Hydrological Sciences Journal

ISSN: 0262-6667 (Print) 2150-3435 (Online) Journal homepage: www.tandfonline.com/journals/thsj20

Comparison between the peaks-over-threshold

method and the annual maximum method for
flood frequency analysis

Nejc Bezak, Mitja Brilly & Mojca Šraj

To cite this article: Nejc Bezak, Mitja Brilly & Mojca Šraj (2014) Comparison between the
peaks-over-threshold method and the annual maximum method for flood frequency analysis,
Hydrological Sciences Journal, 59:5, 959-977, DOI: 10.1080/02626667.2013.831174

To link to this article: https://doi.org/10.1080/02626667.2013.831174

Published online: 01 May 2014.

Submit your article to this journal

Article views: 13304

View related articles

View Crossmark data

Citing articles: 53 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=thsj20
 Hydrological Sciences Journal – Journal des Sciences Hydrologiques, 59 (5) 2014 959
http://dx.doi.org/10.1080/02626667.2013.831174

Comparison between the peaks-over-threshold method and the annual

maximum method for flood frequency analysis
Nejc Bezak, Mitja Brilly and Mojca Šraj
Faculty of Civil and Geodetic Engineering, University of Ljubljana, Jamova 2, SI-1000 Ljubljana, Slovenia
mojca.sraj@fgg.uni-lj.si

Received 18 January 2013; accepted 19 June 2013; open for discussion until 1 November 2014

Editor D. Koutsoyiannis

Citation Bezak, N., Brilly, M., and Šraj, M., 2014. Comparison between the peaks-over-threshold method and the annual maximum
method for flood frequency analysis. Hydrological Sciences Journal, 59 (5), 959–977.

Abstract Flood frequency analysis can be made by using two types of flood peak series, i.e. the annual maximum
(AM) and peaks-over-threshold (POT) series. This study presents a comparison of the results of both methods for
data from the Litija 1 gauging station on the Sava River in Slovenia. Six commonly used distribution functions
and three different parameter estimation techniques were considered in the AM analyses. The results showed a
better performance for the method of L-moments (ML) when compared with the conventional moments and
maximum likelihood estimation. The combination of the ML and the log-Pearson type 3 distribution gave the best
results of all the considered AM cases. The POT method gave better results than the AM method. The binomial
distribution did not offer any noticeable improvement over the Poisson distribution for modelling the annual
number of exceedences above the threshold.
Key words flood frequency analysis; annual maximum series; peaks-over-threshold method; threshold selection; method of
L-moments; maximum likelihood

Comparaison entre les méthodes de dépassement de seuil et du maximum annuel pour les
analyses de fréquence des crues
Résumé Les analyses de fréquence des crues peuvent être réalisées en utilisant deux types de séries de débits de
crue, soit les maximums annuels (MA), soit les débits dépassant un certain seuil (POT pour Peak Over
Threshold). La présente étude présente une comparaison des résultats des deux méthodes pour les données de
la station de Litija 1 sur la rivière Sava en Slovénie. Six fonctions de distribution couramment utilisées et trois
techniques d’estimation des paramètres ont été pris en compte dans les analyses des MA. Les résultats ont montré
une meilleure performance de la méthode des L-moments par rapport aux moments classiques et à la méthode du
maximum de vraisemblance. La combinaison de la méthode des L-moments et de la distribution log-Pearson 3 a
donné les meilleurs résultats parmi tous les cas considérés sur les MA. La méthode POT a donné de meilleurs
résultats que la méthode MA. La distribution binomiale n’a offert aucune amélioration notable par rapport à la
distribution de Poisson pour modéliser le nombre annuel de dépassement du seuil.
Mots clefs analyse de fréquence des crues ; série des maximums annuels ; méthode des débits dépassant un seuil ; choix d’un
seuil ; méthode des L-moments ; maximum de vraisemblance

INTRODUCTION failure of a structure is an important part of engineer-

Quality discharge measurements or water-level mea-
surements and effective flood frequency analyses
(FFA) are the precondition for reliably defining
design discharge (Bonacci and Ljubenkov 2008,
Šraj et al. 2008). Estimation of a flood magnitude
corresponding to a chosen risk of flooding or even
ing practice (Khaliq et al. 2006). Improvements to
FFA techniques are still the goal of hydrologists,
because efficient planning and design of water struc-
tures require good estimates of the behaviour of
extreme hydrological events, which can be highly
destructive and cause major economic losses, envir-
onmental damage and even loss of life (Mikoš et al.

