Journal of Hydrology 544 (2017) 479–487

Contents lists available at ScienceDirect

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

Research papers

Developing and testing a global-scale regression model to quantify mean

annual streamflow
Valerio Barbarossa a,⇑, Mark A.J. Huijbregts a,b, A. Jan Hendriks a, Arthur H.W. Beusen c, Julie Clavreul d,
Henry King d, Aafke M. Schipper b
a
Radboud University, Institute for Water and Wetland Research, Department of Environmental Science, P.O. Box 9010, 6500 GL Nijmegen, The Netherlands
b
PBL Netherlands Environmental Assessment Agency, Department of Nature and Rural Areas, P.O. Box 303, 3720 AH Bilthoven, The Netherlands
c
PBL Netherlands Environmental Assessment Agency, Department of Information, Data and Methodology, P.O. Box 303, 3720 AH Bilthoven, The Netherlands
d
Unilever R&D, Safety and Environmental Assurance Centre, Colworth Science Park, Sharnbrook, Bedfordshire MK44 1LQ, UK

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

Article history: Quantifying mean annual flow of rivers (MAF) at ungauged sites is essential for assessments of global
Received 20 July 2016 water supply, ecosystem integrity and water footprints. MAF can be quantified with spatially explicit
Received in revised form 20 September process-based models, which might be overly time-consuming and data-intensive for this purpose, or
2016
with empirical regression models that predict MAF based on climate and catchment characteristics.
Accepted 24 November 2016
Available online 25 November 2016
Yet, regression models have mostly been developed at a regional scale and the extent to which they
This manuscript was handled by Corrado can be extrapolated to other regions is not known. In this study, we developed a global-scale regression
Corradini, Editor-in-Chief, with the model for MAF based on a dataset unprecedented in size, using observations of discharge and catchment
assistance of Subashisa Dutta, Associate characteristics from 1885 catchments worldwide, measuring between 2 and 106 km2. In addition, we
Editor compared the performance of the regression model with the predictive ability of the spatially explicit glo-
bal hydrological model PCR-GLOBWB by comparing results from both models to independent measure-
Keywords: ments. We obtained a regression model explaining 89% of the variance in MAF based on catchment area
Mean annual discharge and catchment averaged mean annual precipitation and air temperature, slope and elevation. The regres-
River discharge sion model performed better than PCR-GLOBWB for the prediction of MAF, as root-mean-square error
Global hydrology (RMSE) values were lower (0.29–0.38 compared to 0.49–0.57) and the modified index of agreement (d)
Empirical modelling
was higher (0.80–0.83 compared to 0.72–0.75). Our regression model can be applied globally to estimate
Predictions in ungauged basins
MAF at any point of the river network, thus providing a feasible alternative to spatially explicit process-
Scaling relationships
Model comparison based global hydrological models.
PCR-GLOBWB Ó 2016 Elsevier B.V. All rights reserved.
Spatial error model

1. Introduction toring is in rapid decline since the mid-1980s (Shiklomanov

et al., 2002). Modelling approaches have long been used to esti-
Mean annual discharge or flow of rivers (hereafter abbreviated mate MAF at ungauged sites and are generally divided into two
as MAF) is an important indicator of global water supply, with categories: spatially explicit process-based models and
applications in irrigation supply assessment, climate change vul- regression-based empirical models.
nerability assessment (Chang, 2003; Santini and di Paola, 2015), State-of-the-art spatially explicit numerical models for global-
hydropower assessment (Hall et al., 2004), water footprinting scale calculations of streamflow are Macroscale Hydrological Mod-
(Hanafiah et al., 2011; Hoekstra et al., 2011; Jefferies et al., 2012; els (MHM) or Global Hydrology and Water Resources Models
Pfister et al., 2009; Tendall et al., 2014), and for quantifying sedi- (GHWM) (Alcamo et al., 2003; Gosling and Arnell, 2011;
ment fluxes (Syvitski et al., 2003). It also represents one of the most Hanasaki et al., 2008; van Beek and Bierkens, 2008; Van Der
important factors determining the integrity of freshwater biodiver- Knijff et al., 2010; Widén-Nilsson et al., 2007; Wisser et al.,
sity (Oberdorff et al., 1995, 2011; Poff and Zimmerman, 2010; 2010). As these models account for the spatial variability of the
Xenopoulos and Lodge, 2006; Xenopoulos et al., 2005). Despite physical processes involved within catchment hydrology and are
its importance, streamflow data availability is limited, and moni- capable of predicting streamflow even at the daily time scale, they
are computationally and data intensive.
Regression-based approaches to calculate MAF are less time-
⇑ Corresponding author.
consuming and computationally less intensive. Moreover, regres-
E-mail address: vbarbarossa@science.ru.nl (V. Barbarossa).

http://dx.doi.org/10.1016/j.jhydrol.2016.11.053
0022-1694/Ó 2016 Elsevier B.V. All rights reserved.
480 V. Barbarossa et al. / Journal of Hydrology 544 (2017) 479–487

sion equations relating streamflow to explanatory catchment char- elevation of the catchment. Drainage area, mean annual precipita-
acteristics like upstream drainage area, precipitation and tempera- tion and mean annual temperature are often used as predictors of
ture may help to better understand general hydrological patterns MAF in regional regression modelling studies (Verdin and Worstell,
and processes across different scales (Burgers et al., 2013; Farmer 2008; Vogel and Sankarasubramanian, 2000; Vogel et al., 1999).
et al., 2015). However, to date, regression-based approaches relat- The dependence of MAF on drainage area is a well-accepted power
ing mean annual streamflow to catchment characteristics have relationship reflecting the self-similarity of river systems
been mainly applied at a regional scale (Hortness and (Rodríguez-Iturbe and Rinaldo, 2001). Mean annual precipitation
Berenbrock, 2001; Stuckey, 2006; Tran et al., 2015; Verdin and represents the potential runoff of the catchment, as it equals the
Worstell, 2008; Vogel et al., 1999) or to specific climate zones amount of water supplied to the catchment (Thomas and Benson,
(Syvitski et al., 2003), and the extent to which these models can 1970). We selected the mean annual temperature as a proxy for
be extrapolated to other regions is not known. Regression relation- the potential evapotranspiration (PET), because temperature is a
ships at the global scale have hardly been established so far. An major determinant of evapotranspiration (Hamon, 1963; Lu et al.,
exception is Burgers et al. (2013), who derived MAF relationships 2005; Thornthwaite, 1948). Furthermore, previous regression anal-
at a global scale using precipitation and catchment area as predic- yses of MAF have shown an increased explained variance when
tors. However, their model explained only 56% of the variance in additional geomorphologic parameters were considered
MAF, which is low compared to the range of 77–99% achieved by (Hortness and Berenbrock, 2001; Stuckey, 2006; Vogel et al.,
regional regression models (e.g. Verdin and Worstell (2008)). Yet, 1999). Therefore, we included average slope and elevation of the
the regional studies typically included a larger number of predic- catchment as additional predictors in our study. Although eleva-
tors, which suggests that the explanatory power of a global-scale tion and slope alone may not directly influence MAF, they may
regression model may increase if relevant predictors are added. serve as proxies for other factors causing inter-basin streamflow
In addition, the applicability of global regression relationships for variation which are difficult to measure, e.g. radiation, wind, vege-
the prediction of mean annual streamflow has not yet been tested. tation and basin ruggedness (Thomas and Benson, 1970).
Therefore, the aim of this study was twofold: (1) to establish an
empirical regression model relating MAF to easily retrievable
catchment characteristics at the global scale; (2) to test the predic- 2. Materials and methods
tive ability of the regression model in a backcasting analysis and
compare its performance with the predictive performance of 2.1. Mean annual discharge data
PCR-GLOBWB, a spatially explicit MHM (van Beek et al., 2011).
To our knowledge, our study is the first to make an explicit com- We retrieved worldwide MAF data from the Global Runoff Data
parison of the predictive abilities of a process-based and a Centre (GRDC) database, which provides daily or monthly observa-
regression-based global-scale model for MAF. tions of 9213 gauging stations monitored from 1806 to 2015, with
We based our regression model on measured long-term average variable record length (GRDC, 2015). The GRDC has spent more
MAF from 1885 catchments worldwide, ranging from 2 km2 to than 25 years gathering river discharge data from the National
106 km2 in size. We used five predictor variables, including two cli- Hydrological Services of all the World Meteorological Organization
matic variables – mean annual precipitation and air temperature – (WMO) state members, which has resulted in a discharge dataset
and three geomorphologic variables – area, mean slope and mean unprecedented in size. For example, the SAGE Global River

