Professional Documents
Culture Documents
Multivariate Return Periods in Hydrology: A Critical and Practical Review Focusing On Synthetic Design Hydrograph Estimation
Multivariate Return Periods in Hydrology: A Critical and Practical Review Focusing On Synthetic Design Hydrograph Estimation
cess
Hydrology and
Open Access
Hydrol. Earth Syst. Sci., 17, 1281–1296, 2013
www.hydrol-earth-syst-sci.net/17/1281/2013/
doi:10.5194/hess-17-1281-2013 Earth System
© Author(s) 2013. CC Attribution 3.0 License. Sciences
Open Access
Ocean Science
Open Access
review focusing on synthetic design hydrograph estimation
Solid Earth
B. Gräler1 , M. J. van den Berg2 , S. Vandenberghe2 , A. Petroselli3 , S. Grimaldi4,5,6 , B. De Baets7 , and
N. E. C. Verhoest2
1 Institute
for Geoinformatics, University of Münster, Weseler Str. 253, 48151 Münster, Germany
Open Access
2 Laboratory of Hydrology and Water Management, Ghent University, Coupure links 653, 9000 Ghent, Belgium
The Cryosphere
3 Dipartimento di scienze e tecnologie per l’agricoltura, le foreste, la natura e l’energia (DAFNE Department),
Received: 10 May 2012 – Published in Hydrol. Earth Syst. Sci. Discuss.: 31 May 2012
Revised: 27 February 2013 – Accepted: 9 March 2013 – Published: 2 April 2013
Abstract. Most of the hydrological and hydraulic studies highlighting theoretical and practical issues of multivariate
refer to the notion of a return period to quantify design frequency analysis. Also an ensemble-based approach is pre-
variables. When dealing with multiple design variables, the sented. For a given design return period, the approach chosen
well-known univariate statistical analysis is no longer satis- clearly affects the calculated design event, and much atten-
factory, and several issues challenge the practitioner. How tion should be given to the choice of the approach used as
should one incorporate the dependence between variables? this depends on the real-world problem at hand.
How should a multivariate return period be defined and ap-
plied in order to yield a proper design event? In this study
an overview of the state of the art for estimating multivari-
ate design events is given and the different approaches are 1 Introduction
compared. The construction of multivariate distribution func-
tions is done through the use of copulas, given their practi- A very important objective of hydrological studies is to pro-
cality in multivariate frequency analyses and their ability to vide design variables for diverse engineering projects. Re-
model numerous types of dependence structures in a flexi- cently, there has been an increasing interest in, and need for,
ble way. A synthetic case study is used to generate a large simultaneously considering multiple design variables, which
data set of simulated discharges that is used for illustrat- are likely to be associated with each other. In hydrology
ing the effect of different modelling choices on the design and hydraulics, several applications including sewer systems,
events. Based on different uni- and multivariate approaches, dams and flood risk mapping require the selection of storm
the design hydrograph characteristics of a 3-D phenomenon or hydrograph attributes with a predefined return period.
composed of annual maximum peak discharge, its volume, Standard hydrological design approaches are mostly based
and duration are derived. These approaches are based on on well-established univariate frequency analysis methods.
regression analysis, bivariate conditional distributions, bi- Notwithstanding this, approaches to describe hydrological
variate joint distributions and Kendall distribution functions, phenomena involving multiple variables have recently been
proposed, aiding the practitioners in estimating multivariate
return periods. In the literature, as will be described later on, events are selected and their most important variables such
several approaches have evolved over the years. However, it as annual maximum peak discharge, its volume, and duration
is not clear how these compare to each other and which one are analysed, as they form the basis of the analysis. Section 6
is appropriate for a given application. deals with evaluating the performance and differences be-
Recent developments in statistical hydrology have shown tween the investigated approaches in quantifying design hy-
the great potential of copulas for the construction of mul- drograph characteristics and highlights important issues for
tivariate cumulative distribution functions (CDFs) and for practitioners concerned with multivariate frequency analyses
carrying out a multivariate frequency analysis (Favre et al., in hydrology. Finally, conclusions are drawn in Sect. 7.
2004; Salvadori, 2004; Salvadori and De Michele, 2004,
2007; Salvadori et al., 2007; Genest and Favre, 2007;
Salvadori et al., 2011; Vandenberghe et al., 2011). Copulas 2 Constructing multivariate copulas
are functions that combine several univariate marginal cumu-
2.1 Choice of construction method
lative distribution functions into their joint cumulative distri-
bution function. As such, copulas describe the dependence Most of the copula-based research in hydrology addresses
structure between random variables and allow for the calcu- the application of 2-D copulas, for which several fitting and
lation of joint probabilities, independently of the marginal evaluation criteria are becoming more and more widespread.
behaviour of the involved variables. For more theoretical In contrast, the use of multidimensional copulas remains a
details, we refer to Sklar (1959) and Nelsen (2006). Sev- more challenging task. Only a few hydrological studies ad-
eral studies have been dedicated to the frequency analysis dress this issue and almost always face severe (practical)
of multivariate hydrological phenomena such as storms and drawbacks of the available high-dimensional copula families.
floods, often within the context of design. However, limited Most work has been done in the trivariate analysis of rain-
applications have been developed with more than two vari- fall (Zhang and Singh, 2007; Kao and Govindaraju, 2008;
ables (Vandenberghe et al., 2010; Pinya et al., 2009; Kao Salvadori and De Michele, 2006; Grimaldi and Serinaldi,
and Govindaraju, 2008, 2007; Genest et al., 2007; Serinaldi 2006b), floods (Serinaldi and Grimaldi, 2007; Genest et al.,
and Grimaldi, 2007; Zhang and Singh, 2007; Grimaldi and 2007) and droughts (Kao and Govindaraju, 2010; Song and
Serinaldi, 2006a,b). For a complete and continuously up- Singh, 2010; Wong et al., 2010).
dated list of papers about copula applications in hydrology Recently, a flexible construction method for high-
see the website of the International Commission on Statisti- dimensional copulas, based on the mixing of (conditional)
cal Hydrology of International Association of Hydrological 2-D copulas, has been introduced and has been shown to
Sciences1 . have a large potential for hydrological applications. In the
Multivariate frequency analysis is becoming more and literature, this construction is known as the vine copula (or
more widespread and several papers provide insight into gen- pair-copula) construction (Kurowicka and Cooke, 2007; Aas
eralizations of the univariate case and into new definitions of et al., 2009; Aas and Berg, 2009; Hobæk Haff et al., 2010).
the multivariate return period (see e.g. Salvadori et al., 2011; The underlying theory for the vine copula construction is de-
Salvadori and De Michele, 2004; Shiau, 2003; Yue and Ras- scribed in Bedford and Cooke (2001, 2002). This construc-
mussen, 2002). Since some of the proposed approaches are in tion method originates from work presented by Joe (1997) on
contradiction and others are introduced within specific con- which also the method of conditional mixtures, as applied by
texts, there exists a need to clarify the definitions provided so De Michele et al. (2007), is based. In this paper the vine cop-
far and to highlight their differences. This study is devoted ula method will be used to construct the 3-D copula for peak
to this issue and compares a set of different approaches on a discharge Qp , duration D, and volume Vp . The construction
large simulated data set, allowing illustration of the implica- and fitting is discussed in the next section.
tions of different modelling choices.