© 2014 IAHS Press

960 Nejc Bezak et al.
2004, Brilly and Polič 2005). The relationship
between discharge and return period is unique for
every gauging station. The traditional annual maxi-
mum (AM) series approach is still the most used
method in FFA in many countries, including
Slovenia. An AM sample is defined by the maximum
peak flow of each year. However, defining the sam-
ples in such a way can result in loss of information
(Langbein 1949, Lang et al. 1999, Bačová-Mitková
and Onderka 2010). For example, some peaks that
were not AMs, but are still relatively high, are not
considered in AM analyses. However, some very low
discharge values can be part of the AM series sample.
Another shortcoming of the AM series is that the
sample is usually rather small. An alternative to the
AM method is the peaks-over-threshold (POT)
approach (also called the partial duration series
approach). The POT sample is defined by all peak
values that lie above a certain truncation level
(usually called the threshold or base level) (Lang
et al. 1999). Major difficulties in using the POT
method are assuring the independence of the data
series and choosing an appropriate threshold value.
The next step is to select an appropriate statisti-
cal distribution function for the time series of extreme
discharges, which describes the random process best.
Several distribution functions are available for mod-
elling magnitudes of exceedences. Generally more
than one distribution may fit the data well and select-
ing the best model can be difficult (Salas et al. 2012).
Many countries have developed guidelines suggest-
ing which distribution to use. However, finding the
best method is still a challenge for engineers. Next,
the estimation technique for distribution parameters
has to be chosen. The distribution parameters can be
estimated through different methods, such as the
method of moments (MOM), the method of
L-moments (ML) and the maximum likelihood esti-
mation (MLE).
Sankarasubramanian and Srinivasan (1999) com-
pared the MOM and the ML for parameter estimation
of the generalized normal, generalized extreme value
(GEV), Pearson type 3 (P3) and generalized Pareto
distributions. The performance of the methods
depended on the sample size and skewness of data.
Ahilan et al. (2012) analysed data from 172 gauging
stations in Ireland and investigated which of the
extreme value (EV) distributions is most appropriate
for FFA. The Gumbel distribution was preferred over
two other types of EV distributions (EV II or Frechet,
and EV III or Weibull). Önöz and Bayazit (1995)
investigated data from 19 gauging stations. Seven
probability distribution functions were compared and
the parameters were estimated with the MOM and the
probability weighted moments (PWM). Chi-square,
Kolmogorov-Smirnov (K-S), Anderson-Darling
(A-D), probability plot correlation coefficient (PPCC)
and other tests were used to assess the performance of
the individual cases. Generally, the GEV distribution
gave better results than the Gumbel, log-normal (LN),
Pearson type 3 (P3), log-Pearson type 3 (LP3) and
other distribution functions. Seckin et al. (2011) made
FFA on the data from 543 gauging stations in Turkey.
The ML was used for estimation of the distribution
parameters. The GEV distribution gave better good-
ness-of-fit values than other distributions, such as the
generalized logistic (GL) and LP3 distributions.
Shane and Lynn (1964) were the first to propose
the Poisson distribution for modelling the annual
number of exceedences and an exponential distribu-
tion for the magnitudes of exceedences. Some other
distributions were also used for modelling the annual
number of events above the threshold and description
of the magnitudes of exceedences. Besides the
Poisson distribution (Todorovic and Zelenhasic
1970, Cunnane 1979, Tavares and da Silva 1983,
Kjeldsen et al. 2000), one can also choose between
the negative binomial (Cunnane 1979, Lang et al.
1997, Önöz and Bayazit 2001, Bhunya et al. 2012)
and binomial (Lang et al. 1997, Robson and Reed
1999, Önöz and Bayazit 2001) distributions.
Cunnane (1979) made analyses on the data from 26
gauging stations in Great Britain; however, the nega-
tive binomial distribution function did not seem to
offer any substantial improvement over the Poisson
distribution. Bhunya et al. (2012) found that the
Poisson distribution was better than the negative
binomial distribution in cases when the difference
between the mean and the variance of the annual
number of exceedences was small. They used the
GL distribution function for modelling the magni-
tudes of events above the threshold. Other distribu-
tions, such as the generalized Pareto (Van Montfort
and Witter 1985, Rosbjerg et al. 1992, Lang et al.
1999), LN (Adamson and Zucchini 1984, Rosbjerg
1987) and Weibull (Pekarova et al. 2008, Bačová-
Mitková and Onderka 2010), can also be used for
modelling magnitudes. Rosbjerg et al. (1992) applied
the MOM and PWM for parameter estimation for the
partial duration series. They found that if the para-
meters were estimated with MOM, the result of the
statistic mean square error (MSE) was smaller than in
the case of PWM. Önöz and Bayazit (2001) derived
expressions for the asymptotic sampling variance of
 Comparison between POT and AM methods
the estimate of the T-year event. Cunnane (1979)
suggested a simple test statistic, which checks the
suitability of the Poisson distribution for modelling
the annual number of events above threshold. Van
Montfort and Witter (1985) proposed two statistical
tests to check the adequacy of the exponential dis-
tribution for modelling the magnitudes of
exceedences.
The main drawbacks of partial duration series
are the threshold selection and assuring independence
criteria. Cunnane (1979) and USWRC (1982) pro-
posed two different types of independence criteria.
Tavares and da Silva (1983) suggested that the
threshold value should be as low as possible provid-
ing that model assumptions are still valid. Ashkar and
Rousselle (1983a) found that if the Poisson and
exponential distributions are applicable with a certain
threshold, this should remain so with any higher
threshold value. They also showed that the threshold
should be so high that the Poisson model is still valid.
Langbein (1949) suggested that the base level value
should be equal to the lowest AM series event. Lang
et al. (1999) suggested some graphical tests that
should be used to make the appropriate threshold
selection. Begueria (2005) also investigated the influ-
ence of threshold selection on partial duration series
model assumptions, stating that a unique optimum
truncation level cannot be found. Silva et al. (2012)
also investigated the influence of threshold selection
on FFA results. They analysed streamflow and rain-
fall data from Portuguese watersheds, while the tests
proposed by Lang et al. (1999) were used to assess
the POT analysis results. The Flood Estimation
Handbook (Robson and Reed 1999) states that the
main advantage of the POT method over the AM
method is for the case of short data series (less than
14 years of data).
The dilemma of which approach to use is not
new. In one of the first articles concerning partial
duration series, Langbein (1949) investigated the
relationship of recurrence intervals of floods on
annual and partial duration series bases. He found
that the difference between return periods con-
verges to 0.5 years as the return period increases.
The POT method contains more information about
flood peaks than the AM series, because the sam-
ple is larger. Much research has compared the AM
and partial duration series methods. Tavares and da
Silva (1983) found that the partial duration series
method had a lower estimation of the asymptotic
sampling variance than the AM method, assuming
that, at least on average, two peaks per year were
961
selected. The magnitudes were modelled with the
exponential distribution, while the Poisson distribu-
tion was used for the distribution of the annual
number of exceedences. Cunnane (1973) reported
that on average at least 1.65 peaks per year should
be selected to achieve an advantage over the AM
series method. An EV type 1 distribution was
applied to the AM series, the MLE was used for
estimation of the distribution’s parameters and the
sampling variance was used for comparison.
Madsen (1996) performed similar analyses for the
MOM and PWM. He found out that if a partial
duration series contains more than, on average,
0.91 peaks per year, the variance, which was
again chosen as the performance index, is smaller
than in the case of the Gumbel distribution.
Recently, Bhunya et al. (2012) compared different
combinations of the partial duration and AM series
methods. The POT model has more physical back-
ground because it is compounded from the distri-
bution of the annual number of exceedences above
the threshold and their magnitudes (Önöz and
Bayazit 2001). Bačová-Mitková and Onderka
(2010) used data from five gauging stations on
the Danube River to compare the AM and partial
duration series. Which of these two methods gave
higher estimated discharge values depended on the
data properties and not just on the method selected
for FFA.
The objectives of this study were: (a) to explore
the influence of the threshold value in the POT
method on FFA results, (b) to compare different
statistical distribution functions and evaluate three
parameter estimation techniques (MOM, ML and
MLE) and (c) to compare the POT and the AM series
method.
DATA
The Litija gauging station on the Sava River is one of
the oldest stations in Slovenia (Kobold 2011)
(Fig. 1). The Sava River basin is the largest in
Slovenia (Šraj 2001) covering an area of
10 746 km2 and is part of the Danube River basin,
which is located in Central and Eastern Europe
(Pekarova et al. 2008). The Litija 1 station has a
drainage area of 4281.43 km2 and lies at 230 m
a.s.l. (Frantar and Hrvatin 2008). The station is
equipped with a Seba Omega recorder (limnigraph).
Measurements began in 1893 and daily discharge
values are available from 1895. However, in 1953
the station was relocated and the station’s name was
962
Fig. 1 Location of the Litija 1 gauging station.
changed from Litija to Litija 1. A period of 116 years
of data was used for FFA (1895–2010) to compare
the AM series and partial duration series methods. A
comparison between distribution functions, parameter
estimation techniques and different threshold values
was performed on data for the period 1953–2010.
An Alpine pluvial-nival discharge regime is typi-
cal for this station. The characteristic of this regime is
two fairly equal peaks in autumn and spring. In the
case of Litija station, the main peak is in autumn
(Frantar and Hrvatin 2005). Below average water
quantities usually occur in winter (January–March)
and in summer (July–September). The seasonal beha-
viour of AM events is shown in Fig. 2, which indi-
cates that most of the annual peaks occurred between
October and December, when the density of events is
greatest. The resultant of all AMs also occurs in the
autumn period (Fig. 2).
Fig. 2 Seasonality of annual maximums.
Nejc Bezak et al.
METHODS
AM series, partial duration series and time series
models can be used for FFA (WMO 1989). The
POT modelling provides additional flexibility and a
more comprehensive description of the flood process
(Lang et al. 1999). Besides the magnitudes of excee-
dences, the POT model also considers the annual
number of peaks above the threshold. The return
period for the AM model differs from that of the
POT model: TPOT is the average interval between
the high flows that exceed threshold value Q; TAM
is the average interval between the years containing a
flood with flow of at least Q. This definition of the
return period is mostly used in hydrology. The rela-
tionship between the return periods is defined using
the following equation (Langbein 1949):