Fig. 1. Distribution of the 1885 GRDC gauging stations monitored for at least 15 years in the 1981–2010 period. The stations are grouped based on the mean annual flow
(MAF) recorded at each station. Next to each MAF category, the number of observations is provided in brackets.
 V. Barbarossa et al. / Journal of Hydrology 544 (2017) 479–487 481

Discharge Database (http://nelson.wisc.edu/sage/data-and-mod- watershed boundaries were established. We then calculated a sin-
els/riverdata/) and the RivDis database (Vorosmarty et al., 1998) gle mean value of each variable for each catchment corresponding
provide discharge data for 3500 and 1018 stations, respectively. with a GRDC gauging station, as described by Syvitski et al. (2003).
The accuracy of the discharge measurements included in the GRDC Resampling and averaging were performed in ArcGIS 10.3. An over-
database is estimated to be about 10–20% (Syvitski et al., 2005). For view of the summary statistics of the variables is available in
our model development we selected discharge data for the period Table 1.
1981–2010. We used a 30 year period because this is in accordance
with the recommendations for climate analyses (World 2.3. Model fitting
Meteorological Organization, 1992). We excluded years after
2010 because of a decrease in data availability for the most recent Methods available for correlative modelling range from para-
years. We averaged the daily discharge data over each year, using metric and non-parametric regression-based approaches to
only those years where 100% of the daily observations were avail- machine-learning techniques (Chen et al., 2015; Danandeh Mehr
able. Next, we averaged the yearly observations for the period et al., 2013; Fan et al., 2015; Okkan and Serbes, 2012; Wang
1981–2010 in order to obtain a long-term mean annual discharge et al., 2015; Wu et al., 2009). For the present study we selected
for each catchment. We selected monitoring stations with at least ordinary least squares (OLS) regression because it results in an
15 yearly average discharge values in order to obtain representa- explicit equation, which facilitates interpretation and comparison
tive long-term mean values (Kennard et al., 2010). This resulted with other studies. All the variables except for temperature were
in a dataset of 1885 observations out of the 2759 available GRDC log-transformed to avoid heteroscedasticity as they revealed a
gauged catchments with observations in the time range 1981– right-skewed distribution (Table 1), in agreement with the choices
2010 (Fig. 1). made in previous studies for similar variables (Burgers et al., 2013;
Hendriks et al., 2012; Syvitski et al., 2003; Verdin and Worstell,
2.2. Catchment characteristics 2008; Vogel et al., 1999). This resulted in the following linear
regression equation:
We retrieved catchment-specific values for catchment area (A)
and catchment-averaged altitude (H), slope (S), 30-years mean log10 Q ¼ b0 þ bA  log10 A þ bP  log10 P þ bT  T þ bS  log10 S
temperature (T) and 30-years mean precipitation (P) from a com- þ bH  log10 H þ e ð1Þ
bination of data sources (Table 1). Catchment area was retrieved
from the GRDC database, which includes a georeferenced map of where b0 is the intercept, <bA, bP, bT, bS, bH> is the vector of the
the upstream catchment corresponding with each gauging station regression coefficients associated with the predictor variables (sym-
(GRDC, 2011). Catchment boundaries in this map have been estab- bols are described in Table 1) and e is the error term. The back-
lished based on the HydroSHEDS drainage network, a 15 arc- transformed form to real scale of Eq. (1) yields the nonlinear
seconds hydrological map derived from 3 arc-seconds elevation formulation:
data of the National Aeronautics and Space Administration (NASA)
Q ¼ 10b0  AbA  PbP  10bT T  SbS  HbH  10e ð2Þ
Shuttle Radar Topography Mission (SRTM), extended with the
hydro1k hydrological network for latitudes above 60 N, which Prior to performing the regression analysis, we assessed multi-
are not covered by the SRTM data (GRDC, 2011; Lehner et al., collinearity among the predictors using Variance Inflation Factors
2008). (VIFs), employing the function ‘‘vif” of the package ‘‘HH”
We derived altitude from the WorldClim digital elevation (Heiberger, 2015) in the R environment (Development Core
model (DEM), which is a 30 arc-seconds DEM based upon the SRTM Team, 2005). We preferred VIFs over bivariate correlation analysis
elevation data extended with GTOPO30 elevation data for latitudes because pairwise correlation coefficients do not reveal more subtle
above 60 N (Hijmans et al., 2005). We derived a raster slope map forms of multicollinearity (Field, 2009). The maximum VIF was 2.8,
from the WorldClim DEM using the ‘average maximum technique’ well below the standard threshold of 5 (Zuur et al., 2009). We then
in ArcGIS, similarly to Hortness and Berenbrock (2001). For precip- fitted OLS regression models for an increasing number of predic-
itation and temperature, we averaged 30 arc-minutes resolution tors. To identify the best regression model for each of the five sets
monthly raster maps from the Climatic Research Unit time series of predictors (first set with one predictor variable, second set with
(CRU TS) 3.23 to annual values, and consequently averaged over two predictor variables, etc.) as well as the best overall model, we
the period 1981–2010 (Harris et al., 2014; University of East employed the function ‘‘dredge” of the package ‘‘MuMIn” in the R
Anglia Climatic Research Unit, 2013). environment (Barton, 2015). Within each set, the algorithm ana-
We resampled the raster maps obtained for precipitation, tem- lyzes all possible combinations of predictor variables and ranks
perature, altitude and slope in order to match the 15 arc-seconds the regression models based on a user-defined criterion. To iden-
resolution of the HydroSHEDS drainage network, upon which the tify the most parsimonious model for each set of predictors, we

Table 1
Summary statistics of the mean annual streamflow (MAF) and catchment characteristics of 1885 gauging stations in the period 1981–2010.

