Stochastic Environmental Research and Risk Assessment 17 (2003) 42–57 Ó Springer-Verlag 2003

DOI 10.1007/s00477-003-0125-9

Return period of bivariate distributed extreme

hydrological events
J. T. Shiau

42
Abstract. Extreme hydrological events are inevitable and stochastic in nature.
Characterized by multiple properties, the multivariate distribution is a better
approach to represent this complex phenomenon than the univariate frequency
analysis. However, it requires considerably more data and more sophisticated
mathematical analysis. Therefore, a bivariate distribution is the most common
method for modeling these extreme events. The return periods for a bivariate
distribution can be defined using either separate single random variables or two
joint random variables. In the latter case, the return periods can be defined using
one random variable equaling or exceeding a certain magnitude and/or another
random variable equaling or exceeding another magnitude or the conditional
return periods of one random variable given another random variable equaling or
exceeding a certain magnitude. In this study, the bivariate extreme value distri-
bution with the Gumbel marginal distributions is used to model extreme flood
events characterized by flood volume and flood peak. The proposed methodology
is applied to the recorded daily streamflow from Ichu of the Pachang River
located in Southern Taiwan. The results show a good agreement between the
theoretical models and observed flood data.

Keywords: Return period, Bivariate extreme value distribution, Gumbel distri-

bution

1
Introduction
The return period for extreme hydrological events, such as floods and droughts, is
a common criterion employed in the design of hydraulic structures and water
supply systems. Traditionally, a univariate distribution is used to describe the
extreme hydrological phenomena, such as flood peak or rainfall intensity. How-
ever, a complex phenomenon is often characterized by multiple aspects. For
example, Goel et al. (1998) and Yue et al. (1999) indicated that flood flows appear
as multivariate events described by peak, volume and duration. Hence, the best
approach for analyzing such complex events is through the joint distribution of
several random variables, considering the correlations among them. Anderson
and Nadarajah (1993) and Anderson et al. (1994) used the multivariate extreme

J. T. Shiau
Department of Water Resources and Environmental Engineering,
Tamkang University, 151 Ying-chuan Road,
Tamsui, 251, Taiwan, ROC
e-mail: [email protected]
The author wishes to thank the two anonymous reviewers for their
constructive comments that improving the quality of this work.
model to study dependent structure between the maximum reservoir water level
and environmental factors including rainfall, catchment wetness, snowmelt, wind,
and reservoir fullness. Although a multivariate modeling approach offers im-
proved understanding and application, it requires considerably more data and
more sophisticated mathematical analysis. A bivariate distribution is the more
common and easier model for describing complex extreme hydrological events.
Gupta et al. (1976) developed an expression for the joint distribution function of
the largest flood peak and its time of occurrence. Singh and Singh (1991) derived
a bivariate probability density function with exponential marginal distributions to
study rainfall intensity and the corresponding depth. Bacchi et al. (1994) used the
bivariate distribution with marginal exponential distributions to model extreme 43
rainfall duration and severity. Goel et al. (1998) used the bivariate normal dis-
tribution to represent the joint distribution of flood peaks and volumes based on
a partial duration series. Yue et al. (1999) used the Gumbel mixed model, the
bivariate extreme value distribution with the standard Gumbel marginal distri-
butions, to represent the joint probability distribution of flood peaks and volumes
and the joint probability distribution of flood volumes and durations. Yue (2001)
used bivariate logistic distributions with the standard Gumbel marginal distri-
butions to model the El Niño maximum intensity and magnitude, El Niño mag-
nitude and duration, and El Niño maximum intensity and duration. Nadarajah
and Withers (2001) studied the joint behavior of the climate annual maxima for
New Zealand by the bivariate extreme value model.
The bivariate extreme hydrological event distributions and corresponding
return periods have been extensively studied. However, little attempt has been
made to interpret the return period characteristics defined using bivariate
distribution and the relationships between the return periods defined using
univariate and bivariate distributions. Fitting a specific bivariate distribution to
a specific extreme hydrological phenomenon was not the major purpose in this
study. Instead, after selecting a bivariate distribution to model an extreme
hydrological event, the return period and related probability distributions are
derived and the relationships between the univariate and bivariate distribution
return periods are investigated.
In this study, a theoretical derivation of the univariate and bivariate distri-
bution return periods is introduced based on the stochastic process concept.
Under the assumption that the bivariate extreme value distribution with the
Gumbel marginal distributions can be used to model the flood peak and volume,
the flood peak and volume return periods and the joint return period for the flood
peak and volume are derived. The associated properties and the relationships
between the return periods defined by the univariate and bivariate distributions
are also discussed. A recorded daily streamflow from Pachang River located in
Southern Taiwan is used as a case study to illustrate the proposed methodology.