In this paper the construction of multivariate distribu- 2.2 Construction of a 3-D vine copula
tion functions based on vine copulas (also referred to as
pair-copulas by Aas et al., 2009) is first briefly intro- In this paper the focus will be on a 3-D vine copula joining
duced (Sect. 2.2), followed by an overview of several ap- the three marginal distributions of three random variables:
proaches commonly used to estimate multivariate design X, Y and Z. In general, the approach can be extended to
events based upon different definitions of joint return peri- any number of dimensions, although limitations may be in-
ods (JRP) (Sect. 3). Subsequently, a synthetic case study ad- troduced by the computational power and available data. In
dressing the selection of a design hydrograph is presented, the following, we assume that the samples of all three vari-
which will serve as a test case for evaluating the different ap- ables have each been transformed using the following rank
proaches. Section 4 provides all details on the practical con- order transformation S in order to obtain the marginal empir-
text of this case study. Then, in Sect. 5, extreme discharge ical distribution functions:
rank(x)
1 Available at www.stahy.org. S(x) := ,
n+1
3 Estimating design events: definitions and methods 3.2 Design events derived from a bivariate conditional
distribution
In the literature and in practice, several approaches exist for
estimating multivariate design events for a given design re- A second approach (denoted by MAR) consists of con-
turn period. The following sections provide a short overview ditioning the bivariate cumulative distribution function
of the most popular approaches, focusing on how a multivari- (CDF) FXY (x, y) on the univariate marginal design quan-
ate design event for a given return period could be calculated. tile xMAR = xUNI corresponding to the chosen univariate de-
In the specific case of multivariate joint return periods (JRP), sign return period TUNI . The resulting (univariate) condi-
typically a set of possible design events is found. In order tional CDF FY |X (y|x = xUNI ) can then be used to calculate
to be able to assess the differences among the described ap- the value yMAR for the conditional univariate design return
proaches, we select the most probable of all possible design period TMAR .
events. An ensemble-based design approach, in contrast to a Advantage will be taken of the bivariate copula CUV (u, v)
single design event, will also be presented. to perform the calculation. With uMAR = FX (xMAR )
It is important to note that we present different classes and vMAR = FY (yMAR ), the procedure can be ex-
of approaches: univariate (Sects. 3.1 and 3.2), bivariate pressed as follows. We can rewrite the initial definition
µT
(Sects. 3.3 and 3.4.1) and trivariate approaches (Sect. 3.4.2). TMAR = 1−FY |X (y|x = xMAR ) in terms of a copula with
In all cases, multivariate design events are provided; how- U : = FX (X) and V := FY (Y ) as
ever, in the first case the procedure is based on the concept
of a univariate return period, while in the second and third µT
TMAR = ∂CUV (u,vMAR )
case the procedure is based on the concept of a bivariate 1− ∂u |uMAR :=1− µT
TUNI
and trivariate joint return period, respectively. This premise is µT
pivotal since statistically these classes are incomparable due =
1 − CV|U=uMAR (vMAR )
to the different intrinsic nature of the return period concepts.
However, it is important to illustrate the differences in design −1 µT
⇔ vMAR = CV|U=uMAR 1 − .
events that stem from these modelling choices. TMAR
3.1 Design events derived from a regression analysis Inverse transformation yields
of P [X > x ∨ Y > y] following the notation introduced by of the Kendall function to project a multidimensional distri-
Vandenberghe et al. (2011): bution to a univariate one is similarly exploited by Kao and
µT Govindaraju (2010) in the context of a joint deficit index for
TOR = droughts.
1 − FXY (xOR , yOR )
µT The use of the Kendall distribution function to define the
= probability measure for calculating a JRP is advocated by
1 − CUV (FX (xOR ) , FY (yOR ))
µT Salvadori et al. (2011) as it is a theoretically sound multi-
= . variate approach sharing the notion of a critical layer, defined
1 − CUV (uOR , vOR )
through the cumulative distribution function, with the uni-
This approach is in fact an intuitive extension of the defini- variate approach. The definition of the return period in both
tion of a univariate return period. All couples (u, v) that are the univariate and in the multivariate Kendall approach is
at the same probability level tOR = CUV (u, v) of the copula characterized by making a distinction between super-critical
will have the same bivariate return period TOR . For a given and non-critical events based on a critical cumulative prob-
design return period, the corresponding level tOR can easily ability level. The only way to extend this to a multivariate
be calculated, the most likely design point (uOR , vOR ) of all context is by using the Kendall distribution function. Prob-
possible events at this level can be obtained by selecting the ability measures that are constructed differently always en-
point with the largest joint probability density: tail events that will have a joint cumulative distribution func-
tion value that is larger or smaller than the critical proba-
(uOR , vOR ) = argmax fXY FX−1 (u), FY−1 (v) . (4) bility level, and thus fail in subdividing the space between
CUV (u,v) = tOR super-critical and non-critical events with respect to the joint
cumulative distribution function. Following this avenue, any
The corresponding design values xOR and yOR are easily cal- critical probability level tKEN uniquely corresponds to a sub-
culated through the inverse CDFs: division of the space into super-critical and non-critical re-
gions. This is different from the OR case mentioned before,
xOR = FX−1 (uOR ) and yOR = FY−1 (vOR ) .
where in general different choices of critical events from the
Once the joint density along the level curve is derived, one same critical probability level tOR subdivide the space dif-
may consider different alternative approaches. Instead of the ferently. From a return period point of view, the copula ap-
most likely event, one may calculate the expected value of proach refers to super-critical events where at least one of the
the conditional distribution or calculate quantiles for given margins is larger than the design event, but the joint cumula-
probabilities that might lead to a design approach incorpo- tive probability may be lower than the designated level yield-
rating more than a single design event. To limit the number ing a shorter return period. On the other hand, the Kendall-
of approaches, we will focus on the most likely event only, based approach ensures that all super-critical events have a
as e.g. used by Salvadori and De Michele (2012). longer return period than the limit value, while some non-
critical events might have larger marginal values than any
3.4 Design events derived from a copula’s Kendall selected design event.
distribution function For any given copula of any dimension, the Kendall distri-
bution function can be calculated either analytically (e.g. for
Another definition of the bivariate return period is given Archimedean copulas) or estimated numerically, and can
by Salvadori and De Michele (2004); Salvadori (2004) and thus be used to calculate the Kendall joint return period. Until
Salvadori et al. (2007). Recently, the concept of this bivariate now, only a very limited number of studies actually applied
secondary return period was extended to a complete multidi- this kind of return period (e.g. Vandenberghe et al., 2010).
mensional setting by Salvadori et al. (2011), called Kendall In the following sections, the procedure for the 2-D and 3-D
return period (denoted by KEN). This return period corre- cases is outlined.
sponds to the mean interarrival time of events more critical
than the design event, the so-called super-critical or danger- 3.4.1 2-D Kendall joint return period
ous events. The super-critical events are potential threats to
the structure and will appear more rarely than the given de- After choosing the design return period TKEN2 , the corre-
sign return period. This partitioning of the probability distri- sponding probability level tKEN2 of the copula can be calcu-
bution into a super-critical and non-critical region is based on lated by means of the inverse of the 2-D Kendall distribution
the Kendall distribution function KC . This function is a uni- function (Eq. 5). In 2-D this corresponds to finding an isoline
variate representation of multivariate information as it is the on the copula.
CDF of the copula’s level curves: KC (t) = P [C(u, v) ≤ t].
It allows for the calculation of the probability that a random
point (u, v) in the unit square has a smaller (or larger) copula
value than a given critical probability level tKEN . The ability
µT
TKEN2 =
1 − KC (tKEN2 )
µT
⇔ KC (tKEN2 ) = 1 −
TKEN2
µT
⇔ tKEN2 = KC−1 1 − . (5)
TKEN2
of view than uniformly sampling over the isoline (as done by the 2-D analyses in this paper focus on these variables. How-
Vandenberghe et al., 2010). ever, as described in Sect. 3, there are several approaches that
Eventually, one will end up with an ensemble of pairs lead to the design values for Qp and Vp , including a 3-D ap-
(ui , vi ) with i ranging from 1 to N, the ensemble size. By proach. Applying the proposed approaches to the same data
means of the inverse marginal CDFs, these pairs are eas- set allows comparison of the different underlying definitions
ily transformed to real values. This ensemble could then be and implications of the model selection. However, in a prac-
used to run simulations from which the variability of specific tical context one is typically tied to a specific frequency anal-
design variables (e.g. thickness or height of a dam) can be ysis that corresponds to the unique design characteristics.