Variable Symbol Unit Mean Median SDa Mina Maxa c1a Source database
3 1 2 1 3 3 4
MAF Q m s 3.72  10 3.38  10 1.99  10 3.47  10 4.73  10 14.2 GRDCb
Catchment area A m2 4.58  1010 4.61  109 2.05  1011 2.00  106 3.63  1012 9.7 GRDCc
Altitude H m 6.84  102 4.52  102 6.05  102 1.35  101 4.76  103 1.6 WorldClimd
Slope S (°) 2.31  100 1.27  100 2.65  100 4.71  102 1.56  101 2.0 WorldClimd
Precipitation P m  s1 2.78  108 2.46  108 1.46  108 3.53  109 1.09  107 1.3 CRU TS 3.23e
Temperature T °C 8.96  100 8.24  100 7.51  100 1.67  101 2.76  101 0.4 CRU TS 3.23e
a
SD = standard deviation; Min = minimum; Max = maximum; c1 = skewness.
b
GRDC (2015).
c
GRDC (2011).
d
Hijmans et al. (2005).
e
Harris et al. (2014).
482 V. Barbarossa et al. / Journal of Hydrology 544 (2017) 479–487

used the Akaike Information Criterion (AIC) as well as the Bayesian an absolute as well as a relative error measure, following the rec-
Information Criterion (BIC), which employs a larger penalty term ommendations for hydrological model evaluation as provided by
for additional predictor variables. Further, we used the Cooks D Legates and McCabe (1999). We adopted the index of agreement
influence statistic in order to identify observations that may have d2 in the modified form d to avoid inflation of errors by squared
biased the coefficients of the regression (Cook and Weisberg, values (Legates and McCabe, 1999). In addition, the d represents
1982). an improvement over the coefficient of determination (R2)
In order to assess potential bias in the regression coefficients (Legates and McCabe, 1999).
induced by spatial autocorrelation resulting from the nested struc- Given the six orders of magnitude covered by the data, we log
ture of the catchments, we compared the regression coefficients transformed the mean annual streamflow values. The RMSE is cal-
with the coefficients of a spatial error (SE) model. Since spatial culated as:
autocorrelation can exist within either the residuals (spatial error) rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
or the response variable (spatial lag), we performed a preliminary 1 Xm Xn 2
RMSE ¼  ðlog10 Ox;t  log10 Px;t Þ ð3Þ
test for spatial autocorrelation based on the Lagrange Multiplier nm x¼1 t¼1

test (LM test) using the R package ‘spdep’ (Anselin, 1988; Bivand where n  m are the dimensions of the matrix of observations of the
et al., 2013; Bivand and Piras, 2015). We preferred the LM test to m GRDC stations over the n years of the backcasting period, Ox,t is
the more commonly employed Moran’s I test, for the LM test has the observed value for the station x at time t, and Px,t is the pre-
a higher power to discriminate among either spatial error autocor- dicted value for the station x at time t.
relation or spatial lag (Anselin and Rey, 1991). The LM test showed The d is formulated as:
significantly higher autocorrelation in the error term (LM test
value of 465, robust LM test value of 375) than in the response X
m
Pn
t¼1 jlog10 Ox;t  log10 Px;t j
variable (with values of 115 and 25, respectively). Therefore we fit-
d¼P  x¼1  ð4Þ
ted an SE model that accounts for spatial autocorrelation in the m Pn
residuals, expressing the error term of Eq. (1) as: e = kWe + l, x¼1 t¼1 jlog10 P x;t  log10 Oj þ jlog10 Ox;t  log10 Oj

where k is the coefficient in the spatial autoregressive structure,

where O stands for the overall average of the observed streamflow
W is a weight matrix defined by the inverse distance between
of all the stations across the 10 years period. The index of agree-
observations, and l is the vector of identically distributed random
ment varies between 0 and 1, with higher values indicating a better
errors (Ord, 1975). We calculated the distances between the GRDC
fit. When the mean annual streamflow averaged over the 10 years is
stations across the stream network thereby considering as neigh-
considered, n becomes 1 in Eqs. (3) and (4).
bors only those stations belonging to the same encompassing
hydrologic basin. As we employed HydroSHEDS as rivers network
for the calculation of W, we fitted and compared the OLS and SE 3. Results
regression coefficients based on a subset of observations within
58S–60 N latitude (n = 1748). 3.1. Model fitting

2.4. Comparison with PCR-GLOBWB: backcasting analysis The OLS regression analysis revealed that the model with the
full set of predictors was the most parsimonious (i.e., lowest AIC;
We compared the predictive performance of our best regression Table S1). Nearly 90% of the variation in MAF could be explained
model with the global hydrological model PCR-GLOBWB (van Beek by the five catchment characteristics (Table 2), indicating that
and Bierkens, 2008; van Beek et al., 2011). Defined as a ‘‘leaky the most relevant predictors for MAF were covered by the regres-
bucket” type of model, PCR-GLOBWB calculates changes in water sion model. According to the standardized regression coefficients,
storage between two different soil layers, groundwater reservoir which can be compared across explanatory variables to assess their
and atmosphere, forced by CRU TS 2.1 data, on a cell-by-cell basis relative importance (Bring, 1994), catchment area was the most
at 30 arc-minutes resolution, for daily time steps. PCR-GLOBWB important predictor of MAF, followed by precipitation, tempera-
has been widely employed for assessments of global surface water ture, slope and elevation.
and groundwater availability, nutrient transport modelling and The model performed better for higher MAF values (Fig. 2). Fur-
biodiversity impact calculations (Beusen et al., 2016; Gleeson thermore, residual errors were slightly larger for catchments with
et al., 2012; Janse et al., 2015; Wada et al., 2011; Wanders and lower precipitation values and at higher altitudes (Fig. S1). Residu-
Wada, 2015). Compared to other GHMs, PCR-GLOBWB is a purely als tended to be randomly distributed in relation to catchment
process-based model, as opposed to for example WaterGAP which area, precipitation or slope (Fig. S1). Only about 1.3% of the pre-
is partially calibrated (Alcamo et al., 2003; Döll et al., 2003). There- dicted values showed errors greater than one order of magnitude
fore, we considered PCR-GLOBWB a more suitable benchmark for (Fig. 2). For the Cooks D statistic, a maximum value of 0.03 was
the comparison. found, well below the threshold of 1 (Cook and Weisberg, 1982),
We considered monitoring data of GRDC stations continuously meaning that none of the observations biased the regression
monitored from 1971 to 1980 as an independent and common coefficients.
basis for the comparison between the regression model and PCR- The comparison of the regression coefficients between the OLS
GLOBWB. From the 2219 GRDC stations used for the testing of and SE regression models revealed a large overlap of the confi-
PCR-GLOBWB (van Beek et al., 2011), we selected the 543 stations dence intervals of the coefficients (49–86% CI overlap; Table S2).
that were continuously monitored from 1971 through 1980 This indicated that the OLS regression coefficients were not signif-
(Fig. S2). We derived mean annual values of temperature and pre- icantly influenced by spatial autocorrelation (type I error).
cipitation from the CRU TS 3.23 for each catchment, according to
the approach described in Section 2.2, and calculated mean annual 3.2. Performance testing on independent data and comparison with
streamflow for each year in the time span 1971–1980, and as a PCR-GLOBWB
10 years average.
We evaluated and compared the performances of the regression The testing of the regression model on independent data in the
model and PCR-GLOBWB employing root mean square error time period 1971–1980 (backcasting analysis), for both single year
(RMSE) and modified index of agreement (d). Thus, we employed and 10-years average MAF, revealed that the predictions of the
 V. Barbarossa et al. / Journal of Hydrology 544 (2017) 479–487 483