1
1
TAM ¼ 1  exp  (1)
TPOT
POT model
The POT model can be composed of the Poisson,
binomial and negative binomial distributions for
modelling the annual number of events above thresh-
old, and of exponential or generalized Pareto distri-
butions for magnitudes of exceedences. The
dispersion index test (Cunnane 1979) and the Chi-
square test were used to check the adequacy of the
 Comparison between POT and AM methods
chosen distribution function for modelling the annual
number of exceedences.
pffiffiffi The statistic G (maximum/
median) and r n test, both proposed by Van
Montfort and Witter (1985), were used to check the
adequacy of the exponential distribution against the
Pareto distribution for modelling the magnitudes of
peaks above threshold. The POT method includes the
steps discussed in the continuation.
Independence criteria The first step of the POT
method is consideration of the independence condi-
tions. Cunnane (1979) introduced the criterion that
consecutive peaks must be separated by three times
TP (average time to peak of five hydrographs), or that
the smallest discharge value between two consecutive
peaks must be higher than two-thirds of the first peak
value. These two conditions were also considered in
the POT analyses on the Danube River (Bačová-
Mitková and Onderka 2010) and in analyses of the
data from Portuguese watersheds (Silva et al. 2012).
The Water Resources Council suggested (USWRC
1982) the following criteria:
θ < 5 days þ logðAÞ
(2)
or xMIN < ð3=4Þ min½xS1 ; xS2 
where A is basin area in square miles, and xS1 and xS2
are two consecutive peak values. The second peak
should be rejected if one of the conditions in equation
(2) is valid. These conditions were also considered in
the POT analyses for Litija 1 station.
Ashkar and Rousselle (1983b) suggested that
independence criteria should be considered only if
successive peaks are correlated and none of the other
possible steps for reducing correlation is possible.
They also showed that the main weakness of the
first condition in equation (2) occurs in the case of
larger basin areas. It will be shown in the Results
Section that the consideration of independence cri-
teria for Litija 1 station is reasonable. The difference
between the average daily discharge value and the
local maximum is relatively small. This means that
only in the case of very large threshold values is the
average daily discharge value rejected and so only
the local maximum is considered in the POT sample.
If a lower threshold value is chosen, both the average
daily discharge value and the local maximum are
considered. Both values are part of one flood event
and it is not reasonable to consider one event more
than once. This is one of the reasons why the use of
independence criteria is necessary.
963
Threshold selection Defining the threshold
value x0 is one of the main disadvantages of the
POT method. The selection of the threshold is a
subjective process. It is important that the threshold
value selected is high enough, so that the model
assumptions are not violated (Lang et al. 1999). But
the truncation level should be selected as low as
possible, so that the highest number of exceedences
is selected and more reliable parameter estimates can
be made (Ashkar and Rousselle 1987).
Langbein (1949) suggested that the truncation
level be selected as equal to the lowest AM event.
This means that at least one event per year is
included in the analyses. Madsen et al. (1993) recom-
mended the use of the standard frequency factor k
and data properties (mean value and standard devia-
tion). The threshold value can be calculated using the
following expression:
x0 ¼ μx þ kσ x (3)
The frequency factor k with value 3 can be selected.
The low k value means that a low threshold value is
selected and more events are included in the FFA. In
the Flood Estimation Handbook (Robson and Reed
1999), the threshold value is defined so that, on
average, one, three or five events per year are
selected. On average, four events per year were
selected for the Danube River analyses (Bačová-
Mitková and Onderka 2010). Some researchers have
suggested using a return period as a basis for thresh-
old selection. A return period of 1.2–2 years was
used by Irvine and Waylen (1986); Lang et al.
(1999) proposed three tests for threshold selection.
The results of the tests give an interval of possible
threshold values. One of these tests confirms the
hypothesis of the Poisson process. This graphical
test is based on the dispersion index (Cunnane
1979). An alternative to these possible truncation
levels that have no physical background is to set the
threshold as the overflowing water level value (bank-
full level). In the latter case, only floods are consid-
ered in analyses.
Distributions for modelling the annual num-
ber of exceedences above threshold x0 The Poisson
distribution is the first choice for modelling the
annual number of exceedences m. The alternatives
are the binomial and negative binomial distributions.
If the mean value of m is the same as the var-
iance of m, then the Poisson distribution is appropri-
ate for modelling the annual number of events above
964 Nejc Bezak et al.