assessed. This approach needs additional analysis as it will The case study proposed in this paper consists of apply-
yield several design vectors going beyond the standard no- ing a continuous simulation model on a small, ungauged
tion of a single design event. As an example, one could route basin for which 500 yr of synthetic direct run-off time se-
an ensemble of 1000 pairs of peak discharge and volume ries at a 5 min resolution are simulated. From this series the
through a dam model and consider the water height in the 500 maximum annual peaks are selected together with their
reservoir. Using just one design event only one water height corresponding hydrograph (identified as the continuous se-
is obtained. However, using the ensemble, information on the quence of non-zero direct discharge values including the an-
range and likelihood of possible water heights for the given nual peak). Note that as direct discharge is considered, a zero
design return period is obtained, making it possible to incor- discharge value does not imply a dry river. Consequently,
porate the variability within the design variables stemming 500 (Qp , D, Vp ) triplets are available to which the described
from multiple design events along the critical level. approaches estimating design events are applied. By consid-
In the trivariate case (see Sect. 3.4.2) no isoline is obtained ering a real case study, the obtained differences and hence
but an isosurface. Similar to the 2-D case, the full weight the implications of a modelling choice can be evaluated in a
function over this isosurface could be rescaled to a bivari- practical context. In order to simulate the 500 yr run-off time
ate probability density function out of which an ensemble series, the COSMO4SUB model, described in the following
of triplets could be sampled. The higher the dimensionality section, is applied.
of the design problem, the more advantageous the ensem-
ble approach becomes: in three dimensions more informa- 4.2 The COSMO4SUB framework
tion is lost than in two dimensions by selecting just one de-
sign event. The drawback of the ensemble approach is the The synthetic data set on which the previously described
increasing need for run time when higher dimensions are approaches are applied is obtained through the use of
considered. the COSMO4SUB framework (Grimaldi et al., 2012d,c).
COSMO4SUB is a continuous model which allows the sim-
ulation of synthetic direct run-off time series using mini-
4 Differences among multivariate design events in the mal input information from rainfall data and digital terrain
synthetic design hydrograph application support. Specifically, the watershed digital elevation model
(DEM) with a standard resolution used in hydrological mod-
4.1 Experimental set-up elling, the soil use and type, daily (preferably at least 30 yr
long) and sub-daily (preferably at least 5 yr long) rainfall
In order to illustrate differences among estimated design observations are the only data necessary to run the model.
events by the approaches described in the previous sections, a COSMO4SUB includes three modules: a rainfall time series
simulation experiment is set up and analysed with respect to simulator, a rainfall excess scheme and a geomorphological
the synthetic design hydrograph (SDH) attributes. The SDH rainfall–run-off model. Next, the general principles are ex-
is defined as a hydrograph with an assigned return period plained and in Sect. 5.1 specific details of the calibration are
(uni- or multivariate), which can be characterized by random presented.
variables such as the peak discharge Qp , the duration D and The first module is based on a single-site copula-
the volume Vp . Specifically, given an observed or simulated based daily rainfall generator (Serinaldi, 2009) and on the
run-off time series from which a set of extreme hydrographs continuous-in-scale universal multifractal model (Schertzer
is selected, one can determine the SDH shape in several ways and Lovejoy, 1987) for disaggregating the daily rainfall to
(see Serinaldi and Grimaldi, 2011 and references therein). In the desired time scale (up to 5 min). The parameters included
a 2-D set-up, two hydrograph parameters (peak discharge and in this first module (six for each month for the daily rainfall
volume, peak discharge and duration or volume and duration) simulator and three for the disaggregation model) are cali-
should be fixed, while the third one is obtained from the cho- brated on the basis of the available rainfall observations (at
sen hydrograph shape distribution. In a 3-D set-up the three two different scales).
characteristic parameters are obtained jointly. The second module is related to the rainfall excess step.
In most common hydrological applications the interest is A new mixed Green–Ampt Curve Number (CN4GA Curve
in the peak discharge (Qp ) and volume (Vp ). Consequently, Number for Green Ampt) procedure was recently proposed
(Grimaldi et al., 2012b) and included in the present version Cellesi rain gauge station for a period of 49 yr at a daily
of the COSMO4SUB framework. The key concept is to use time scale, and for a period of 10 yr at a 5 min resolution
the initial abstraction (i.e. all the losses due to initial satura- (Serinaldi, 2009, 2010). For a description and evaluation of
tion, filling terrain gaps, interception, etc.) and the total SCS- the 500 yr rainfall synthetic time series, we refer to Grimaldi
CN excess rainfall volume to estimate the effective saturated et al. (2012c,d).
hydraulic conductivity and the ponding time of the Green–
Ampt model. Consequently, the CN4GA approach tries to 5.2 Annual extreme discharge events
appropriately distribute the volume estimated by the SCS-
CN method over time. This module is characterized by five Once the 500 yr synthetic direct run-off time series is deter-
parameters (specified in Sect. 5.1) which are empirically as- mined, as described in Sect. 4.1, the 500 maximum annual
signed using the soil use and soil type map information. In peak discharge events are selected and characterized by their
addition, the event separation time (Ts ) is included in this peak discharge Qp , duration D, and volume Vp . For only six
module since the continuous implementation of the SCS- years the model provides zero direct run-off, which is rea-
CN method requires to fix a no-rain time interval for which sonable considering the limited size of the watershed. These
the cumulative gross and excess precipitation can be reset to values are excluded in the following analyses.
zero. As shown in Grimaldi et al. (2012d,c), this parameter All approaches rely on the marginal distribution functions
has a limited influence on the final results, and the value can of Qp , D and Vp that need to be fitted in the first place.
be arbitrarily assigned in the range of 12–36 h. As the peak discharge variable consists of annual extreme
The third module allows a continuous convolution of values selected from the simulated 500 yr discharge series,
the rainfall excess to be carried out for obtaining the di- and the other two variables are closely correlated (but not
rect run-off time series through an advanced version of the necessarily annual maxima), the fit of several extreme value
width function instantaneous unit hydrograph (WFIUH). The distributions is considered, i.e. the exponential, the Weibull
adopted model, named WFIUH-1par (Grimaldi et al., 2010, and the generalized extreme value (GEV) distribution func-
2012a), identifies the watershed IUH through the topographic tions. These distributions are, respectively, a one-, two- and
information, and needs only one parameter that can be quan- three-parameter distribution, allowing for various degrees of
tified referring to the watershed concentration time (Tc ), esti- model complexity. Furthermore, the GEV distribution gen-
mated using empirical equations. Following the application erally encompasses three different distributions, namely the
of the three described modules, a continuous run-off sce- Fréchet, the (reversed) Weibull and Gumbel distributions ei-
nario is obtained from which maximum annual hydrographs ther directly, or through a transformation, as in the case of the
in terms of their peak discharge are selected. It is important Weibull distribution which corresponds to a reversed Weibull
to note that the variables duration and volume in the selected distribution. These different distribution types each repre-
triplets do not necessarily reflect annual maxima. sent a different kind of tail behaviour, namely a light tail
(Gumbel), a heavy tail (Fréchet) and a bounded upper tail
(Weibull). These behaviours can be separated based on the
5 Data and materials shape parameter ξ of the GEV. Furthermore, the Weibull dis-
tribution is fitted separately as well, as it only corresponds
This study is based on simulated data and a statistical model to a GEV distribution after transformation. Finally, most ex-
is fitted to this data set. This way a data set of sufficient size treme value distributions are of the exponential type, and
to compare the various approaches presented in this paper is cannot deal with an offset, i.e. when the smallest value of
obtained. the variable in the CDF is larger than zero. However, as a re-
sult of censoring the zeros, the smallest value of the variables
5.1 Model set-up tends to be significantly higher than zero, leading to poor fits
of the CDF. Therefore, a location parameter has been intro-
In order to provide a realistic scenario that can be duced in the distributions to ensure a proper fit in the tails.
used to evaluate the previously described approaches, the A first test to ascertain the appropriate distribution for the
COSMO4SUB model was applied on the Torbido River, a three marginal variables is to display the empirical CDFs to-
small tributary of the Tiber River located in central Italy (wa- gether with the directly fitted distribution. This is shown in
tershed area: 61.67 km2 ). Basin elevations range from 85 to Fig. 3, in which only the upper tail of the CDF is shown,
625 m, the average slope is 22 %, and the maximum distance i.e. the interval [0.80, 1] as the focus is on the extremes.
between divide and outlet is 25.8 km. The watershed DEM It can immediately be seen that not all the distributions fit
at a 20 m spatial resolution was provided by the Italian Ge- these tails equally well. This is corroborated by the Akaike
ographic Military Institute (IGMI, 2003), while land cover information criterion (AIC) computed for all different mod-
was extracted from the CORINE database (EEA, 2000). els, shown in Table 1, as well as the log-likelihood of each
Observed rainfall data, useful for calibrating the two-stage model (not shown). Based on these criteria and Fig. 3, we
rainfall simulator parameters, are available from the Castel select the Weibull distribution for Qp and the exponential
Table 1. The values of the AIC for the various distributions of the 0.00 0.50 1.00
respective variables.