2
Derivation of return period

2.1
Return period for a univariate distributed annual maximum series
According to the definition and notation used by Chow et al. (1988), the return
period for a univariate distributed annual maximum series is described below.
Suppose that an extreme event occurs if the random variable X is greater than or
equal to some magnitude xT . The recurrence interval TX is defined as the time
 period between occurrences for the event X
xT . The return period for the event
X
xT is the expected value of TX , denoted by EðTX Þ in this study. The return
period for an event of a given magnitude is thus defined as the average recurrence
interval between events equaling or exceeding a specific magnitude.
The return period for the event X
xT can be related to the probability of
occurrence for such events in the following way. It is assumed that the probability
of occurrence for the event X
xT in any year is PðX
xT Þ. Because the annual
maximum series observation in each year is independent, the probability of a
recurrence interval TX is the product of the probabilities for TX  1X < xT events
followed by one X
xT event, that is PðX < xT ÞTX 1 PðX
xT Þ. The expected
44 value for TX is then given by

X
1
1 1
EðTX Þ ¼ TX PðX < xT ÞTX 1 PðX
xT Þ ¼ ¼ ð1Þ
TX ¼1
PðX
xT Þ 1  PðX < xT Þ

The above equation is the return period for an extreme event described by a
single random variable and based on the annual maximum series. Obviously, the
return period depends on the distribution of the selected random variables, that
is, the longer the return period, EðTX Þ, the less the frequency, PðX
xT Þ, and the
larger the magnitude of the random variable, xT .

2.2
Return period for a univariate distributed partial duration series
The return period for a univariate distributed partial duration series can be
derived based on the stochastic process concept under the assumption that the
events are independently and identically distributed. The occurrences of inde-
pendent event X for different magnitudes x are shown in Fig. 1. Let L denote the
time period between any two successive events without consideration of the
magnitude, called the interarrival time in this study. The events with a magnitude
equal to or greater than any value x; X
x, are denoted by  in Fig. 1, while the
events with a magnitude less than any value x; X < x, are denoted by  in Fig. 1.
Hence, the time period between two events with magnitudes equaling or ex-
ceeding any magnitude x, namely the recurrence interval, denoted by TX , is equal
to the summation of the interarrival time for all events between them. This
relationship is expressed as

X
NX
TX ¼ Li ð2Þ
i¼1

where Li is the interarrival time between any two successive events and NX is the
number of events until the occurrence of the next event X
x.
Obviously, the recurrence interval TX is also a random variable and its
expected value is called the return period for X
x. Hence

Fig. 1. The occurrences of event X,  denoting X < x, and  denoting X
x

 !
X
NX
EðTX Þ ¼ E Li ¼ EðNX ÞEðLi Þ ð3Þ
i¼1

If the interarrival time, Li , is assumed to have an identical and independent

distribution, then the above equation can be simplified to

EðTX Þ ¼ EðNX ÞEðLÞ ð4Þ

The magnitude of an event is inversely related to its frequency of occurrence, with

very severe events occurring less frequently than more moderate events. NX is 45
therefore a random variable that depends on the distribution of X. Let F(x) denote
the cumulative distribution function of X, i.e. FðxÞ ¼ PðX xÞ. Because X is
considered as a continuous random variable, therefore, PðX xÞ ¼ PðX < xÞ
(Stone, 1996). The probability of an event with a magnitude equaling or exceeding
x, namely the probability of X
x, is

PðX
xÞ ¼ 1  PðX < xÞ ¼ 1  FðxÞ ð5Þ

The recurrence interval for event X
x is NX  1 X < x followed by one X
x.