a certain threshold level. The probability mass func- Instead of the above-mentioned distributions, one
tion of the Poisson distribution and the first three can also use the LN distribution (Adamson and
moments are (Önöz and Bayazit 2001): Zucchini 1984, Rosbjerg 1987), the Weibull
(Bačová-Mitková and Onderka 2010) or the GL
eμ μk (Bhunya et al. 2012) distribution functions. The
Pðm ¼ kÞ ¼ k ¼ 0; 1; 2; :::
k! cumulative distribution function of the exponential
EðmÞ ¼ E ¼ μ (4) distribution and the first two moments are (Önöz
and Bayazit 2001):
varðmÞ ¼ V ¼ μ
μ3 ¼ μ

y
GY ð yÞ ¼ 1  exp  y  0
β
If the mean value of m is larger than the variance of (8)
the annual number of exceedences, the binomial dis- Eð yÞ ¼ E ¼ β
tribution can be used. This means that the time dis- varð yÞ ¼ V ¼ β2
tance between the events above the threshold is
approximately constant. The probability mass func- If the magnitudes are not exponentially distributed,
tion of the binomial distribution and the first three then the generalized Pareto distribution is one of the
moments of m are (Önöz and Bayazit 2001): possible alternatives. The cumulative distribution

 function and the first two moments of the Pareto
γ k
Pðm ¼ kÞ ¼ α ð1  αÞγk k ¼ 0; 1; 2; ::: distribution are (Lang et al. 1999):
k
 1=k
EðmÞ ¼ E ¼ μγ (5) k
GY ð yÞ ¼ 1  1  ð yÞ kÞ0
varðmÞ ¼ V ¼ αð1  αÞγ β
μ3 ðmÞ ¼ αð1  αÞð1  2αÞγ β
Eð yÞ ¼ (9)
1þk
If the mean value of the annual number of excee- β2
dences is smaller than the variance of m, the negative varð yÞ ¼
ð1 þ k Þ2 ð1 þ 2k Þ
binomial distribution is used as an alternative to the
Poisson distribution. This means that the peaks some-
If the distribution for modelling the magnitudes of
times occur in bunches. The probability mass func-
exceedences is combined with the distribution for
tion of the negative binomial distribution and the first
modelling the annual number of peaks above thresh-
three moments are (Önöz and Bayazit 2001)
old, the following expression can be derived (Shane