1.00
●
● ●●
●● ● ●●●●
● ●● ● ●
●
●●
●
●
● ●● ● ● ● ● ●● ●●● ● ●●●●
●
●● ●
●
● ●
● ● ●● ● ●●●●
● ●●
● ●●
●●
● ●
●●
●●● ●●
● ● ● ●● ●● ● ●●● ●●● ●● ●
●●●●●
● ●●● ● ●●●
●● ● ●● ● ●●●
● ●
● ●●●
●
● ●●●●●●● ●●●● ●
●●●
●●●●●●●
●●
● ● ● ●
● ●
● ● ● ●● ● ●●● ●●●
● ● ●● ●
● ●● ● ● ● ●● ●
● ● ●
● ● ● ●● ●●●●● ●● ● ●● ●
●●●
● ● ●●●
● ●● ●
●● ●● ● ●
●●●
●
● ● ●●●●●●
● ● ●
●●
●●
●●●
● ●
●●
●● ●●
●
●
● ● ●● ● ●● ●
●
GEV Exponential Weibull ●
●●● ● ●● ●●
●
● ● ● ●● ●
●●●
●●
●
●● ● ●
● ● ●●●
● ●●●
●
●● ●●● ●
●●●
●● ●●
● ● ●●
●
●●●● ●
●
●●
●●
●
●
●● ●●●
● ●
●●
●
●
●
●
●●
●
●
●●
●●
●
●●●
●●
●●
●●●●
●
●●●
●
●●●
●●
●●
●● ● ●
●
●● ● ●
●
●●
●●
●
●
●
●●
0.50
●● ●● ●● ●● ●●● ●●● ●
●● ●
●● ●● ● ●● ● ●●● ● ● ● ●●
●● ● ● ● ●●●●●
Qp ● ● ●
●
● ●● ●
●
●
●
●
● ●●
●
●●
● ●
● ● ●●
●●
● ●●●
●● ●
●●
●
●●
●
●
●●
●●
●●
●
●●
●
●
●●
●●●●
●●
●● ●
● ●● ●●●● ●
●● ● ● ●
●
● ●● ●
● ●●
●●● ●● ● ●
●
● ●●●● ●
Qp 5370 5360 5326 ● ● ●
●
● ●
● ●●
● ● ●● ●● ● ●
●● ● ● ●
● ●
● ●●● ●
● ●●●
●●● ●
●● ●● ● ● ●
●●
● ●
●●● ●●● ● ●
●
●● ● ●
●●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●●●
●● ●
●
●●● ● ●●
●● ●
● ● ●
● ●
●●●
●
● ●● ●●● ● ● ●● ● ●
●● ●
●
●
●
●
●
● ●●●●●● ● ●
● ● ● ●● ● ● ●● ●●●
● ●
● ●●
D 2610 2646 2928 ●
●
●
●
●
●
●
●
●
●
●●
●●
●●
●
● ●● ●●●
●● ● ●
●● ● ● ●
●● ● ●
● ●● ● ●
●●
● ●
●
●
●
● ●
●
● ● ● ● ● ●●
●●
● ●
● ●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●●
●●●
●
●●
●●● ●
● ●
●
●●●●
0.00
●
● ●
●●●● ● ●● ● ●
●●
●●
Vp 14 641 14 599 14 601 ●●
●
●
● ● ● ●
●
●
●
●
●
●
●
●
●
1.00
● ● ●●● ●● ●●●
●●
●●
●● ●● ● ●● ●● ● ●●
●● ●●●● ● ●
●
● ● ●● ●●●
●●
● ● ●● ●●● ●● ●●● ● ● ●● ●●● ● ●●
●●●●
●
●●●●●●
●●●
●
●● ●● ●●●
● ●●●● ● ● ●● ●● ● ● ●
● ● ●
● ●● ● ●●
● ●●● ●● ● ●●● ●● ●
●● ● ● ● ● ● ● ● ● ● ● ● ●
● ● ●● ●●● ● ● ●●
● ● ●●
● ●
● ● ●● ● ●
●● ● ● ●● ●●● ● ●●● ● ●
● ● ● ●● ●●
● ●● ●
●●●● ●
●● ● ●● ●●● ● ●● ●
●●●●
●
●
●
●● ●● ●● ●● ● ●
●
●
● ● ● ● ● ● ● ●●● ● ●●● ● ●●
●● ●● ● ● ● ● ●●● ● ●●● ● ●
●● ● ● ●● ●●●● ●
●
●● ● ●●●● ●● ●● ●
Qp: Weib D: Exp. Vp: Exp. ● ●● ●
●
●
● ●● ●
●
● ●●
●● ●
●
●●● ● ●
●
●●●●
●
● ●●● ●●
●●● ●●
●● ●● ●
●
●
●
● ● ●● ● ●
● ●
●●● ●● ●
●●● ●
●●
● ●
●● ● ● ●
●● ●● ● ●
●●● ●●
●●
● ● ●● ●● ● ● ●●●● ● ● ●● ● ● ● ●● ● ●● ●●● ●●●
0.50
● ● ●● ●●
●●●●●●●● ●
● ●● ●● ● ● ●● ● ●● ●
●
●● ● ● ●●● ●
●● ●●
1.00
1.00
1.00
●●●●
● ● ● ● ● ● ●
● ●● ●● ● ● ●●
● ● ●
● ●●●● ●
●●
●●
●
● ● ●● ● ●●
● ● ●●
●
●
●●
● ● ● ●●
●● ● ●●
●●
●
●● D ●●● ● ●
● ●●
●● ●
●
●
●●●● ●
● ● ● ●
●●●● ● ●
● ●●
●
● ● ● ● ●●
●● ● ●● ●● ●● ● ●● ● ● ●● ●
● ● ●● ●●●● ●●● ● ●
●● ●
●
●●
● ● ●
●
● ● ● ● ● ●●●
●● ●
●● ●
● ●●
● ● ● ●●● ●● ● ● ● ●● ● ● ● ● ●●●
● ● ●●●● ●● ●
● ●
●
●●● ●●● ● ●● ● ●● ● ●● ●●● ● ● ●●● ●●
● ● ● ● ● ●●● ● ●
● ●
● ●● ●●●
●●
●● ● ●● ● ●● ●●●●●●
● ●
● ●● ●
● ● ●●●●●● ● ●● ● ●
● ● ●
● ● ●●●●● ● ● ● ● ●
● ●● ● ●● ●
● ● ● ●● ● ● ● ●●● ●●●● ●●● ● ●
●● ● ● ● ●
●●●
●
●● ● ●● ● ● ● ●● ●●
●● ● ●
●●
● ● ●●
● ●
●● ●
●
●●●
● ● ●●●● ●●●●●● ●● ● ● ●● ●●●●● ●● ●●●●● ●●●●● ● ●●
● ● ● ● ● ●
●
●●●●●●●● ●● ●●●● ● ●
●●●●●●●● ●● ●● ●● ●
0.00
0.95
0.95
0.95 ●
●
●
●
●
●●
●●
●●
●●●●●●●● ● ● ●
●
●
●
●
●●
●●
●●
●
●●●●● ●●● ● ●
1.00
● ● ●
●●
●
●● ● ● ●
●● ● ●
●
●
●
●●●●●●●
●●●●●
●●
●
●
● ● ●● ● ●●
● ●●● ● ●
●●
●
●
●
●●
●
●
●●
● ● ●●● ●
●● ● ● ●●●
●
● ● ● ●●●
●
● ●
●● ● ● ● ● ●●●● ●●●
● ●
●
●
●●●●● ● ● ● ● ● ●
●● ●●
●●●
● ● ●● ● ● ●●● ● ● ●●●●●●
● ●● ●●● ●●●● ● ● ●●● ●
●● ●● ●●
● ●●● ●●
●●●●●
● ●
●● ●● ● ●●
●●●●● ●● ● ●● ●
● ●● ●●● ● ●●●●●● ● ● ● ●● ●
● ● ● ● ●
● ●●●
●●● ●●
●
● ● ●● ● ● ● ● ● ●
● ●
●●●
●
● ● ● ● ●● ●● ● ● ● ● ● ●●●●●●
0.90
0.90
0.90
●● ● ●●●●●●●● ●●
●● ●●●
●
● ●●
● ● ●
●● ● ●● ● ●●●
● ● ● ● ●● ●●
●● ● ●● ●● ●
● ●
●●
●●
●●● ● ● ● ●● ● ● ● ● ●● ●●●
●
● ●●
●●
● ● ● ● ● ● ●●●● ●● ● ●● ●
●● ● ●
●
●●●● ●●
●●
● ● ● ● ●● ●● ● ●●●
●● ●● ● ●
0.50
●●●● ● ● ●●● ● ●●● ● ●●●● ●
● ● ●● ●● ●●
●●
●●
● ●● ● ●●
● ● ● ●●●● ●● ● ● ●● ●● ● ●
●● ● ● ●●●
● ●● ●
●●
●
●
●
●●
●
●●
●●●
●
●●
●
●
●
●●
●● ●●● ●
●●● ●●● ●
● ●● ●
●
● ●
●
●●●● ●●
● ●●
● Vp
●●● ● ●●●●●●
● ● ● ●● ●● ●
● ●
●
●●● ●●
●●●
●●● ● ● ● ● ● ● ●● ●●● ●
●● ●●
● ● ●● ●●●● ●● ●● ●
●●●● ● ●
●● ● ●
●
●●●●
●● ●●●●
●● ● ● ● ●●● ● ●● ● ● ●
● ●●● ●
● ● ●● ●
●●●
●● ● ●●
●
●
●
●
●
●
●
●
● ● ● ●● ●● ●● ● ● ● ●● ●● ● ● ● ●
●●
●
●
●●
● ●
● ●● ●●●●● ●●● ●
●●●
●●
●●
●
● ●● ●
●
●●
● ●●●● ●●●● ● ●
0.85
0.85
0.85
●● ●
●●●●●
●
●
●
●
●
●
●● ● ●●
● ●● ●●●● ●●●●● ● ●
●
●●●●
● ● ●● ● ●● ● ●
●● ●●●
● ●
● ● ●●● ●
●●
●
●
●
●
●
●
●● ● ● ●●●● ● ●● ● ● ● ● ●●
●●
●●●
●
●
●●
●● ●
● ●
●● ● ● ● ● ●● ●
●
●●●● ● ●●
0.00
●
●
●● ●
●
●● ●
GEV ●
●
●
●
●
●
●●
●●
●
●
●
● ●
●
Exp.
Weib.
0.80
0.80
0.80
0 200 400 0 20 40 0e+00 4e+06 8e+06 Fig. 4. Normalized rank scatter plots for all pairs of variables.
3
peak discharge [m /s] duration [h] volume [m ] 3
Kendall’s tau is 0.85 for (Qp , Vp ), 0.42 for (Qp , D) and 0.54 for
(Vp , D).
Fig. 3. The various cumulative distribution functions together with
the empirical cumulative distribution function for the three vari-