Table 2
Coefficients (raw and standardized), goodness of fit (R2) and number of underlying observations (m) of the most parsimonious regression model Q = 10b0  AbA  PbP  10bTT 
HbH  SbS. CI = confidence interval.

Coefficient Value (95% CI) Std. value R2 m

b0 9.066 (8.503 to 9.630) 0 0.89 1885
bA 1.018 (1.001 to 1.035) 0.961
bP 2.070 (1.991 to 2.148) 0.486
bT 0.038 (0.040 to 0.035) 0.290
bH 0.509 (0.565 to 0.454) 0.212
bS 0.464 (0.421 to 0.507) 0.237

on catchment area and catchment-averaged precipitation, temper-

ature, slope and altitude. The analysis revealed the catchment area
to be the most important predictor of MAF (Table 2). Indeed, a sin-
gle regression analysis based on catchment area alone already
explains 61% of the variation in MAF (Table S1), in agreement with
scaling relationships reflecting self-similarity across catchments.
Nevertheless, additional predictors considerably increased the
explained variance (Table S1). This suggests that multiple regres-
sion improves the interpretation of the spatial scaling of MAF at
a global level, in agreement with recent findings that point toward
multiple regression (called ‘‘multiscaling”) to improve the inter-
pretation of scaling behavior for daily streamflow of the Southeast
United States (Farmer et al., 2015).
The MAF scaled to catchment area with an exponent of about 1,
implying a linear relationship between MAF and drainage area.
This is in agreement with the coefficients reported by a number
of regional studies using multiple regression analysis (Table 3),
and close to the value of 0.86 reported by the global study of
Burgers et al. (2013). Further, MAF scaled to precipitation with
Fig. 2. Predicted versus observed MAF values of the 1885 GRDC stations employed an exponent of about two. This is highly similar to the findings
in the regression analysis. The solid line represents perfect model fit (1:1 line) and
reported by Tran et al. (2015) (Table 3), who conducted a regional
the dashed lines represent a difference of one order of magnitude.
study covering 533 catchments. On the other hand, the coefficient
for precipitation found in this study is in contrast with the linear
regression model were characterized by lower RMSE values and relationship observed in Burgers et al. (2013). However, in the cur-
higher d values than the predictions of the global hydrological rent study we employed about three times more observations and
model PCR-GLOBWB (Fig. 3). our regression model explained about 33% more of the variance. In
The PCR-GLOBWB simulation resulted in a greater number of general, coefficients reported for precipitation tend to be greater
outliers, in agreement with the higher RMSE values (Fig. 3). For than 1 and converging towards 2 (Table 3), reflecting nonlinearity
the OLS model, about 3% of the observations had residuals greater in the physical process responsible for the runoff generation (see
than one order of magnitude, whereas about 7% of the PCR- e.g. Yu et al. (2015)). Therefore, given the global coverage, the
GLOBWB results deviated more than one order of magnitude from heterogeneity of the input variables values employed to calibrate
the measurements. PCR-GLOBWB performed slightly better than the coefficients and the large amount of explained variance, the
the OLS model for the highest MAF values (>10,000 m3/s; Fig. 3). value of about 2 is considered to be a reasonable estimate of the
Both models performed poorly for MAF values lower than 10 m3/ exponent for precipitation.
s. In general, residuals of PCR-GLOBWB revealed a cone-shaped For temperature a negative coefficient was obtained, reflecting
distribution, with larger errors at lower discharge values. In con- decreasing MAF with increasing temperature. Indeed, increases in
trast, the OLS model showed a tendency to overestimate low dis- temperature would lead to an increased evapotranspiration, which
charge values. eventually implies that less water is routed through the drainage
Both models performed better when backcasting the 10-years network and poured at the discharge point. The negative relation-
average of MAF than MAF for single years. For the 10-years average ship is in agreement with the results of regional studies, but as
MAF, the residuals greater than one order of magnitude reduced to these studies were based on log-transformed temperature data,
about 1% (4 observations) and 4% (22 observations) for the regres- the values of the coefficients are not directly comparable with ours
sion model and PCR-GLOBWB, respectively. (Tran et al., 2015; Verdin and Worstell, 2008; Vogel et al., 1999).
The relationship between MAF and altitude is rather complex.
The exponent resulting from the regression analysis was negative,
4. Discussion which can be explained by the fact that at higher altitudes the solar
radiation and wind are more intense, therefore enhancing the
4.1. Regression coefficients interpretation evapotranspiration process (Tran et al., 2015). In contrast, some
studies reported positive exponents for altitude (Table 3). A possi-
We developed a global-scale multiple regression model for pre- ble explanation could be that these studies did not include temper-
dicting mean annual flow of rivers based on easily retrievable input ature as a predictor, which implies that the positive exponent for
parameters. We calibrated the model on long-term average dis- the altitude term may reflect an effect of temperature. At high alti-
charge data (1981–2010) from 1885 catchments worldwide, tudes, the lower temperature likely results in less evapotranspira-
resulting in a model explaining 89% of the variance in MAF based tion hence an increase in MAF. However, if the model does not
484 V. Barbarossa et al. / Journal of Hydrology 544 (2017) 479–487

Fig. 3. Results of the backcasting analysis for the period 1971–1980, showing predicted versus observed MAF for the OLS regression model (left) and the GHM PCR-GLOBWB
(right), based on yearly values (top) as well as 10-year average values (bottom). Within each chart, the solid line represents perfect model fit (1:1 line) and the dashed lines
define a range of accuracy of plus/minus one order of magnitude. RMSE = root mean square error; d = modified index of agreement.