 and Lynn 1964):
γþk1 k
Pðm ¼ kÞ ¼ α ð1  αÞγ
k X
1  k
k ¼ 0; 1; 2; ::: FX ð x Þ ¼ Pðm ¼ k Þ Gy ð yÞ (10)
αγ k¼0
EðmÞ ¼ E ¼
ð1  αÞ (6)
αγ The POT model with exponentially distributed peaks
varðmÞ ¼ V ¼ yields the EV type I distribution of AMs, and, when
ð1  αÞ2 the peaks are modelled with the generalized Pareto
αγð1 þ αÞ distribution, this leads to the GEV distribution
μ3 ðmÞ ¼
ð1  αÞ3 (Madsen et al. 1997, Lang et al. 1999, Önöz and
Bayazit 2001). The following probability distribution
functions are combinations of Poisson and exponen-
Probability distributions for POT modelling tial distributions (Önöz and Bayazit 2001), binomial
Exponential and generalized Pareto distributions can and exponential distributions (Önöz and Bayazit
be used for modelling the magnitudes of 2001), negative binomial and exponential distribu-
exceedences: tions (Önöz and Bayazit 2001) and Poisson and gen-
eralized Pareto distributions (Lang et al. 1999),
y ¼ x  x0 (7) respectively:
 Comparison between POT and AM methods
 
x  x0
FX ð xÞ ¼ exp μ exp  (11a)
β

 γ
γ α x  x0
FX ð xÞ ¼ ð1  αÞ 1þ 1  exp 
1α β
(11b)

 γ
x  x0
FX ð xÞ ¼ ð1  αÞγ 1  α 1  exp 
β Danube River (Bačová-Mitková and Onderka 2010).
(11c)
( 
 1=k )
k
FX ð xÞ ¼ exp μ 1  ðx  x0 Þ (11d)
β are shown in Table 1.
where FX(x) is the distribution function. The expres-
sions for the estimate of the T-year event (discharge)
can be easily obtained from equations (11a–d).
Annual maximum series model
The AM series method is still frequently used in
FFA. Using the AM series method there are no pro-
blems with independence criteria (usually), threshold
selection and distribution of exceedences. The LN,
P3, LP3, Gumbel, GEV and GL distributions were
used for FFA with the AM series on the Sava River
(Litija 1 station). The Gumbel distribution is a special
case of the GEV distribution, while the GL and GEV
distributions are special cases of the four-parameter
Kappa distribution. The MOM, the ML and the MLE
were used for parameter estimation in the AM
method. The MOM gives good results in the case
of symmetrical data, but, in most practical cases,
flood data are asymmetrical, in which case the ML
gives better results. L-moments are analogous to
conventional moments and are more robust and less
sensitive to outliers; they also give better parameter
estimates for small samples (Hosking 1990). L-
moments are linear functions of PWM and are
described in detail by Hosking and Wallis (2005).
The MOM equates the sample moments with theore-
tical moments of the distribution function. This
method is simple, but it is known that the numerical
values of sample moments can be rather different
from those of the population from which the sample
has been drawn (Sankarasubramanian and Srinivasan
1999). The MLE defines the distribution parameters
in such way that the probability of occurrence of our
sample is maximal (Haktanir and Horlacher 1993).
965
The LP3 distribution is mostly used in FFA in
Slovenia. Using frequency factors is the simplest way
of obtaining design discharges (Haktanir 1991). This
distribution function was also suggested by the Water
Resources Council (USWRC 1982). The GEV and GL
distributions were proposed in the Flood Estimation
Handbook (Robson and Reed 1999) for FFA on data
from the rivers in Great Britain. LN distribution and
LP3 distribution were used for the AM analyses on the
Karim and Chowdhury (1995) used the LN, Gumbel,
LP3 and GEV distributions for FFA in Bangladesh.
The distribution functions and equations for para-
meters estimation that were used for the AM analyses
Goodness-of-fit tests
One of the purposes of this article was to find which
distribution function gives the best fit with the AM
series sample. The Kolmogorov-Smirnov, Anderson-
Darling and PPCC tests were used to check the
adequacy of the tested distribution functions. The
root mean square error (RMSE), mean absolute
error (MAE), relative mean absolute error (RMAE)
and Akaike information criterion (AIC) were used to
define the best fit distribution. These tests were also
used to compare the MOM, the ML and the MLE for
estimation of the distribution parameters.
The RMSE, MAE, RMAE and PPCC tests were
also used to define with which truncation level the
POT model gave the best fit to data. Different thresh-
old values were used to define different POT samples.
The results of K-S, A-D and PPCC tests can also be
used to compare different distribution functions.
Critical values for different significance levels are
easily found in the literature (Haan 2002, Kottegoda
and Rosso 2008), but if the distribution function para-
meters are estimated from the sample used in the test,
the critical values are modified. Different authors have
provided critical values for the PPCC test for different
distributions: Vogel (1986) for the Gumbel, normal
and LN distributions, Chowdhury et al. (1991) for
the GEV distribution, Vogel and McMartin (1991)
for P3 and LP3 distributions and Kim et al. (2008)
for GL and generalized Pareto distributions. Modified
critical values for the K-S test can be found in the
literature for the following distribution functions: nor-
mal and LN (Lilliefors 1967, Crutcher 1975), GEV
(Chandra et al. 1981, Chowdhury et al. 1991),
Gumbel and gamma (Crutcher 1975) and the logistic
distribution (Stephens 1979). Different authors have
966 Nejc Bezak et al.