ables. The best fitting distribution is denoted in the title of each calculated, and normalized rank scatterplots are evaluated for
graph. each pair of variables (Fig. 4). Evidently, there are strong
positive associations. Also, some ties are present, especially
for D, which have been assigned with their mean rank in the
distribution for Vp . Seemingly, the GEV provides the overall
transformation. The next section deals with the modelling of
best fit for D according to the AIC, despite the poor repre-
these associations.
sentation of the upper tail (see Fig. 3). As the focal point of
this study is set around a return period of ten years addressing 5.3 Fitting of the 2-D and 3-D copulas
the top 10 % of the CDF, we chose to select the exponential
distribution because of its better fit in this region. More in- As described in Sect. 2.2, we used maximum likelihood es-
depth testing through Q-Q plots (not shown here) indicates timation to fit a copula from each investigated family for
that this is indeed a better approximation of the distribution. every pair of variables and selected the best fitting one by
Further investigation of additional distribution families and the highest log-likelihood value. The copula families inves-
combinations of these might improve the fit of the marginals, tigated include Gaussian, Student, Gumbel, Frank, Clayton,
but is out of scope of this paper. Nevertheless, a good fit of BB1, BB6, BB7, BB8 and the survival copulas of the 4 lat-
the marginal distributions is key to the practical application. ter ones (details on all these families can be found in Nelsen
Hence, the following models are selected: (2006) and Joe (1997)). Table 2 gives an overview of the pa-
– Qp : Weibull (Anderson–Darling p = 0.59), rameters and goodness-of-fit results. The p values are esti-
mated from 1000 iterations each.
– D: exponential (Anderson–Darling p = 0.08), The following approaches in the 2-D case make use of the
fitted BB7 copula C13 which models the dependence between
– Vp : exponential (Anderson–Darling p = 0.18).
Qp and Vp . It should be noted that this copula is not able to
Here, the Anderson–Darling test was used to determine represent the boundary effect present in the rank scatter plot
whether the samples were significantly different from the fit- (Fig. 4). To the authors’ knowledge, no copula is available
ted distributions. It should be understood that a consistency in the literature that would be able to model such a bound-
in marginal distribution functions across the different ap- ary effect. As the BB7 copula family belongs to the class of
proaches is far more important for comparison reasons than Archimedean copulas, its Kendall distribution function can
a perfect fit, considering the underlying data are simulations. easily be obtained analytically.
To analyse the association between the variables, which For the 3-D case, the three fitted bivariate copulas C12 ,
will be modelled by means of copulas, Kendall’s tau is C23 and C13|2 are then composed into the 3-D vine copula
Table 2. An overview of the fitted bivariate copulas in the 2-D copula-based and 3-D vine copula-based approach.
6
parameter Clayton, Frank, and Gumbel copulas have also ●
● ● ●
been performed. The log-likelihood shows a 10% increase
volume [106m3]
●
for the fitted vine copula (1047) with respect to the Gaussian ● ●
●
4
●
● ● ●
one (935), while the three one-parameter Archimedean cop- ●●
●
● ● ●
● ●
● ●
● ●
● ● ●
ulas have far smaller values (432–532). Thus, the vine copula ●
● ●
● ●
● ●
● ●
●● ●
●
● ● ●●
●●
2
●● ● ●
yields the best fit within this set of copula families in terms ●
● ●● ●
●
● ●●
● ●
●
●●
● ● ●●
●
●●
●
●
●●
●
●
●●●
●
●
●
●
●
●●
●●●
●
●
●
●● ●●● ●
●
● ●● ● ●
●●●● ●●●● ●●●● ● ●
●● ●●● ●●●
● ●●●●
●
●●●
of the log-likelihood. As no closed form exists for the cumu- ●●
●
●●
●●●
●
●
●
●●
●●
●
●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●●●
●●●
●●●●
●
● ●
●●
●●
●●●
●
●
●●
●
●
●●
●
●●
●●
●
●●●●●
●●
●●
●
●●
●
●
lative distribution function of this vine copula, a numerical ●
●
●●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●●
●
●
●●
●
●●
●●
0
evaluation based on a sample of 100 000 points was carried
0 100 200 300 400 500
out in order to be able to calculate the (inverse of the) Kendall
peak discharge [m3/s]
distribution function.
However, the singularity appearing in Fig. 4 for the pair Fig. 5. Illustration of the derivation of the design quantiles based on
(Qp , Vp ) is neglected in both, the bivariate and the vine cop- the regression approach.
ula (as well as in the other considered copulas). In the bi-
variate case, no copula family with such a limited support
could be found, while the vine copula’s decomposition is
based on the bivariate copulas C12 and C23 addressing the visualized as grey dots in Fig. 6 that summarizes all de-
pair (Qp , Vp ) only through the conditional joint distribution. scribed approaches.
