Y., 1997. Regional low-ow studies in South Africa. In: Gustard, A.,
Blazkova, S., Brilly, M. (Eds.), FRINEDS97, Regional Hydrology: Concepts and
Models for Sustainable Water Resource Management. IAHS Publication 246,
Wallingford, pp. 125132.
Teng, J., Chiew, F.H.S., Vaze, J., Marvanek, S., Kirono, D.G.C., 2012. Estimation of
climate change impact on mean annual runoff across continental Australia
using Budyko and Fu equations and hydrological models. J. Hydrometeor. 13,
10941106.
Tuteja, N.K., Vaze, J., Teng, J., Mutendeudzi, M., 2007. Partitioning the effects of pine
plantations and climate variability on runoff from a large catchment in
southeastern Australia. Water Resour. Res. 43 (8), W08415. http://dx.doi.org/
10.1029/2006WR005016.
Vaze, J., Teng, J., 2011. Future climate and runoff projections across New South
Wales, Australia results and practical applications. Hydrol. Process. 25 (1), 18
35. http://dx.doi.org/10.1002/hyp.7812.
86 Y. Zhang et al. / Journal of Hydrology 525 (2015) 7286
Vaze, J., Chiew, F.H.S., Perraud, J.M., Viney, N., Post, D.A., Teng, J., Wang, B.D., Lerat, J.,
Goswami, M., 2010a. Rainfallrunoff modelling across southeast Australia:
datasets, models and results. Aust. J. Water Resour. 14 (2), 101116.
Vaze, J., Post, D.A., Chiew, F.H.S., Perraud, J.M., Viney, N.R., Teng, J., 2010b. Climate
non-stationarity validity of calibrated rainfallrunoff models for use in
climate change studies. J. Hydrol. 394 (34), 447457.
Vaze, J., Post, D.A., Chiew, F.H.S., Perraud, J.-M., Teng, J., Viney, N., 2011. Conceptual
rainfallrunoff model performance with different spatial rainfall inputs. J.
Hydrometeor. 12 (5), 11001112. http://dx.doi.org/10.1175/2011JHM1340.1.
Viney, N.R., Perraud, J., Vaze, J., Chiew, F.H.S., Post, D.A., Yang, A., 2009. The
usefulness of bias constraints in model calibration for regionalisation to
ungauged catchments, MODSIM. In: 18th World IMACS/MODSIM Congress,
Cairns, pp. 34213427.
Vogel, R.M., Fennessey, N.M., 1995. Flow duration curves. II: a review of applications
in water-resources planning. Water Resour. Bull. 31 (6), 10291039.
Wagener, T., McIntyre, N., 2005. Identication of rainfallrunoff models for
operational applications. Hydrol. Sci. J. J. Sci. Hydrol. 50 (5), 735751.
Wang, E.L., Zhang, Y.Q., Luo, J.M., Chiew, F.H.S., Wang, Q.J., 2011. Monthly and
seasonal streamow forecasts using rainfallrunoff modeling and historical
weather data. Water Resour. Res. 47, W05516. http://dx.doi.org/10.1029/
2010wr009922.
Westerberg, I.K., Guerrero, J.L., Younger, P.M., Beven, K.J., Seibert, J., Halldin, S., Freer,
J.E., Xu, C.Y., 2011. Calibration of hydrological models using ow-duration
curves. Hydrol. Earth Syst. Sci. 15 (7), 22052227.
Zhang, Y.Q., Chiew, F.H.S., 2009. Relative merits of different methods for runoff
predictions in ungauged catchments. Water Resour. Res. 45, W07412. http://
dx.doi.org/10.1029/2008WR007504.
Zhang, Y.Q., Chiew, F.H.S., 2012. Estimation of mean annual runoff across Southeast
Australia by incorporating vegetation types into Budyko framework. Aust. J.
Water Resour. 15 (2), 109120.
Zhang, L., Hickel, K., Dawes, W.R., Chiew, F.H.S., Western, A.W., Briggs, P.R., 2004. A
rational function approach for estimating mean annual evapotranspiration.
Water Resour. Res. 40 (2), W02502. http://dx.doi.org/10.1029/2003wr002710.
Zhang, Y.Q., Leuning, R., Hutley, L.B., Beringer, J., McHugh, I., Walker, J.P., 2010.
Using long-term water balances to parameterize surface conductances and
calculate evaporation at 0.05 degrees spatial resolution. Water Resour. Res. 46,
W05512. http://dx.doi.org/10.1029/2009wr008716.
Zhang, Y.Q., Viney, N.R., Chiew, F.H.S., van Dijk, A., Liu, Y.Y., 2011. Improving
hydrological and vegetation modelling using regional model calibration
schemes together with remote sensing data. In: 19th International Congress
on Modelling and Simulation, pp. 34483454.
Zhang, Y., Vaze, J., Chiew, F.H., Teng, J., Li, M., 2014. Predicting hydrological
signatures in ungauged catchments using spatial interpolation, index model,
and rainfallrunoff modelling. J. Hydrol. 517, 936948.
Zhou, Y.C., Zhang, Y.Q., Vaze, J., Lane, P., Xu, S.G., 2013. Improving runoff estimates
using remote sensing vegetation data for bushre impacted catchments. Agric.
For. Meteorol. http://dx.doi.org/10.1016/j.agrformet.2013.04.018.