That is, NX has a geometric distribution with parameter 1  FðxÞ and its proba-
bility mass function is given by

PðNX ¼ nÞ ¼ PðX < xÞn1 PðX
xÞ ¼ FðxÞn1 ½1  FðxÞ ; n ¼ 1; 2; 3; . . . ð6Þ

The expected value of NX is

1 1
EðNX Þ ¼ ¼ ð7Þ
PðX
xÞ 1  FðxÞ

The return period for events with a magnitude equal to or greater than x therefore
becomes

EðLÞ
EðTX Þ ¼ EðNX ÞEðLÞ ¼ ð8Þ
1  FðxÞ
Shiau and Shen (2001) applied this method to derive the return period of hy-
drological droughts of differing severity. The above procedure can be applied to
the partial duration series as well as the annual series. When applying this to the
annual maximum series, EðLÞ ¼ 1 year because each event is selected from one
year. Hence, Eq. (8) and Eq. (1) are identical.

2.3
Return period for a bivariate distributed partial duration series
Without a doubt, the above procedure can be directly applied to a bivariate dis-
tributed event. However, the correlation structure between random variables must
be considered. It is assumed that an extreme hydrological event can be characterized
using two random variables X and Y and the correlation coefficient q between them.
The return periods for a bivariate distributed event can be derived in two ways. The
first method treats each random variable separately, namely this method derives the
return period for random variable X and the return period for random variable Y.
The correlation between X and Y is not considered if X and Y are independent or
 treated in other ways. The occurrence of X
x and the occurrence of Y
y are
shown in Fig. 2(a) and (b), respectively. Apparently, the condition X
x and Y
y
may not occur simultaneously in one event. The return periods for X
x and Y
y
are therefore derived separately and given by

EðLÞ
EðTX Þ ¼ ð9aÞ
1  FðxÞ

EðLÞ
EðTY Þ ¼ ð9bÞ
1  FðyÞ
46
where E(L) is the expected value of interarrival time; EðTX Þ and EðTY Þ are the
return periods of X
x and Y
y, respectively; FðxÞ and F(y) are the cumulative
distribution functions of X and Y, respectively.
The above univariate frequency analysis is useful when only one extreme random
variable events is significant in the design criterion or these two random variables
are less dependent. However, a separate analysis of random variables X and Y
cannot reveal the significant correlation relationship between them if the correla-
tion is important information in the design criterion. The second method, on the
other hand, considers the random variables X and Y jointly. This can be done either
by defining the joint return periods for X and Y or by defining the conditional return
period for X given Y or vice versa. In this study, the joint return periods for X and Y
were defined in two cases: the return period for X
x or Y
y and the return
period for X
x and Y
y. Both methods are given below
EðLÞ EðLÞ
EðTXY Þ ¼ ¼ ð10Þ
PðX
x or Y
yÞ 1  Fðx; yÞ

0 EðLÞ EðLÞ
EðTXY Þ¼ ¼ ð11Þ
PðX
x and Y
yÞ 1  FðxÞ  FðyÞ þ Fðx; yÞ
Both of the above equations are defined using the magnitudes of X and Y si-
multaneously. These relationships imply that various combinations of values, x
and y, can result in the same return period. In addition, the return period can also
be defined by the event for X given Y
y or event for Y given X
x that are
called the conditional return period for X given Y
y and the conditional return
period for Y given X
x, respectively.