Table 1 Equations for parameter estimation of different distributions with the MOM, ML and MLE methods for the log-
normal (LN), Pearson type 3 (P3), log-Pearson type 3 (LP3), Gumbel, generalized extreme value (GEV), generalized
logistic (GL) distributions, respectively.
Distribution CDF, equations for MOM, ML and log-likelihood function, respectively
ðx 
1=2
LN 1 ðlnxμY Þ2 =2σ 2Y σ 2X σ2
FX ð xÞ ¼ e dx; σ Y ¼ ln 1 þ ; μY ¼ lnμX  Y
xσ Y ð2πÞ 1=2 μX2 2
0
μY ¼ y; σ Y ¼ Sy
μY ¼ l1 ; σ Y ¼ π1=2 l2
X
n
1 X n
ln LðμY ; σ Y Þ ¼ n lnð2πÞ  n ln σ Y  ln xi  2
ðln xi  μY Þ2
i¼1
2σ Y i¼1
ðx

1 x  c α1 ðxcÞ=β
P3 FX ð xÞ ¼ e dx
βΓðαÞ β
c
4 σ x csx
α¼ 2 ;β¼ ; c ¼ μx  αβ
csx 2
1 þ 0:2906z
For: 0 < |t3| < 1/3: z ¼ 3πt32 ; α ¼ .
z þ 0:1882z2 þ 0:0442z3
0:36067z  0:59567z2 þ 0:25361z3
For: 1/3 < |t3| < 1: z ¼ 1  t3 ; α ¼
1  2:78861z þ 2:56096z2  0:77045z3
ΓðaÞ
For all t3values: β ¼ signðt3 Þπ1=2 l2 ; c ¼ l1  αβ
Γða þ 0:5Þ
X
n
1X n
ln Lðα; β; cÞ ¼ n lnðβÞ  n lnðΓðαÞÞ  nðα  1Þ lnðβÞ þ ðα  1Þ lnðxi  cÞ  ðxi  cÞ
ðy
 i¼1
β i¼1
1 y  c α1 ðycÞ=β
LP3 FY ð y Þ ¼ e dy; y ¼ log x
ΓðαÞ β
0
4 σ y csy
α¼ 2 ;β¼ ; c ¼ μy  αβ
csy 2
Same equations as for P3 distribution
X
n X
n
1X n
ln Lðα; β; cÞ ¼ nα lnðβÞ  n lnðΓðαÞÞ  yi þ ðα  1Þ lnðyi  cÞ  ðyi  cÞ
i¼1 i¼1
β i¼1
ðxuÞ=α
Gumbel FX ð xÞ ¼ ee
ðð61=2 Þ σ x Þ
α¼ ; u ¼ μx  0:5772α
π
l2
α¼ ; u ¼ l1  0:5772α
ln 2
1X n X
n
ln Lðu; αÞ ¼ n lnðαÞ  ðxi  uÞ  eðxi uÞ=α
α i¼1 ! i¼1
 
x   1=k
GEV FX ð xÞ ¼ exp  1  k
α

1=2
k Γð1 þ 3k Þ þ 3Γð1 þ k ÞΓð1 þ 2k Þ  2Γ3 ð1 þ k Þ k 2 σ 2X α
cs ¼  3=2 ;α¼ ;  ¼ μX  ½1  Γð1 þ k Þ
jk j Γð1 þ 2k Þ  Γ2 ð1 þ k Þ Γð1 þ 2k Þ  Γ2 ð1 þ k Þ k
2 ln2 kl2 αðΓð1 þ k Þ  1Þ
c¼  ; k ¼ 7:8590c þ 2:9554c2 ; α ¼ ;  ¼ l1 þ .
3 þ t3 ln3 Γð1 þ k Þ 1  2k k
X n Xn

 
1 xi  
ln Lðk; α; Þ ¼ n lnðαÞ  ð1  k Þ yi  eyi ; y ¼  ln 1  k
k α

i¼1
1
i¼1
k
GL FX ð xÞ ¼ 1 þ 1  ½x  1=k
α
α
k ¼ 3π
2
tan1 ð0:59484cs Þ; α ¼ signðk Þ kσ X
1=2 ;  ¼ μX  k ½1  Γð1 þ k ÞΓð1  k Þ
½Γð1þ2k ÞΓð12k ÞΓ2 ð1þk ÞΓ2 ð1k Þ
l2 l2  α
k ¼ t3 ; α ¼ ;  ¼ l1 þ
Γð1 þ k ÞΓð1 

k Þ X
k 
 !
1 n
xi   X n
xi   1=k
ln Lðk; α; Þ ¼ n lnðαÞ þ 1 ln 1  k 2 ln 1 þ 1  k
k i¼1
α i¼1
α
 Comparison between POT and AM methods 967

Table 2 Test statistics for the applied tests. Kendall test (Kendall 1975, Douglas et al. 2000)
Test Test statistic
was used to detect the presence of trends in selected
samples for the periods 1895–2010, 1895–1952 and
K-S D ¼ max Fx Qmax;i  Fi ð xÞ 1953–2010. The AM sample and three POT samples
Pn ð2i1ÞlnF Q
x ð max;i Þþlnf1Fx ðQmax;niþ1 Þg (POT 1, POT 3 and POT 5) were analysed. Samples
A-D A2 ¼ n  n
i¼1 Pn were defined based on daily discharge series with
ð x 
x Þð Q  max Þ
Q
PPCC ¼ Pn
i max;i
1=2 Pn included local maximums. Negative trends could be
PPCC i¼1
2 1=2
ð i¼1 ðxi xÞ2 Þ ð Q max;i Qmax Þ