Thus, all investigated copula families would in general sam- First the 2-D case is considered, in which the focus is on
ple unrealistic point pairs (Qp , Vp ) beyond the border appear- the couple (Qp , Vp ). In the regression-based approach (REG,
ing in the scatter plot. A discussion on this singularity and its Sect. 3.1) the starting point is the univariately derived quan-
underlying process can be found in Serinaldi (2013). tile qp,UNI , being usually the driving variable in many hy-
drological applications (see Eq. 3). Based on a regression
between Qp and Vp , as shown in Fig. 5, the design volume
6 Results and discussion vp,REG is easily estimated as 2.14 × 106 m3 . This volume is
lower than the one obtained by a purely univariate analysis
6.1 Calculation of single design events partly due to the different definition based on the expectation
instead of a quantile.
In this section the design values for the SDH with a design The second 2-D approach is based on the conditional cop-
return period of 10 yr are calculated based on the 2-D and ula (MAR, Sect. 3.2). The conditioning of the bivariate cop-
3-D approaches presented in Sect. 3. The triplet (Qp , D, Vp ) ula CUW (denoted as C13 in Sect. 5.3) for uUNI = 0.9 results in
is considered for which the following transformations hold the function CW|U (w, u = 0.9). The value of wMAR = 0.9521
corresponds with a probability level of 0.9. By means of the
U = FQp Qp , V = FD (D) and W = FVp Vp . inverse FV−1p
(wMAR ), the design volume vp,MAR is calculated
as 2.92 × 106 m3 , which is considerably larger than the for-
As a reference the univariate case is analysed first. mer design volumes.
Based on the inverse of the CDFs FQp , FD and FVp , The true joint return period approaches based on the bi-
at a probability level of 1 − TµUNI
T 1
= 1 − 10 = 0.9, the de- variate copula CUW (OR, Sect. 3.3) is the third 2-D ap-
sign values qp,UNI = 174 m s , vp,UNI = 2.21 × 106 m3 and
3 −1 proach. For TOR = 10 yr, the corresponding copula level
dUNI = 16.02 h are obtained. In the following, Table 3 and tOR equals 0.9, and corresponds to an isoline. Using the
Fig. 6 provide a way to compare these and all further es- marginal CDFs for Qp and Vp , Eq. (4) can be solved to
timated design events. In order to be able to compare de- find the point (uOR , wOR ) with the highest joint likelihood,
sign events with the data, the simulated pairs (Qp , Vp ) are i.e. (uOR , wOR ) = (0.927, 0.925). Using the inverse CDFs
Table 3. Overview of the calculated design event for T = 10 yr, based on several approaches. The values are rounded to address the limited
numerical precision and ease comparison.
●
● ●
●
●
● ● the 3-D vine copula. In contrast to the 2-D approaches, the
3
●
●
●
● ● tKEN3 level corresponds to a surface. Using the marginal
●
● ●
●
●
●
●
● ●
●● ●data
●
CDFs in combination with Eq. (6), the most likely point
● ●
●●
●
● ● ● UNI on this surface is found as (uKEN3 , vKEN3 , wKEN3 ) = (0.844,
volume [106m3]
● ● ●
● ● ● ● ● ● ● ●
REG
● ● ● ●● ● ●● ●
●
●
●●
● ●
●●
● ●
●●●●
● ●
●● MAR 0.820, 0.851). Using the inverse CDFs this results in the de-
● ● ● ●● ●● ● ●●●
sign event (qp,KEN3 , dKEN3 , vp,KEN3 ) = (147 m3 s−1 , 12.90 h,
● ● ●●● OR
● ●
● ● ●●
●● ●
●
●● ● ● ● ● ●
●● ●● ● ●
● ● ●
● ●● ●
● ●
● KEN2D
1.83 × 106 m3 ). Note that the Kendall distribution function
● ●●● ● ●● ● ●● ●●●●●
● ● ●● ● KEN3D
● ● ● ●● ●●
1
● ●
●● ● ●●●●●●
● ●●
● ●●●●●●●
●● ●
● ●
● ● ● ●●●●● ●
●● ●●● ●●● ● ●●
●●● ●
●● ●●●
● ●
●
●
● ●●
●
●●
●
●●●●
●●●●
●●
●●●
●●●●●
●●
●●
●
●●●
●●●
●
●
●●
●
●●●●
is a univariate representation of multivariate information and
● ●●●
● ●
●●●
●
●●●
●
●
● ●●
●●●●
● ●
●
●
●●
●●
●
●
●
●
●●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
● ●● ●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●●●
●●
●
that its form is different in the 2-D and 3-D cases.
●
●●
●
●●
●
●●
●
●●
●
●●●
● ●
●●
●
●●
●●
●
●●
●
●●
●
●●
●●
●
●
●●
●
0
0.02
5
OR
volume [106m3]
KEN2D OR
4
KEN2D
density
0.01
3
2
0.00
140 160 180 200 220 240 260
150 200 250 300 350
peak discharge [m3/s]
3
peak discharge [m /s]
1.2e−06
OR
of [150 m3 s−1 , 238 m3 s−1 ] and a volume in [183 × 106 m3 , KEN2D
326 × 106 m3 ]. Note from Fig. 7 that lower volumes occur
6.0e−07
density
together with higher peak discharge values and vice versa. As
briefly mentioned in Sect. 3.6, the ensemble of design events
0.0e+00
can also be used to calculate another design variable, such
as the water height in a reservoir, for which again a PDF of
2 2.5 3 3.5 4 4.5 5
possible design values can easily be obtained. However, this
exercise is beyond the scope of this paper. volume [106m3]
3.5
MAR
OR
KEN2D
3
volume [106m3]
2.5
2
1.5
Fig. 10. 500 simulations of different approaches to obtain a single design event with a design return period of 10 yr. Straight line segments
indicate the univariate approaches and curved lines represent the bivariate approaches, with dashed lines approximating the ensemble con-
fidence band estimates for their corresponding point cloud each. The estimated design events are denoted as filled shapes. The open grey
shapes represent the most and least extreme 2.5 % for each approach.
quantile, whereas for the approaches using the Kendall-based correspond to the most severe one for a given hydraulic
JRP a smaller quantile is found. The other approaches use structure. However, there is the full potential to set a step
the univariate quantile as a starting point, resulting in identi- aside from this one-event-design approach to a full ensemble-
cal quantiles. Considering the design quantiles vp,· , the MAR based design approach. Therefore, this paper includes an ap-
and OR approaches yield a larger quantile than the univariate proach for the generation of a design ensemble. It is clear
quantile, while the REG and Kendall-based approaches yield that the ensemble approach provides a lot more informa-
smaller values. tion on the possible outcome of design events. The pro-
The three true JRP approaches use information of the full posed ensemble-based approach entails the most likely de-
bivariate (copula-based and 2-D Kendall-based approach) or sign event, but furthermore provides a clear idea on the prob-
trivariate (3-D Kendall-based approach) distribution func- ability that other events (but all having the same JRP) will
tion. The Kendall-based approaches have the advantage of occur. Checking these ensembles against the desired design
using a mathematically consistent way of defining the proba- of the hydraulic structure will illustrate the real threat to the
bility of extremes or dangerous events relying on the CDF as structure. It therefore provides a way of assessing some un-
in the univariate approach, unlike the JRP approach based certainty of the design variables associated with the selection
on the copula solely. For a full discussion please refer to of a single design event. If a single design event is sought,
Salvadori et al. (2011). However, there is no universal choice the pure copula-based approach (Sect. 3.3) has the advantage
of an appropriate approach to all real-world problems. The of guaranteeing that only a fraction of 1/TOR events exceeds
most important is to address the problem from a probabilis- the margins of any of the possible estimated design events.