Fig. 2. The occurrences of events, (a)  denoting X < x,  denoting X
x, (b)  denoting
Y < y,  denoting Y
y
 EðTY Þ EðLÞ 1
EðTXjY
y Þ ¼ ¼
PðX
x; Y
yÞ 1  FðyÞ 1  FðxÞ  FðyÞ þ Fðx; yÞ
ð12aÞ
EðLÞ
¼
½1  FðyÞ ½1  FðxÞ  FðyÞ þ Fðx; yÞ

EðTX Þ EðLÞ 1
EðTYjX
x Þ ¼ ¼
PðX
x; Y
yÞ 1  FðxÞ 1  FðxÞ  FðyÞ þ Fðx; yÞ
ð12bÞ
EðLÞ
¼
½1  FðxÞ ½1  FðxÞ  FðyÞ þ Fðx; yÞ
47
where EðTX Þ and EðTY Þ are the return periods for X
x and Y
y; respectively.
Because the return period is the expected period for the recurrence interval, the
conditional distribution of Y given X
x and the conditional distribution of X
given Y
y can also be derived in the following.

PðX
x; Y < yÞ FðyÞ  Fðx; yÞ

FðyjX
xÞ ¼ ¼ ð13aÞ
PðX
xÞ 1  FðxÞ

PðX < x; Y
yÞ FðxÞ  Fðx; yÞ

FðxjY
yÞ ¼ ¼ ð13bÞ
PðY
yÞ 1  FðyÞ

3
Return period for a bivariate extreme value distribution
Kotz and Nadarajah (2000) provided a comprehensive survey of the bivariate
and multivariate extreme value models. However, not every bivariate distri-
bution can be applied to hydrological processes. Gumbel and Mustafi (1967)
proposed two general forms for bivariate extreme value distributions in term
of the marginal distributions. The cumulative distribution functions for the
bivariate extreme value distribution for Type A and Type B have the following
forms, respectively.
n h i1 o
1
h þ 1
ln FðxÞ ln FðyÞ
Fðx; yÞ ¼ FðxÞFðyÞe ð14Þ

where F(x) and F(y) are the marginal distributions for random variables X and Y,
respectively; and h is a parameter, 0 h < 1.
m 1
þð ln FðyÞÞm m
Fðx; yÞ ¼ e½ð ln FðxÞÞ ð15Þ

where m is a parameter, m
1.
Because the marginal distributions of the above two bivariate distributions are
the extreme value distributions, i.e. the Gumbel distribution, many researchers
used them to represent the extreme hydrological events. For instance, Yue et al.
(1999) used the Type A distribution, called the Gumbel mixed model, to represent
flood events. Yue (2001) used the Type B distribution, called the bivariate logistic
distributions with standard Gumbel marginal distributions, to model El Niño
events. The limitation of the Type A distribution is that the correlation of the
random variables must lies between 0 and 2/3. Hence, the Type B distribution was
 used as an example to illustrate how to derive the return period and associated
probability properties for a bivariate distribution in this study.
In Eq. (15), m is the parameter describing the correlation between X and Y. The
estimators for m is given by
1
m ¼ pffiffiffiffiffiffiffiffiffiffiffi ð16Þ
1q
where q is the correlation coefficient for X between Y.
The marginal distributions for X and Y take the Type I extreme value distri-
bution, namely the Gumbel distribution, expressed as
48
xb
 ax x
FðxÞ ¼ ee ð17aÞ
yby
 ay
FðyÞ ¼ ee ð17bÞ

where ax ; bx ; ay ; by are parameters.

Hence, the cumulative distribution function for the bivariate extreme value
distribution becomes
h mðxbx Þ mðyby Þ im1
 ay
 e ax þe
Fðx; yÞ ¼ e ð18Þ
Kotz et al. (2000) stated that such a distribution is also called the logistic model,
and Yue (2001) called it the bivariate logistic distributions with Gumbel marginal
distributions.
The probability density function, therefore, is
h mðxbx Þ mðyby Þ im1 1
m2
 ay
1  e ax þe mðxbx Þ
 
mðyby Þ

f ðx; yÞ ¼ e e ax þe ay
ax ay
" 1
# ð19Þ
mðxbx Þ mðyby Þ m mðxbx Þ mðyby Þ
    ay
e ax þe ay
þm  1 e ax