 n
i¼1
1=2 observed for the samples of the whole discussed
P 2
RMSE RMSE ¼ 1n xi  Qmax;i period and for the first part of the period 1895–
i¼1
1952, whereas positive trend could be observed for
P
n
MAE MAE ¼ n 1
xi  Qmax;i the samples of the second part of the period (Fig. 7);
i¼1
in all cases the Mann-Kendall test results were not
RMAE RMAE ¼ MAE
 max
Q statistically significant.
AIC AIC ¼ nlog RMSE2 þ 2k The Von Neumann ratio test (Von Neumann
1941, Sahin and Cigizoglu 2010) and the standard
Notes: Fi, empirical distribution function; Fx, tested distribution function;
Qmax,i, sample discharge value; xi, calculated discharge value for the normal homogeneity test (Alexandersson 1986, Sahin
chosen distribution function; n, sample size; k, number of distribution and Cigizoglu 2010) were used to assess the homo-
parameters; K-S, Kolmogorov-Smirnov; A-D, Anderson-Darling; PPCC,
probability plot correlation coefficient; RMSE, root mean square error; geneity of data from the Litija hydrological station on
MAE, mean absolute error; RMAE, relative mean absolute error; AIC, the Sava River. AM series samples for the periods
Akaike information criterion.
1895–2010, 1895–1952 and 1953–2010 were used to
check if the data are homogeneous. Critical values
provided critical values for the A-D test where the provided by Sahin and Cigizoglu (2010) were used
parameters of the distribution are estimated from the for the Von Neumann test and critical levels calcu-
tested sample: Kottegoda and Rosso (2008) for nor- lated by Khaliq and Ouarda (2007) were applied with
mal, LN and exponential distributions, Zheng (2002) the standard normal homogeneity test. Both tests
for gamma distributions, Charles Annis has published showed that the analysed data are homogeneous for
critical values for Gumbel and Weibull distributions the whole of the discussed period. Because the ana-
based on D’Agostino and Stephens (1986) on his web lysed sample is homogenous and stationary, FFA can
page (http://www.statisticalengineering.com) and be performed.
Ahmad et al. (1988) for GEV and GL distributions.
All the critical values could also be obtained using
Monte Carlo simulations. POT model
All the test statistics used here are shown in
Table 2. This type of AIC test was used by The POT analyses were made for different threshold
Karmakar and Simonovic (2008, 2009). values. Many possible choices for threshold values
that were found in the literature were applied to Litija
1 data for the period 1953–2010. The RMSE, MAE,
RESULTS AND DISCUSSION RMAE and PPCC tests were used to find out which
of these possible truncation levels gave the best fit to
Data analyses
the POT sample. The results of the tests for different
The characteristics of the AM samples and some threshold values for partial duration series are shown
POT samples are shown in Table 3. The Mann- in Fig. 3.

Table 3 Characteristics of the AM and some POT samples in selected periods.