include temperature as predictor, the larger MAF at higher alti- area (p-value < 0.01, R2 = 0.03), while for the regression model no
tudes may result in a positive regression coefficient for altitude significant correlation was found. However, new global hydrologi-
instead. The positive exponent of about ½ obtained for slope (S) cal models with greater spatial resolution than the 30 arc-minutes
compares well with the values reported in regional studies and version of PCR-GLOBWB employed in this study may achieve bet-
reflects that in catchments with steeper slopes, the runoff and con- ter results, especially for smaller catchments (e.g. see the list of
sequently the MAF is enhanced. models provided in Bierkens (2015)). Yet, such refined models
are more demanding in terms of computational costs (Bierkens,
4.2. Performance comparison with PCR-GLOBWB 2015), and might therefore be more suitable when monthly or
daily discharge values are needed.
Overall, our regression model performed better than PCR- Both the regression model and PCR-GLOBWB performed worse
GLOBWB when applied to an independent test dataset. Differences for water-scarce regions, as revealed by larger errors at higher dry-
were apparent in particular for smaller catchments (Fig. 3). The ness ratio values (Fig. 4). The dryness ratio reflects water losses due
residuals (absolute values) of PCR-GLOBWB for the 10-years to evapotranspiration relative to the amount of precipitation. It is
average MAF revealed a significant negative trend in relation to defined as the actual annual evapotranspiration divided by the
 V. Barbarossa et al. / Journal of Hydrology 544 (2017) 479–487 485

Table 3
Regression coefficients found in this study compared with coefficients reported in regional and global studies available from the literature. n represents the number of catchments
employed to calibrate the regression coefficients; Area range is the range of the catchment areas employed in the respective study.

This study Burgers et al. Gyawali et al. (2015) Hortness and Berenbrock Stuckey (2006) Tran et al. (2015) Vogel et al.
(2013) (2001) (1999)
bA 1.02 0.86 0.87 0.83 to 1.10 1.01 1.01 0.58 to 1.14
bP 2.07 1.01 3.68 1.64 to 2.70 1.80 2.04 1.21 to 6.42
bT 0.04 – – – – 0.49a 7.66 to
0.51a
bS 0.46 – 0.30b 3.44 to 7.52b – – 0.33 to 0.51
bH 0.51 – – 2.36 to 2.30 0.13 – 1.66c
n 1885 663 93 200 195 533 1553
Area range 2.0100 to 7.3  103 to 6.2  100 to 4.4  103 8.0  100 to 3.5  104 5.6  100 to 1.3  101 to –
[km2] 3.6106 4.6  106 4.5  103 1.7  102
Extent Global Global western Great Lakes, Idaho, USA Pennsylvania, Upper Mississippi, USA
USA USA USA
a
Values refer to the log-transformed form and are therefore not directly comparable with the coefficient obtained in this study.
b
Values refer to slope in percent instead of degrees and are therefore not directly comparable with the coefficient obtained in this study.
c
Used only for region 16, ‘‘Great Basin” in Vogel et al. (1999).

Fig. 4. Residuals plot of the backcasting of the 10-year average MAF vs the dryness ratio for (a) the regression model and (b) PCR-GLOBWB. Note that 18 values (about 3% of
total data) with dryness ratios smaller than 0 were excluded for clarity of representation.

total annual precipitation, where the actual evapotranspiration is tion and temperature values (Table 1). This indicates that the model
calculated as the annual precipitation minus the unit discharge, developed by this study can be applied to predict MAF at any point
in turn obtained by dividing the discharge by the area of the catch- of any river network globally, taking into account the weaker pre-
ment (Vogel et al., 1999). The larger errors with higher dryness diction power for water-scarce regions (see Section 4.2). As such,
ratios are likely due to a combination of higher uncertainty in the model is suitable for assessments of water availability and eco-
the precipitation values for water-scarce regions and hydrological logical integrity in relation to changing future climatic conditions.
processes that are particularly relevant in dry regions yet not The model requires only a small number of input parameters,
described by the models (Döll et al., 2003; van Beek et al., 2011; namely catchment area, catchment-averaged precipitation,
Vogel et al., 1999). Examples of such processes include the almost catchment-averaged temperature, catchment-averaged altitude
instantaneous evaporation from many ephemeral post-rainfall and catchment-averaged slope. Catchment area as well as slope
ponds and relatively large losses from the river channel to ground- and altitude can be easily derived from a flow direction raster
water (Döll et al., 2003). In addition, water abstraction by human map and a digital elevation model with standard Geographic Infor-
activities is likely to affect the natural flow in water-scarce regions mation System (GIS) tools. Annual mean precipitation and temper-
more than in wet regions, thus providing an additional possible ature can be obtained from observations covering a given period of
explanation for the overestimation of MAF in dry regions. interest. If predictor variable values are within the range of values
employed in the calibration phase (Table 1), the uncertainty of the
4.3. Applicability of the regression model predictions is known. Application of the model to predictor values
outside the calibration domain results in MAF values with
The residuals of our regression model were not related to area unknown uncertainty. In addition to this, given the fact that the
(Fig. S1-b), suggesting that the model is area-independent and model was calibrated on the CRU TS 3.23 climatic data (Harris
maintains similar performance across catchments ranging across et al., 2014), with the HydroSHEDS 15-s river network (Lehner
at least six orders of magnitude in size (2 to 106 km2). Although et al., 2008) and with the DEM provided by the WorldClim data-
the distribution of the monitoring stations employed for the calibra- base (Hijmans et al., 2005), we acknowledge that the model is
tion of the regression coefficients was skewed towards America and not valid with other input sources.
Europe (Fig. 1), the residuals of the model-application (yearly and The regression model performed worse for extreme MAF values
10 years average MAF) were consistent across different continents when applied at finer temporal scales, as exemplified by the
(Fig. S3). This is explained by the wide range of latitudes covered decreased performance of the model on a year-by-year basis
by the monitoring stations, reflected by a large range in precipita- (Fig. 3). This might be due to the fact that the regression
486 V. Barbarossa et al. / Journal of Hydrology 544 (2017) 479–487