Hence, the return period of random variable X equal to or greater than x is

EðLÞ
EðTX Þ ¼ xb ð20Þ
 ax x
1  ee
where E(L) is the average interarrival time for the extreme hydrological events
under consideration.
The conditional distribution for Y given X
x is

h mðxbx Þ mðyby Þ im1

yby  ay
  e ax þe
e ay
e e
FðyjX
xÞ ¼ xb ð21Þ
 ax x
1  ee

The conditional return period for Y given X
x is

 EðLÞ
EðTYjX
x Þ ¼ 2 h im1 3
mðxbx Þ mðyby Þ
yby  ay
xbx
6 xbx   e ax þe 7
1e e ax e ax e ay
41  e e þe 5

ð22Þ

Similarly, the return period for random variable Y, the conditional distribution
for X given Y
y; and the conditional return period for X given Y
y are given
by
49
EðLÞ
EðTY Þ ¼ yby ð23Þ
 ay
1 ee
h mðxbx Þ mðyby Þ im1
 ay
xbx
 e ax þe
e ax
e e
FðxjY
yÞ ¼ yby ð24Þ
 ay
1 ee
EðLÞ
EðTXjY
x Þ ¼ 2 h im1 3
mðxbx Þ mðyby Þ
yby yby  ay
 6 xb
 a x   e ax þe 7
e ay e ay
1e 41  ee x  e þe 5

ð25Þ

The joint return periods for event X
x or Y
y and event X
x and Y
y are

EðLÞ
EðTXY Þ ¼ h im1 ð26Þ
mðxbx Þ mðyby Þ
 ay
 e ax þe
1e

0 EðLÞ
EðTXY Þ¼ h im1 ð27Þ
mðxbx Þ mðyby Þ
yby  ay
xbx   e ax þe
1e e ax e e ay þe
The parameters ax ; bx ; ay ; by and the average interarrival time E(L) must be
estimated from the data used, that is, these parameters are event-specific.

4
Applications

4.1
Data used
Flooding in Taiwan is a common phenomenon and often causes considerable
economic and life losses. As flooding can be represented by an extreme value
distribution (Gumbel, 1958), hence, flood flow is used as an example to illustrate
the proposed methodology. The Pachang River, with an 81 Km in length and
475 Km2 in drainage area, is an important river located in Southern Taiwan.
 Thirty-nine yearly daily streamflow records, from 1961 to 1999, were employed to
determine the return period for extreme flood events. In this study, the flood
events, shown in Fig. 3, were defined as a daily streamflow equaling or exceeding a
certain threshold. Two properties, flood volume (V) and flood peak (Q), were
used to characterize the extreme flood events. The flood duration was not con-
sidered in characterizing the flood events because most of the rivers in Taiwan are
so steep that only floods with a short duration occur. The flood peak is defined as
the maximum daily flow during the flood period, while the flood volume is
defined as the cumulative flow volume during the flood period.
The threshold was considered at 100 cms in this study, i.e. flood events are
50 defined as a daily streamflow equal to or greater than 100 cms. Fifty events were
abstracted from the recorded daily streamflow data. The correlation coefficient
between the flood volume and flood peak for these fifty flood events was 0.403.

4.2
Bivariate extreme value distribution of flood events
It was assumed that the Gumbel distribution could be used to represent both the
flood peak and flood volume for extreme flood events. Therefore, flood events
characterized by flood volume and flood peak can be fitted into a bivariate ex-
treme value distribution with the Gumbel marginal distributions. The correlation
coefficient between flood volume and flood peak was 0.403, hence m ¼ 1:294
according to Eq. (16). The parameters, a and b; of the Gumbel distribution were
estimated using the method of moments suggested by Yue (2001) and given by
pffiffiffi
6
a¼ s ð28Þ
p
b ¼ M  0:577a ð29Þ

where M and s are the mean and standard deviation of the sample data,
respectively.
Estimated from the observed flood events, the flood volume parameters are
aV ¼ 512:6, bV ¼ 1256:7 and the flood peak parameters are aQ ¼ 269:3, bQ ¼ 479:6.
The flood volume and flood peak cumulative distribution function is expressed as
h 1:294ðv1256:7Þ 1:294ðq479:6Þ
i0:773
 e 512:6 þe 269:3
Fðv; qÞ ¼ e ð30Þ

Figure 4 shows the contour plot, with probabilities 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7,
0.8, and 0.9, of cumulative distribution function of flood volume and flood peak.
The observed flood events are also shown in the Fig. 4.