tic point of view and to be aware of the practical implications It should also be noted that the fitting of the copulas (bi-
of the approach chosen (outlined in this paper). It is also ev- variate, trivariate or multivariate) is a very important part
ident, but not necessarily the case, that the more variables of the design event estimation. If the practitioner is not ac-
are included (2-D vs. 3-D), the smaller the design quantiles quainted with this initial aspect of design studies, it is very
become. Salvadori and De Michele (2012) discuss this di- easy to make wrong choices. Naturally, the multidimensional
mensionality paradox and provide a theoretical explanation approaches require a larger data set in order to produce robust
for it. parameter estimates. Thus, the length of the time series and
Furthermore, the issue of selecting just one design event the amount of missing data have to be considered before an
out of a range of events all having the same joint return pe- approach is selected. The authors of this work believe that the
riod (i.e. on an isoline or isosurface of the copula) could be vine copula approach is the way to go for constructing flex-
seen as a drawback of the multivariate approaches available ible multivariate distribution functions, as it enables use of
in the literature, as the most likely event does not necessarily more widely spread bivariate copulas as building blocks for
more complex multivariate distribution functions. Of course, simulation study showed that the investigated approaches
a good balance between the number of variables considered yield statistically different design events. Thus, the predic-
and the (numerical) complexity of the vine copula should be tions are not only different following the theoretical inequal-
sought, keeping in mind that all this also affects the eventual ity (Eq. 7), but do withstand the variability due to uncer-
design. Further studies are necessary to assess the sensitivity tainties associated with the probabilistic model fitted to the
of the JRP analysis to sample size and sample selection. data. Based on the available literature and the case study
In general, the Kendall-based approach can be applied to in this paper the copula-based and Kendall-based JRP ap-
any copula and can be used for both large (e.g. floods) and proaches are valuable multivariate extensions of the univari-
small (e.g. droughts) extremes. However, one should also be ate approaches. However, their applicability always depends
aware of the fact that the approaches in this study are only on the availability of data and the probabilistic nature of the
applied to variables that are positively associated and with actual real-world problem. For constructing multivariate cop-
a focus on extremes in terms of large values. In all other ulas, the vine copula method is advised.
cases, adaptations should be made in order to operate in the Further (joint) research efforts should focus on a shift
right area of the copula. Further applications of the copula- from one-design-event approaches to ensemble-design-event
based and Kendall-based JRP approaches in other case stud- approaches, enabling incorporation of the variability in the
ies should provide more insight on this in the near future. design event selection. A first valuable approach to this
Serinaldi (2012) highlighted the potentially misleading no- ensemble-based design was provided in this paper. The ul-
tion of return periods, and suggests to report return peri- timate goal should be the elaboration of a useful and un-
ods alongside with annual exceedance probabilities as done derstandable framework for multivariate frequency analyses,
e.g. by Theiling and Burant (2012). This aspect should be with clear guidelines to practitioners.
included in further studies as well. From a practical perspective, it is impossible to provide a
general suggestion for an appropriate approach to estimate
multivariate design events applicable to a vast set of design
7 Conclusions exercises. Firstly, as previously described, the available ap-
proaches are different from a statistical point of view. Until
The aim of this study was to provide an overview of state- now many applications are based on the concept of univariate
of-the-art approaches to estimate design events for a given return periods, as the concept of multivariate return periods
return period and to discuss their differences in a practical has a different meaning and is potentially less conservative.
application. Therefore, a synthetic case study focusing on the Secondly, the best approach, in our opinion, is related to the
estimation of design parameters for a synthetic design hydro- hydraulic structure to be designed. Different design exercises
graph (SDH) was considered. As they are the most important might be critical to single variables, which should then be
SDH variables, the peak discharge Qp , its duration D, and selected as the driving component in the data selection and
volume Vp were chosen. in the modelling process. If one is, for example, interested
In first instance a review of several approaches yielding in the hydrograph volume for designing a reservoir for flood
design events available in the recent literature was provided regulation, it is essential to understand whether there is a pre-
focusing on how to apply these. As multiple variables were dominant driving variable. Specifically, if the reservoir is reg-
considered in the different return period approaches, an im- ulated by a levee, the volume design is related to a specific
portant aspect is (the modelling of) the dependence between discharge value. In this case the bivariate conditional distri-
variables. In this context the potential and the use of copulas bution could be preferred, since the discharge analysis is per-
for the construction of multivariate distribution functions was formed with a standard univariate approach and the volume
stressed and illustrated. On the one hand, a bivariate copula return period is estimated conditioned on the discharge de-
of (Qp , Vp ) was fitted. On the other hand, also the fitting of sign value. In similar practical problems, the regression anal-
the trivariate copula of (Qp , D, Vp ) was elaborated in a com- ysis could be preferred when the data availability does not
prehensive way by means of the vine copula approach. justify a richer statistical model application. On the contrary,
Eventually, design events for a 10 yr joint return pe- for instance, when the analyst is estimating the extension of
riod were obtained considering a 2-D regression-based, a flood inundation for which both peak discharge and volume
2-D conditional copula-based, a 2-D copula-based, a 2-D could play a similar role, a joint return period approach could
Kendall-based, and a 3-D Kendall-based approach. The tra- be appealing. Indeed, an ensemble of equally rare scenarios
ditional 1-D return period definition is considered as a ref- (i.e. having the same return period) could be used to assess
erence for comparison purposes. Differences in design quan- the variability of the obtained flood maps due to the selection
tiles were discussed while also the theoretical appropriate- of a single design event. Also in this case, it should be kept
ness was explained. This paper warns practitioners against in mind that the univariate n year return period is different to
blind use of just one available design event estimation ap- the bivariate and trivariate n year return period. Even though
proach, and stresses the importance of good copula fitting it is to be expected that including more variables improves
and the effect on the eventual design event outcome. A
the modelling of the process, one should keep in mind the Grimaldi, S., Petroselli, A., Alonso, G., and Nardi, F.: Flow time
drastically increasing need of data to fit such models. estimation with variable hillslope velocity in ungauged basins,
Adv. Water Resour., 33, 1216–1223, 2010.
Grimaldi, S., Petroselli, A., and Nardi, F.: A parsimonious geomor-
phological unit hydrograph for rainfall runoff modeling in small
Acknowledgements. The critical and useful comments of two
ungauged basins, Hydrol. Process., 57, 73–83, 2012a.
anonymous reviewers as well as G. Salvadori greatly improved this
Grimaldi, S., Petroselli, A., and Romano, N.: Green-Ampt Curve
work, for which the authors are very grateful. The interested reader
Number mixed procedure as an empirical tool for rainfall-
is referred to the public review process of this paper at HESSD for
runoff modelling in small and ungauged basins, Hydrol. Process.,
a discussion on JRP going beyond this publication. The research
doi:10.1002/hyp.9303, in press, 2012b.
of the first author is partially funded by the German Research
Grimaldi, S., Petroselli, A., and Serinaldi, F.: Design hydrograph
Foundation (DFG) under project PE 1632/4-1. S. Vandenberghe
estimation in small and ungauged watersheds: continuous sim-
was a doctoral research fellow of the Research Foundation Flanders
ulation method versus event-based approach, Hydrol. Process.,
(FWO). The Special Research Fund of Ghent University financially
26, 3124–3134, 2012c.
enabled the research of M. J. van den Berg.
Grimaldi, S., Petroselli, A., and Serinaldi, F.: A continuous simula-
tion model for design-hydrograph estimation in ungauged water-
Edited by: N. Romano
sheds, Hydrolog. Sci. J., 57, 1035–1051, 2012d.
Hobæk Haff, I., Aas, K., and Frigessi, A.: On the simplified pair-
copula construction – Simply useful or too simplistic?, J. Multi-
References var. Anal., 101, 1296–1310, 2010.
IGMI: Raster (Matrix) numerical DEM of Italy, http://www.igmi.
Aas, K. and Berg, D.: Models for construction of multivariate de- org/pdf/info matrix2003.png, 26. Feb. 2013, 2003.
pendence – a comparison study, Eur. J. Finance, 15, 639–659, Joe, H.: Multivariate Models and Dependence Concepts, Chapman
doi:10.1080/13518470802588767, 2009. and Hall, 1997.
Aas, K., Czado, C., Frigessi, A., and Bakken, H.: Pair-copula con- Kao, S.-C. and Govindaraju, R. S.: A bivariate frequency analysis of
structions of multiple dependence, Insur. Math. Econ., 44, 182– extreme rainfall with implications for design, J. Geophys. Res.,
198, doi:10.1016/j.insmatheco.2007.02.001, 2009. 112, D13119, doi:10.1029/2007JD008522, 2007.
Bedford, T. and Cooke, R.: Probability density decomposition for Kao, S.-C. and Govindaraju, R. S.: Trivariate statistical analysis of
conditionally dependent random variables modeled by vines, extreme rainfall events via the Plackett family of copulas, Water
Ann. Math. Artif. Intell., 32, 245–268, 2001. Resour. Res., 44, W02415, doi:10.1029/2007WR006261, 2008.