coefficients have been calibrated on 30 years averaged data and Acknowledgments

therefore when the model is applied at a finer temporal scale it
may underestimate high MAF values and overestimate low MAF This project has received funding from the Europeans Union’s
values. Yet, the bias due to temporal downscaling was relatively Horizon 2020 research and innovation programme under the
low at high MAF values (Fig. 3), suggesting that the overestimation Marie Sklodowska-Curie grant agreement No. 641459. The authors
of MAF at low values is related to the decreased model perfor- would like to thank the Global Runoff Data Center for providing the
mance in water-scarce regions (Fig. 4) rather than the temporal data of daily rivers discharge from worldwide gauging stations, and
downscaling. Implementing catchment aridity as predictor in a Rens van Beek from Utrecht University for providing global
regression model is, however, a non-trivial problem. Attempts to monthly gridded discharge values from the model PCR-GLOBWB.
include an interaction term of precipitation and temperature to The data used in this work are listed in the references and provided
describe catchment aridity within this study were inconclusive in the supporting information.
due to the scarce increase in the explained variance and multi-
collinearity issues. Furthermore, including complex multi-
variable predictors would go beyond the scope of developing a Appendix A. Supplementary material
conceptually simple and easily applicable model. In addition,
process-based models like PCR-GLOBWB which implement the Supplementary data associated with this article can be found, in
evapotranspiration process in a more mechanistic way reveal the the online version, at http://dx.doi.org/10.1016/j.jhydrol.2016.11.
same limitations when predicting MAF in arid catchments. There- 053.
fore, the scarcity and uncertainty of monitored precipitation and
discharge values in such regions represent a major source of bias. References
Hence, further research should focus on improving models for pre-
dicting mean annual flow in water-scarce regions, which is partic- Alcamo, J. et al., 2003. Development and testing of the WaterGAP 2 global model of
water use and availability. Hydrol. Sci. J.-Journal Des Sciences Hydrologiques 48
ularly relevant in assessments of water availability and water (3), 317–337. http://dx.doi.org/10.1623/hysj.48.3.317.45290.
footprints. In addition, future work would benefit from an increase Anselin, L., 1988. Lagrange multiplier test diagnostics for spatial dependence and
in the quality and quantity of monitoring data, for precipitation spatial heterogeneity. Geogr. Anal. 20 (1), 1–17. http://dx.doi.org/10.1111/
j.1538-4632.1988.tb00159.x.
and discharge in particular. Anselin, L., Rey, S., 1991. Properties of tests for spatial dependence in linear
regression models. Geogr. Anal. 23 (2), 112–131. http://dx.doi.org/10.1111/
j.1538-4632.1991.tb00228.x.
Barton, K., 2015. MuMIn: Multi-Model Inference. R package version 1.15.1.
5. Conclusions
Beusen, A.H.W., Bouwman, A.F., Van Beek, L.P.H., Mogollón, J.M., Middelburg, J.J.,
2016. Global riverine N and P transport to ocean increased during the 20th
We presented a conceptually simple model for predicting mean century despite increased retention along the aquatic continuum.
annual flow of rivers globally at any point of the river network. The Biogeosciences 13 (8), 2441–2451. http://dx.doi.org/10.5194/bg-13-2441-2016.
Bierkens, M.F.P., 2015. Global hydrology 2015: state, trends, and directions. Water
model explained 89% of the variance in MAF based on observations Resour. Res. 51 (7), 4923–4947. http://dx.doi.org/10.1002/2015WR017173.
retrieved from 1885 catchments worldwide. The regression coeffi- Bivand, R., Hauke, J., Kossowski, T., 2013. Computing the Jacobian in Gaussian
cients obtained were within the ranges reported by previous spatial autoregressive models: an illustrated comparison of available methods.
Geogr. Anal. 45 (2), 150–179. http://dx.doi.org/10.1111/gean.12008.
regional-scale studies and indicated that MAF scales linearly to Bivand, R., Piras, G., 2015. Comparing implementations of estimation methods for
catchment area, while it scales nonlinearly to precipitation with spatial econometrics. 63(18), 36. http://dx.doi.org/10.18637/jss.v063.i18.
an exponent of about 2. Temperature, slope and altitude, which Bring, J., 1994. How to standardize regression coefficients. Am. Stat. 48 (3), 209–
213. http://dx.doi.org/10.2307/2684719.
have not been used before in global regression models for MAF, Burgers, H.E., Schipper, A.M., Hendriks, A.J., 2013. Size relationships of water
further improve the explained variance. Our model can be applied discharge in rivers: scaling of discharge with catchment area, main-stem length
to estimate MAF at any ungauged site in the river network globally. and precipitation. Hydrol. Process. 28 (23), 5769–5775. http://dx.doi.org/
10.1002/hyp.10087.
However, it should be noted that the model is valid only for input Chang, H., 2003. Basin hydrologic response to changes in climate and land use: the
parameters within the range of the calibration variables and there- Conestoga River Basin, Pennsylvania. Phys. Geogr. 24 (3), 222–247. http://dx.
fore, outside the applicability domain the uncertainty in the esti- doi.org/10.2747/0272-3646.24.3.222.
Chen, X.Y., Chau, K.W., Busari, A.O., 2015. A comparative study of population-based
mation of MAF is unknown. Application of the model to predict
optimization algorithms for downstream river flow forecasting by a hybrid
1-year and 10-year average MAF based on independent test data neural network model. Eng. Appl. Artif. Intell. 46 (Part A), 258–268. http://dx.
revealed that only 3% and 1% of the simulated MAF values deviated doi.org/10.1016/j.engappai.2015.09.010.
more than one order of magnitude from the measurements, Cook, R.D., Weisberg, S., 1982. Residuals and Influence in Regression. Chapman &
Hall.
respectively. In addition to this, our model performed slightly bet- Danandeh Mehr, A., Kahya, E., Olyaie, E., 2013. Streamflow prediction using linear
ter than the widely employed macro hydrological model PCR- genetic programming in comparison with a neuro-wavelet technique. J. Hydrol.
GLOBWB, particularly for smaller catchments. Both the model 505, 240–249. http://dx.doi.org/10.1016/j.jhydrol.2013.10.003.
Development Core Team, 2005. R: A Language and Environment for Statistical
developed in this study and PCR-GLOBWB performed worse for Computing. R Foundation for Statistical Computing, Vienna, Austria.
water-scarce regions by overestimating the MAF, due to the Döll, P., Kaspar, F., Lehner, B., 2003. A global hydrological model for deriving water
increased uncertainty in rainfall and discharge observations and availability indicators: model tuning and validation. J. Hydrol. 270 (1–2), 105–
134. http://dx.doi.org/10.1016/S0022-1694(02)00283-4.
to the difficulty in describing the catchment hydrology in such Fan, Y.R. et al., 2015. A stepwise-cluster forecasting approach for monthly
regions. This implies that in dry regions, our model should be streamflows based on climate teleconnections. Stoch. Env. Res. Risk Assess. 29
applied very cautiously. (6), 1557–1569. http://dx.doi.org/10.1007/s00477-015-1048-y.
Farmer, W.H., Over, T.M., Vogel, R.M., 2015. Multiple regression and inverse
It is recommended that forthcoming studies on global models moments improve the characterization of the spatial scaling behavior of daily
for the prediction of MAF concentrate on including a better repre- streamflows in the Southeast United States. Water Resour. Res. 51 (3), 1775–
sentation of water-scarce regions within the model. Nonetheless, 1796. http://dx.doi.org/10.1002/2014WR015924.
Field, A., 2009. Discovering Statistics Using SPSS. Sage, London, UK.
streamflow prediction models would in general benefit from the
Gleeson, T., Wada, Y., Bierkens, M.F.P., van Beek, L.P.H., 2012. Water balance of
improvement of the quality and quantity of monitored precipita- global aquifers revealed by groundwater footprint. Nature 488 (7410), 197–200.
tion and discharge data especially in arid regions. In addition, http://www.nature.com/nature/journal/v488/n7410/abs/nature11295.
future research should focus on whether such a simplified frame- html#supplementary-information.
Gosling, S.N., Arnell, N.W., 2011. Simulating current global river runoff with a global
work for describing catchment hydrology worldwide can be hydrological model: model revisions, validation, and sensitivity analysis.
derived for finer temporal resolutions (daily or monthly). Hydrol. Process. 25 (7), 1129–1145. http://dx.doi.org/10.1002/hyp.7727.
 V. Barbarossa et al. / Journal of Hydrology 544 (2017) 479–487 487