Fig. 3. Characteristics of a
flood event
 The theoretical marginal distributions for flood volume and flood peak are
expressed in the following and shown in Figs. 5 and 6, respectively.
v1256:7
FðvÞ ¼ ee 512:6
ð31aÞ
q479:6
 269:3
FðqÞ ¼ ee ð31bÞ

The observed flood volume and flood peak for these fifty flood events are also
shown in Figs. 5 and 6. Compared to the theoretical Gumbel distributions, Eqs.
(31a) and (31b), the observed data are well fitted and accepted by the chi-square 51
goodness of fit test.

4.3
Return periods
The average interarrival time for these fifty flood events, defined as streamflow
equaling to or exceeding 100 cms, for the Ichu gauge station was 283.2 days,
namely 0.776 years. Based on the information estimated from the observed floods,
the return periods, expressed in years, for flood volume and flood peak are given by

0:776
EðTV Þ ¼ v1256:7
ð32aÞ
1  ee 512:6

0:776
EðTQ Þ ¼ q479:6 ð32bÞ

1  ee 269:3

Figure 7 illustrates the relationship between the return period and flood volume
and peak defined by Eqs. (32a) and (32b), respectively. If the n-year flood events
are defined as a return period equal to n years, the flood volume and flood peak
for 2, 5, 10, 20, 50, and 100-year flood events are listed in Table 1.

Fig. 4. The contour plots of bivariate distribution of flood volume and flood peak and
observed flood events
52

Fig. 5. The marginal distribution of flood volume

Fig. 6. The marginal distribution of flood peak

The conditional flood peak distribution given flood volume equaling or

exceeding a specific threshold is
h 1:294ðv1256:7Þ 1:294ðq479:6Þ
i0:773
q479:6
 e 512:6 þe 269:3
e 269:3
e e
FðqjV
vÞ ¼ v1256:7
ð33Þ
1  ee 512:6

If the specific threshold for flood volume in Eq. (33) is selected as the same
values in column 2 of Table 1, then the conditional flood peak distribution given 2,
5, 10, 20, 50, and 100-year flood events defined solely by the flood volume is shown
in Fig. 8. According to above information, one can estimate the return period for
flood peak exceeding a certain magnitude given flood volume equaling or ex-
ceeding any specific value. For example, given a 2-year flood event defined solely
by the flood volume, i.e., flood volume exceeding 1621 cms day, the return pe-
riod of flood peak equal to or greater than 704 cms is 2 years. Obviously, the above
information cannot be obtained from the univariate frequency analysis.
Similarly, the conditional flood volume distribution given flood peak equaling
or exceeding a certain magnitude is expressed in Eq. (34). The conditional dis-
tribution for flood volume given 2, 5, 10, 20, 50, and 100-year flood events defined
solely by the flood peak is shown in Fig. 9.
h 1:294ðv1256:7Þ 1:294ðq479:6Þ
i0:773
v1256:7  e 512:6 þe 269:3 53
ee 512:6
e
FðvjQ
qÞ ¼ q479:6 ð34Þ
 269:3
1  ee
The joint return period for flood volume equal to or greater than a certain value
or flood peak equal to or greater than another certain value, i.e. V
v or Q
q, is
shown in Fig. 10 and expressed as
0:776
EðTVQ Þ ¼ h i0:773 ð35Þ
1:294ðv1256:7Þ 1:294ðq479:6Þ
 e 512:6 þe 269:3
1e

Fig. 7. Return periods for flood volume and flood peak

Table 1. The 2, 5, 10, 20, 50, and 100-year flood events defined separately by flood volume
and flood peak

Return period (years) Flood volume (cms Æ day) Flood peak (cms)

2 1621 671
5 2169 959
10 2546 1157
20 2912 1349
50 3388 1599
100 3745 1787
 The joint return period for flood volume equal to or greater than a certain value
and flood peak equal to or greater than another certain value, V
v and Q
q; is
shown in Fig. 11 and expressed as