Bedford, T. and Cooke, R.: Vines – A new graphical model for de- Kao, S.-C. and Govindaraju, R. S.: A copula-based joint
pendent random variables, Ann. Stat., 30, 1031–1068, 2002. deficit index for droughts, J. Hydrol., 380, 121–134,
Berg, D.: Copula Goodness-of-fit testing: An overview and power doi:10.1016/j.jhydrol.2009.10.029, 2010.
comparison, Eur. J. Finance, 15, 675–701, 2009. Kojadinovic, I. and Yan, J.: Modeling multivariate distributions with
Brechmann, E. C. and Schepsmeier, U.: Modeling Dependence with continuous margins using the copula R package, http://www.
C- and D-Vine Copulas: The R Package CDVine, J. Stat. Softw., jstatsoft.org/v34/i09/, J. Stat. Softw., 34, 1–20, 2010.
52, 1–27, 2013. Kurowicka, D. and Cooke, R. M.: Sampling algorithms for
De Michele, C., Salvadori, G., Passoni, G., and Vezzoli, R.: A multi- generating joint uniform distributions using the vine cop-
variate model of sea storms using copulas, Coast. Eng., 54, 734– ula method, Comput. Stat. Data Anal., 51, 2889–2906,
751, 2007. doi:10.1016/j.csda.2006.11.043, 2007.
EEA: CORINE (Coordination of Information on Environment) Nelsen, R. B.: An Introduction to Copulas, Springer, 2006.
Database, a key database for European integrated environmen- Panchenko, V.: Goodness-of-fit test for copulas, Physica A, 355,
tal assessment, Programme of the European Commission, 2000. 176–182, 2005.
Favre, A. C., Adlouni, S. E., Perreault, L., Thiemonge, Pinya, M. A. S., Madsen, H., and Rosbjerg, D.: Assessment of the
N., and Bobee, B.: Multivariate hydrological frequency risk of inland flooding in a tidal sluice regulated catchment using
analysis using copulas, Water Resour. Res., 40, W01101, multi-variate statistical techniques, Phys. Chem. Earth, 34, 662–
doi:10.1029/2003WR002456, 2004. 669, doi:10.1016/j.pce.2009.03.001, 2009.
Genest, C. and Favre, A. C.: Everything you always wanted to know R Core Team: R: A Language and Environment for Statistical
about copula modeling but were afraid to ask, J. Hydrol. Eng., 12, Computing, R Foundation for Statistical Computing, http://www.
347–368, 2007. R-project.org/, Vienna, Austria, 2011.
Genest, C., Favre, A. C., Beliveau, J., and Jacques, C.: Metael- Salvadori, G.: Bivariate return periods via 2-copulas, J. Roy. Stat.
liptical copulas and their use in frequency analysis of mul- Soc. Ser. B, 1, 129–144, 2004.
tivariate hydrological data, Water Resour. Res., 43, W09401, Salvadori, G. and De Michele, C.: Frequency analysis via copulas:
doi:10.1029/2006WR005275, 2007. theoretical aspects and applications to hydrological events, Water
Grimaldi, S. and Serinaldi, F.: Asymmetric copulas in multivariate Resour. Res., 40, W12511, doi:10.1029/2004WR003133, 2004.
flood frequency analysis, Adv. Water Resour., 29, 1155–1167, Salvadori, G. and De Michele, C.: Statistical characterization of
2006a. temporal structure of storms, Adv. Water Resour., 29, 827–842,
Grimaldi, S. and Serinaldi, F.: Design hydrograph analysis with 3- 2006.
copula function, Hydrol. Sci. J. – J. Sci. Hydrol., 51, 223–238,
2006b.
Salvadori, G. and De Michele, C.: On the use of copulas in hydrol- Serinaldi, F. and Grimaldi, S.: Synthetic design hydrographs based
ogy: theory and practice, J. Hydrol. Eng., 12, 369–380, 2007. on distribution functions with finite support, J. Hydrol. Eng., 16,
Salvadori, G. and De Michele, C.: Multivariate Extreme Value 434–446, doi:10.1061/(ASCE)HE.1943-5584.0000339, 2011.
Methods, in: Extremes in a Changing Climate, edited by: Shiau, J. T.: Return period of bivariate distributed extreme hydro-
AghaKouchak, A., Easterling, D., Hsu, K., Schubert, S., and logical events, Stoch. Environ. Res. Risk Assess., 17, 42–57,
Sorooshian, S., Springer, 2013. 2003.
Salvadori, G., De Michele, C., Kottegoda, N., and Rosso, R.: Ex- Sklar, A.: Fonctions de répartition à n dimensions et leurs marges,
tremes in Nature, An Approach using Copulas, Springer, 2007. Publ. Inst. Statist. Univ. Paris, 8, 229–231, 1959.
Salvadori, G., De Michele, C., and Durante, F.: On the return period Song, S. and Singh, V. P.: Meta-elliptical copulas for drought fre-
and design in a multivariate framework, Hydrol. Earth Syst. Sci., quency analysis of periodic hydrologic data, Stoch. Environ.
15, 3293–3305, doi:10.5194/hess-15-3293-2011, 2011. Res. Risk Assess., 24, 425–444, doi:10.1007/s00477-009-0331-
Schertzer, D. and Lovejoy, S.: Physical modeling and analysis of 1, 2010.
rain and clouds by anisotropic scaling multiplicative processes, Theiling, C. H., and Burant, J. T.: Flood inundation mapping for
J. Geophys. Res., 92, 9693–9714, 1987. integrated floodplain management: Upper Mississippi River sys-
Serinaldi, F.: A multisite daily rainfall generator driven by bivariate tem, River Res. Appl., doi:10.1002/rra.2583, in press, 2012.
copula-based mixed distributions, J. Geophys. Res.-Atmos., 114, Vandenberghe, S., Verhoest, N. E. C., Buyse, E., and De Baets,
D10103, doi:10.1029/2008JD011258, 2009. B.: A stochastic design rainfall generator based on copulas
Serinaldi, F.: Multifractality, imperfect scaling and hydrological and mass curves, Hydrol. Earth Syst. Sci., 14, 2429–2442,
properties of rainfall time series simulated by continuous uni- doi:10.5194/hess-14-2429-2010, 2010.
versal multifractal and discrete random cascade models, Non- Vandenberghe, S., Verhoest, N. E. C., Onof, C., and De Baets, B.:
lin. Processes Geophys., 17, 697–714, doi:10.5194/npg-17-697- A comparative copula-based bivariate frequency analysis of ob-
2010, 2010. served and simulated storm events: A case study on Bartlett–
Serinaldi, F.: Interactive comment on “Joint return periods in Lewis modeled rainfall, Water Resour. Res., 47, W07529,
hydrology: a critical and practical review focusing on syn- doi:10.1029/2009WR008388, 2011.
thetic design hydrograph estimation” by S. Vandenberghe Volpi, E. and Fiori, A.: Design event selection in bivariate hydro-
et al., Hydrol. Earth Syst. Sci. Discuss., 9, C3721–C3721, logical frequency analysis, Hydrolog. Sci. J., 57, 1 –10, 2012.
http://www.hydrol-earth-syst-sci-discuss.net/9/C3721/2012/ Wong, G., Lambert, M. F., Leonard, M., and Metcalfe,
hessd-9-C3721-2012.png, 12 August 2012. A. V.: Drought analysis using trivariate copulas condi-
Serinaldi, F.: Closure to “Synthetic design hydrographs based on tional on climatic states, J. Hydrol. Eng., 15, 129–141,
distribution functions with finite support” by F. Serinaldi and doi:10.1061/(ASCE)HE.1943-5584.0000169, 2010.
S. Grimaldi, J. Hydrol. Eng., 18, 126–129, 2013. Yue, S. and Rasmussen, P.: Bivariate frequency analysis: discussion
Serinaldi, F. and Grimaldi, S.: Fully nested 3-copula: procedure and of some useful concepts in hydrological application, Hydrol. Pro-
application on hydrological data, J. Hydrol. Eng., 12, 420–430, cess., 16, 2881–2898, 2002.
2007. Zhang, L. and Singh, V. P.: Gumbel–Hougaard copula for trivariate
rainfall frequency analysis, J. Hydrol. Eng., 12, 409–419, 2007.