GRDC, 2011. Watershed Boundaries of GRDC Stations/Global Runoff Data Centre. Stuckey, M.H., 2006. Low-flow, base-flow, and mean-flow regression equations for
Federal Institute of Hydrology (BfG), Koblenz, Germany. Pennsylvania streams. U.S. Geological Survey Scientific Investigations Report
GRDC, 2015. Long-Term Mean Monthly Discharges and Annual Characteristics of 2006-5130, p. 84.
GRDC Stations/Global Runoff Data Centre. Koblenz, Germany: Federal Institute Syvitski, J.P., Peckham, S.D., Hilberman, R., Mulder, T., 2003. Predicting the
of Hydrology (BfG). terrestrial flux of sediment to the global ocean: a planetary perspective. Sed.
Gyawali, R., Griffis, V.W., Watkins, D.W., Fennessey, N.M., 2015. Regional regression Geol. 162 (1–2), 5–24. http://dx.doi.org/10.1016/S0037-0738(03)00232-X.
models for hydro-climate change impact assessment. Hydrol. Process. 29 (8), Syvitski, J.P., Vörösmarty, C.J., Kettner, A.J., Green, P., 2005. Impact of humans on the
1972–1985. http://dx.doi.org/10.1002/hyp.10312. flux of terrestrial sediment to the global coastal ocean. Science 308 (5720), 376–
Hall, D., Cherry, S., Reeves, K., Lee, R., Carroll, G., Sommers, G., Verdin, K., 2004. 380. http://dx.doi.org/10.1126/science.1109454.
Water Energy Resources of the United States With Emphasis on Low Head/Low Tendall, D.M., Hellweg, S., Pfister, S., Huijbregts, M.A.J., Gaillard, G., 2014. Impacts of
Power Resources, U.S. Department of Energy Efficiency and Renewable Energy, river water consumption on aquatic biodiversity in life cycle assessment—a
Wind and Hydropower Technologies, Idaho Falls, Idaho. Retrieved from <http:// proposed method, and a case study for Europe. Environ. Sci. Technol. 48 (6),
www1.eere.energy.gov/windand hydro/pdfs/doewater-11111.pdf>. 3236–3244. http://dx.doi.org/10.1021/es4048686.
Hamon, W.R., 1963. Computation of Direct Runoff Amounts from Storm Rainfall. Int. Thomas, D.M., Benson, M.A., 1970. Generalization of streamflow characteristics
Assoc. Sci. Hydrol. Publ. 63, 52–62. from drainage-basin characteristics. US Geological Survey WaterSupply Paper
Hanafiah, M.M., Xenopoulos, M.A., Pfister, S., Leuven, R.S.E.W., Huijbregts, M.A.J., 1975. Washington, DC, pp. 62.
2011. Characterization factors for water consumption and greenhouse gas Thornthwaite, C.W., 1948. An approach toward a rational classification of climate.
emissions based on freshwater fish species extinction. Environ. Sci. Technol. 45 Geogr. Rev. 38, 55–94.
(12), 5272–5278. http://dx.doi.org/10.1021/es1039634. Tran, L.T., O’Neill, R.V., Bruins, R.J.F., Smith, E.R., Harden, C., 2015. Linking land
Hanasaki, N. et al., 2008. An integrated model for the assessment of global water use/land cover with climatic and geomorphologic factors in regional mean
resources – Part 1: model description and input meteorological forcing. Hydrol. annual streamflow models with geospatial regression approach. Prog. Phys.
Earth Syst. Sci. 12 (4), 1007–1025. http://dx.doi.org/10.5194/hess-12-1007- Geogr. 39 (2), 258–274. http://dx.doi.org/10.1177/0309133314562441.
2008. University of East Anglia Climatic Research Unit, Jones, P. D., Harris, I., 2013. CRU
Harris, I., Jones, P.D., Osborn, T.J., Lister, D.H., 2014. Updated high-resolution grids of TS3.00: Climatic Research Unit (CRU) Time-Series (TS) Version 3.00 of High
monthly climatic observations – the CRU TS3.10 Dataset. Int. J. Climatol. 34 (3), Resolution Gridded Data of Month-by-month Variation in Climate (Jan. 1901–
623–642. http://dx.doi.org/10.1002/joc.3711. Dec. 2006). NCAS British Atmospheric Data Centre.
Heiberger, R.M., 2015. HH: Statistical Analysis and Data Display: Heiberger and van Beek, L.P.H., Bierkens, M.F.P., 2008. The Global Hydrological Model PCR-
Holland. R package version 3.1-21. GLOBWB: Conceptualization, Parameterization and Verification, Utrecht, The
Hendriks, A.J., Schipper, A.M., Caduff, M., Huijbregts, M.A.J., 2012. Size relationships Netherlands. Retrieved from http://vanbeek.geo.uu.
of water inflow into lakes: empirical regressions suggest geometric scaling. J. nl/suppinfo/vanbeekbierkens2009.pdf.
Hydrol. 414, 482–490. http://dx.doi.org/10.1016/j.jhydrol.2011.11.025. van Beek, L.P.H., Wada, Y., Bierkens, M.F.P., 2011. Global monthly water stress: 1.
Hijmans, R.J., Cameron, S.E., Parra, J.L., Jones, P.G., Jarvis, A., 2005. Very high Water balance and water availability. Water Resour. Res. 47 (7). http://dx.doi.
resolution interpolated climate surfaces for global land areas. Int. J. Climatol. 25 org/10.1029/2010WR009791.
(15), 1965–1978. http://dx.doi.org/10.1002/joc.1276. Van Der Knijff, J.M., Younis, J., De Roo, A.P.J., 2010. LISFLOOD: a GIS-based
Hoekstra, A.Y., Chapagain, A.K., Aldaya, M.M., Mekonnen, M.M., 2011. The Water distributed model for river basin scale water balance and flood simulation.
Footprint Assessment Manual: Setting the Global Standard, London UK. Int. J. Geogr. Inform. Sci. 24 (2), 189–212. http://dx.doi.org/10.1080/
Hortness, J.E., Berenbrock, C., 2001. Estimating monthly and annual streamflow 13658810802549154.
statistics at ungaged sites in Idaho. Water-Resources Investigations Report 01– Verdin, K.L., Worstell, B., 2008. A fully distributed implementation of mean annual
4093, US Department of the Interior, US Geological Survey 2001–4093, Reston, streamflow regional regression equations. JAWRA J. Am. Water Resour. Assoc.
VA, pp. 36. 44 (6), 1537–1547. http://dx.doi.org/10.1111/j.1752-1688.2008.00258.x.
Janse, J.H. et al., 2015. GLOBIO-Aquatic, a global model of human impact on the Vogel, R.M., Sankarasubramanian, A., 2000. Spatial scaling properties of annual