0 0:776
EðTVQ Þ¼ h i0:773 ð36Þ
1:294ðv1256:7Þ 1:294ðq479:6Þ
v1256:7
q479:6  e 512:6 þe 269:3
1  ee 512:6 e 269:3
e þe

54

Fig. 8. Conditional distribution of flood peak given 2, 5, 10, 20, 50, and 100-year flood
event defined by flood volume

Fig. 9. Conditional distribution of flood volume given 2, 5, 10, 20, 50, and 100-year flood
event defined by flood peak
 55

Fig. 10. Joint return period of flood volume and flood peak EðTVQ Þ

0
Fig. 11. Joint return period of flood volume and flood peak EðTVQ Þ

Not like the return period defined by a single random variable, the specific joint
return periods can be achieved using various combinations of the two random
variables. Hence, the joint return period for flood volume and flood peak must be
illustrated using the contour lines. However, the joint return periods EðTVQ Þ and
0
EðTVQ Þ exhibit different characteristics. The contour lines for various specific joint
0
return period years defined by EðTVQ Þ are bounded by the horizontal and vertical
axes, while the contour lines for various specific joint return period years defined by
0
EðTVQ Þ have no bounds. In addition, for the same value of v and q, EðTVQ Þ is greater
than EðTVQ Þ according to Eqs. (10) and (11). For example, v ¼ 1000 cms Æ day and
0
q ¼ 1000 cms, EðTVQ Þ ¼ 0:95 years, and EðTVQ Þ ¼ 6:23 years.
 4.4
Discussion
The bivariate distribution of extreme hydrological events can offer important and
useful information for hydrological structure design. However, comprehensively
interpreting and comparing the univariate frequency analysis is the key issue. If
the capacity of hydrological structures can be determined using flood peak or
flood volume separately and the correlation between them is not significant, then
Eq. (32a) or (32b) will be sufficient for the design criterion. If both the flood
volume and flood peak are important factors for design, then Eq. (32a) or (32b) is
insufficient because the correlation between them is not considered. For example,
56 the flood peak is an essential design factor for a reservoir spillway. If the reservoir
storage capacity is small compared to the flood inflow, then the flood volume is
also an important safety factor in the reservoir design. The joint return periods
for flood volume and flood peak can offer more useful information for design
criterion. The use of Eq. (35) or (36) as the design criterion depends on what
situations will destroy the structure. Under the condition that either flood peak or
flood volume exceeding a certain magnitude will cause damage, then Eq. (35) can
be used to evaluate the average recurrence interval. On the other hand, when the
flood volume and flood peak must exceed a certain magnitude that will cause
damage, then Eq. (36) is used.
It is worthwhile to observe the relationship between the joint return period for
the flood volume and flood peak and the return periods defined solely by the
flood peak or flood volume. Comparing Figs. 10 and 11, the values of horizontal
0
part of specific return period defined by EðTVQ Þ and EðTVQ Þ is the same and equal
to the return period defined by the flood peak solely. The values of vertical part of
0
specific return period defined by EðTVQ Þ and EðTVQ Þ is equal to the return period
defined by the flood volume solely. For example, a 10-year flood event defined by
the flood volume solely is 2546 cms day and a 10-year flood event defined by the
flood peak solely is 1157 cms. These two magnitudes are also the lower limits for a
10-year flood defined by EðTVQ Þ and the upper limits for a 10-year flood defined
0
by EðTVQ Þ.