biodiversity of inland aquatic ecosystems. Environ. Sci. Policy 48, 99–114. streamflow in the United States. Hydrol. Sci. J. 45 (3), 465–476. http://dx.doi.
http://dx.doi.org/10.1016/j.envsci.2014.12.007. org/10.1080/02626660009492342.
Jefferies, D. et al., 2012. Water Footprint and Life Cycle Assessment as approaches to Vogel, R.M., Wilson, I., Daly, C., 1999. Regional regression models of annual
assess potential impacts of products on water consumption. Key learning points streamflow for the United States. J. Irrig. Drain. Eng. 125 (3), 148–157. http://dx.
from pilot studies on tea and margarine. J. Clean. Prod. 33, 155–166. http://dx. doi.org/10.1061/(ASCE)0733-9437(1999) 125:3(148).
doi.org/10.1016/j.jclepro.2012.04.015. Vorosmarty, C.J., Fekete, B.M., Tucker, B.A., 1998. Global River Discharge, 1807–
Kennard, M.J., Mackay, S.J., Pusey, B.J., Olden, J.D., Marsh, N., 2010. Quantifying 1991, V[ersion]. 1.1 (RivDIS). Data set. Available on-line [http://www.daac.
uncertainty in estimation of hydrologic metrics for ecohydrological studies. ornl.gov] from Oak Ridge National Laboratory Distributed Active Archive
River Res. Appl. 26 (2), 137–156. http://dx.doi.org/10.1002/rra.1249. Center, Oak Ridge, Tennessee, U.S.A. http://dx.doi.org/10.3334/ORNLDAAC/199.
Legates, D.R., McCabe, G.J., 1999. Evaluating the use of ‘‘goodness-of-fit” Measures Wada, Y. et al., 2011. Global monthly water stress: 2. Water demand and severity of
in hydrologic and hydroclimatic model validation. Water Resour. Res. 35 (1), water stress. Water Resour. Res. 47 (7). http://dx.doi.org/10.1029/
233–241. http://dx.doi.org/10.1029/1998WR900018. 2010WR009792.
Lehner, B., Verdin, K., Jarvis, A., 2008. New global hydrography derived from Wanders, N., Wada, Y., 2015. Human and climate impacts on the 21st century
spaceborne elevation data. Eos, Trans. Am. Geophys. Union 89 (10), 93–94. hydrological drought. J. Hydrol. 526, 208–220. http://dx.doi.org/10.1016/j.
http://dx.doi.org/10.1029/2008EO100001. jhydrol.2014.10.047.
Lu, J., Sun, G., McNulty, S.G., Amatya, D.M., 2005. A comparison of six potential Wang, W.-C., Chau, K.-W., Xu, D.-M., Chen, X.-Y., 2015. Improving forecasting
evapotranspiration methods for regional use in the Southeastern United States. accuracy of annual runoff time series using ARIMA based on EEMD
JAWRA J. Am. Water Resour. Assoc. 41 (3), 621–633. http://dx.doi.org/10.1111/ decomposition. Water Resour. Manage 29 (8), 2655–2675. http://dx.doi.org/
j.1752-1688.2005.tb03759.x. 10.1007/s11269-015-0962-6.
Oberdorff, T. et al., 1995. Global scale patterns of fish species richness in rivers. Widén-Nilsson, E., Halldin, S., Xu, C.-Y., 2007. Global water-balance modelling with
Ecography 18 (4), 345–352. WASMOD-M: parameter estimation and regionalisation. J. Hydrol. 340 (1–2),
Oberdorff, T. et al., 2011. Global and regional patterns in riverine fish species 105–118. http://dx.doi.org/10.1016/j.jhydrol.2007.04.002.
richness: a review. Int. J. Ecol. 2011, 12. http://dx.doi.org/10.1155/2011/ Wisser, D., Fekete, B.M., Vörösmarty, C.J., Schumann, A.H., 2010. Reconstructing
967631. 20th century global hydrography: a contribution to the Global Terrestrial
Okkan, U., Serbes, Z.A., 2012. Rainfall–runoff modeling using least squares support Network-Hydrology (GTN-H). Hydrol. Earth Syst. Sci. 14 (1), 1–24. http://dx.doi.
vector machines. Environmetrics 23 (6), 549–564. http://dx.doi.org/10.1002/ org/10.5194/hess-14-1-2010.
env.2154. World Meteorological Organization, 1992. International Meteorological Vocabulary.
Ord, K., 1975. Estimation methods for models of spatial interaction. J. Am. Stat. Secretariat of the World Meteorological Organization, Geneva.
Assoc. 70 (349), 120–126. http://dx.doi.org/10.1080/01621459.1975.10480272. Wu, C.L., Chau, K.W., Li, Y.S., 2009. Methods to improve neural network performance
Pfister, S., Koehler, A., Hellweg, S., 2009. Assessing the environmental impacts of in daily flows prediction. J. Hydrol. 372 (1–4), 80–93. http://dx.doi.org/10.1016/
freshwater consumption in LCA. Environ. Sci. Technol. 43 (11), 4098–4104. j.jhydrol.2009.03.038.
http://dx.doi.org/10.1021/es802423e. Xenopoulos, M.A., Lodge, D.M., 2006. Going with the flow: using species-discharge
Poff, N.L., Zimmerman, J.K.H., 2010. Ecological responses to altered flow regimes: a relationships to forecast losses in fish biodiversity. Ecology 87 (8), 1907–1914.
literature review to inform the science and management of environmental http://dx.doi.org/10.1890/0012-9658(2006) 87[1907:GWTFUS]2.0.CO;2.
flows. Freshw. Biol. 55 (1), 194–205. http://dx.doi.org/10.1111/j.1365- Xenopoulos, M.A. et al., 2005. Scenarios of freshwater fish extinctions from climate
2427.2009.02272.x. change and water withdrawal. Glob. Change Biol. 11 (10), 1557–1564. http://dx.
Rodríguez-Iturbe, I., Rinaldo, A., 2001. Fractal River Basins: Chance and Self- doi.org/10.1111/j.1365-2486.2005.001008.x.
Organization. Cambridge University Press, UK. Yu, K.-X., Gottschalk, L., Xiong, L., Li, Z., Li, P., 2015. Estimation of the annual runoff
Santini, M., di Paola, A., 2015. Changes in the world rivers’ discharge projected from distribution from moments of climatic variables. J. Hydrol. 531 (Part 3), 1081–
an updated high resolution dataset of current and future climate zones. J. 1094. http://dx.doi.org/10.1016/j.jhydrol.2015.11.012.
Hydrol. 531 (Part 3), 768–780. http://dx.doi.org/10.1016/j.jhydrol.2015.10.050. Zuur, A., Ieno, E.N., Walker, N., 2009. Mixed Effects Models and Extensions in
Shiklomanov, A.I., Lammers, R.B., Vörösmarty, C.J., 2002. Widespread decline in Ecology with R. Springer, New York.
hydrological monitoring threatens Pan-Arctic Research. Eos, Trans. Am.
Geophys. Union 83 (2), 13–17. http://dx.doi.org/10.1029/2002EO000007.