5
Summary and conclusions
This study presented a methodology to define the return period for bivariate
distributed extreme hydrological events. If the extreme hydrological events must
be described using two random variables, the return periods for the extreme
hydrological events can be defined using a single random variable separately or
jointly using two random variables. If the correlation between random variables is
insignificant in the design criterion, the return periods for the extreme events can
be defined using one random variable. The joint consideration of two random
variables leads to a more complex return period form. The return periods for a
bivariate distribution can be defined in two ways. The first method defines the
return periods using one random variable equaling or exceeding a certain mag-
nitude and/or another random variable equaling or exceeding another certain
magnitude. The second method defines the conditional return periods for one
random variable given that another random variable equals or exceeds a specific
magnitude.
In this study, a bivariate extreme value distribution with the Gumbel marginal
distributions (also known as the logistic model) was used to fit extreme flood
events defined by a daily streamflow exceeding a specific threshold. Both the flood
volume and flood peak were then used to characterize the extreme flood events. If
the flood volume and flood peak are considered separately, the return periods for
extreme events can be defined using the flood volume or flood peak, as in the
traditional univariate flood frequency analysis. Joint consideration of the flood
volume and flood peak results in a return period definition in two different ways.
The first method defines the joint return periods for flood volume and/or flood
peak equaling or exceeding a certain magnitude. For the same magnitudes of
0
flood volume and peak, EðTVQ Þ is greater than EðTVQ Þ. The second method de-
fines the conditional return period for flood volume given flood peak, or vice
versa. In addition to the conditional return periods, the conditional flood volume
distribution given flood peak and the conditional flood peak distribution given 57
flood volume can also be derived. It is worthwhile to note that the contour lines
0
for various specific joint return period years defined by EðTVQ Þ are bounded by
the horizontal and vertical axes, while the contour lines for various specific joint
return period years defined by EðTVQ Þ have no bounds.
The proposed methodology was applied to the recorded daily streamflow from
the Ichu gauge station of the Pachang River located in Southern Taiwan. The
results exhibited a good agreement between the theoretical models and observed
data and provided more useful information than the univariate random variable
frequency analysis.

References
Anderson CW, Dwyer IJ, Nadarajah S, Reed DW, Tawn JA (1994) Maximum reservoir
water level. In Reservoir Safety and Environment, pp. 200–213, Thomas Telford, London
Anderson CW, Nadarajah S (1993) Environmental factors affecting reservoir safety. In
Barnett V, Turkman KF (eds), Statistics for the Environment pp. 163–182. John Wiley &
Sons, Chichester
Bacchi B, Becciu G, Kottegoda NT (1994) Bivariate exponential model applied to intensities
and duration of extreme rainfall. J. Hydrol., 155: 225–236
Chow VT, Maidment DR, Mays LR (1988) Applied Hydrology, McGraw-Hill Inc., New York
Goel NK, Seth SM, Chandra S (1998) Multivariate modeling of flood flows. ASCE J. Hy-
draul. Eng., 124(2): 146–155
Gumbel EJ (1958) Statistics of Extremes. Columbia University Press, New York
Gumbel EJ, Mustafi CK (1967) Some analytical properties of bivariate extremal distribu-
tions. J. Am. Statist. Assoc., 62: 569–588
Gupta VK, Duckstein L, Peebles RW (1976) On the joint distribution of the largest flood
and its time of occurrence. Water Resour. Res., 12(2): 295–304
Nadarajah S, Withers CS (2001) Modeling dependency between climate extremes for New
Zealand. World Resour. Rev., 13(4): 526–539
Kotz S, Nadarajah S (2000) Extreme Value Distributions: Theory and Applications, Im-
perial College Press, London
Kotz S, Balakrishnan N, Johnson NL (2000) Continuous Multivariate Distribution, Vol 1,
Models and Applications, John Wiley & Sons, Inc., New York
Shiau JT, Shen HW (2001) Recurrence analysis of hydrologic droughts of differing severity.
ASCE J. Water Resour. Plng. and Mgmt., 127(1): 30–40
Singh K, Singh VP (1991) Derivation of bivariate probability density functions with ex-
ponential marginals. Stochast. Hydrol. Hydraul., 5(1): 55–68
Stone CJ (1996) A Course of Probability and Statistics, Wadsworth Publishing Company,
New York
Yue S, Ouarda TBMJ, Bobee B, Legendre P, Bruneau P (1999) The Gumbel mixed model for
flood frequency analysis. J. Hydrol., 226: 88–100
Yue S (2001) A statistical measure of severity of El Niño events. Stochast. Envir. Res. and
Risk Assess., 15(2): 153–172