Characteristic 1895–2010 1895–1952 1953–2010

AM POT 5 AM POT 5 AM POT 1 POT 3 POT 5

Sample size 116 580 58 290 58 58 174 290

Mean (m3 s-1) 1234 847 1234 874 1234 1348 982 822
SD (m3 s-1) 397 317 407 314 391 296 321 317
Skewness 0.29 1.56 0.05 1.38 0.57 1.17 1.56 1.78
Kurtosis –0.35 2.53 –0.53 1.89 –0.07 1.19 2.39 3.44
968
Fig. 3 Test results of RMSE, MAE and PPCC for different threshold values for the POT method.
The POT sample that was composed of, on
average, eight peaks per year gave the best fit to
data. Only slightly poorer test statistics were calcu-
lated with the sample that was composed of, on
average, five peaks per year. The Poisson distribu-
tion was used for modelling the annual number of
exceedences. This distribution function could not be
rejected by the Chi-square test and the dispersion
index test (Cunnane 1979) at the significance level
0.05. Magnitudes of exceedences were modelled
with
pffiffiffi an exponential distribution. The G test and
r n test (Van Montfort and Witter 1985) could not
reject the tested distribution function at the 0.05
significance level. Different POT samples were also
checked with a simple graphical test, based on the
dispersion index, as proposed by Cunnane (1979) and
also used in other studies (Lang et al. 1999, Begueria
2005). A similar graphical plot was also used by
Ashkar and Rousselle (1987). The test statistic is
defined using the expression: d = var(m)/E(m).
Fig. 4 Dispersion index test for the original series and the series for which independence criteria were considered.
Nejc Bezak et al.
Ashkar and Rousselle (1987) suggested that the
threshold value should be chosen at the point
where the curve starts to stabilize around the line
d = 1. It can be seen from Fig. 4 that the POT
sample with an average of five peaks per year
(POT 5) gives satisfactory results. Because the
threshold selection is a subjective process, one
could easily select a different truncation level
(higher or even lower values). The POT 5 sample
was also suggested by the Flood Estimation
Handbook (Robson and Reed 1999). The hatched
line in Fig. 4 shows the results of the dispersion
index test when the independence criteria were not
considered. In this case, the variance of the annual
number of exceedences is larger than the mean
value. This means that the events occurred in
bunches, because daily-scale data sets that included
local maximums were considered in the analyses.
This graphical test defines the interval of possible
threshold values.
 Comparison between POT and AM methods
Fig. 5 Comparison of distributions for the annual number of exceedences for the threshold value of 508 m3 s-1.
Ashkar and Rousselle (1983a) found that if the
Poisson distribution is valid for some threshold value,
it should remain so for any higher threshold level. The
results of the dispersion index test for the discussed
data are in agreement with this conclusion (Fig. 4).
If, instead of the Poisson distribution, the nega-
tive binomial (or binomial) distribution was used for
modelling the annual number of exceedences, the
results of the RMSE, MAE, RMAE and PPCC tests
were generally not so good, as in the case of the
combination of the Poisson and exponential distribu-
tions. Figure 5 indicates that the Poisson and bino-
mial distributions gave similar results in the case of
the POT 5 sample. The dispersion index value for
this POT sample was slightly below 1; however, the
Poisson distribution was not rejected by the disper-
sion index test or by the Chi-square test with the
significance level of 0.05. But the results of RMSE,
MAE, RMAE and PPCC tests showed that the
Poisson distribution gave a better fit to the data
than the negative binomial distribution. The differ-
ence in the estimated discharge values is less than 1%
for smaller return periods and converges to 0% for
larger return periods. Cunnane (1979) and Önöz and
Bayazit (2001) also found that the negative binomial
(binomial) distribution does not offer any satisfactory
improvement of the Poisson distribution.
Annual maximum series model
The K-S, A-D, PPCC, RMSE, MAE, RMAE and AIC
tests were applied to the AM series data for the period
1953–2010. The results of the tests for the case when
the parameters were estimated with the MOM, the ML
and the MLE are shown in Table 4. When the para-
meters were estimated with the ML, the results of the
969
test statistics were generally better than those in the
case of the other two methods. The LP3 distribution
gave the best results in most of the applied tests. When
the parameters were estimated with the MOM, the LN
distribution came second and the GEV distribution
third. In the case of ML, the P3 and GEV distributions
came second and third, respectively. When the distri-
butions parameters were estimated using the MLE, the
best results were obtained using the P3 distribution
function. LN and LP3 distributions followed as the
second and the third, respectively. Poor test results
were obtained with the GL distribution. The results
of PPCC, K-S and A-D tests showed that the LP3
distribution could not be rejected at the 0.05 signifi-
cance level for any of the cases of parameter estima-
tion considered. All the tested distribution functions
could not be rejected with the chosen significance
level of 0.05. All critical values for the 0.05 signifi-
cance level are given in parentheses in Table 4.
The ML gave higher estimated discharge values
for larger return periods than the MOM and lower
estimates for smaller return periods. The difference
between the MOM and ML in the estimated discharge
value for the return period of 100 years was 25 m3 s-1
for the LP3 distribution function. In Fig. 6, the results
of AM analyses are shown for all considered distribu-
tion functions and all three parameter estimation tech-
niques. The Weibull plotting position formula was
used for graphical representation of the AM series
sample. In the case of the MLE, the curves that link
the estimated discharge value and the return periods
are more dispersed than when using the ML or MOM.
The L-moments ratio diagram was also used
for data from Litija 1 gauging station. This plot
confirmed the test results in Table 4. The best
results were obtained using LP3 and P3
970 Nejc Bezak et al.

Table 4 Tests results for the AM method for the period 1953–2010 and parameters estimated with the method of moments,
the method of L-moments and the maximum likelihood estimation.
Distribution/Test K-S A-D PPCC RMSE MAE RMAE AIC

Method of moments
LN 0.069 0.311 0.996 38.02 32.65 0.026 187.28
(0.116) (0.762) (0.98)
P3 0.066 0.275 0.996 41.81 33.87 0.027 194.07
(0.117) (0.752) (0.98)
LP3 0.069 0.289 0.997 38.09 32.38 0.026 189.37
(0.117) (0.752) (0.98)
Gumbel 0.088 0.531 0.995 49.96 43.05 0.035 201.04
(0.117) (0.777) (0.97)
GEV 0.068 0.277 0.996 40.98 33.44 0.027 193.05
(0.115) (0.56) (0.98)
GL 0.089 0.545 0.992 40.01 50.31 0.041 212.27
(0.102) (0.698) (0.97)
Method of L-moments
LN 0.062 0.277 0.996 34.87 29.60 0.024 182.92
(0.116) (0.762) (0.98)
P3 0.059 0.239 0.997 34.31 28.58 0.023 184.10
(0.117) (0.752) (0.98)
LP3 0.063 0.251 0.997 33.05 27.60 0.022 182.22
(0.117) (0.752) (0.98)
Gumbel 0.068 0.348 0.995 40.45 35.03 0.028 190.41
(0.117) (0.777) (0.97)
GEV 0.062 0.254 0.997 34.86 29.24 0.024 184.91
(0.115) (0.56) (0.98)
GL 0.076 0.389 0.994 46.42 38.20 0.031 199.34
(0.102) (0.698) (0.97)
Maximum likelihood estimation
LN 0.069 0.311 0.996 38.02 32.65 0.026 187.28
(0.116) (0.762) (0.98)
P3 0.063 0.288 0.996 36.47 31.56 0.026 187.19
(0.119) (0.752) (0.98)
LP3 0.069 0.285 0.997 38.61 32.67 0.026 190.06
(0.117) (0.752) (0.98)
Gumbel 0.068 0.335 0.995 39.46 34.27 0.028 189.17
(0.117) (0.777) (0.97)
GEV 0.071 0.303 0.997 41.55 34.63 0.028 193.76
(0.115) (0.56) (0.98)
GL 0.077 0.389 0.991 53.63 40.09 0.032 206.61
(0.102) (0.698) (0.94)
Note: See Table 1 for explanation of distributions.

distributions. This type of a graph is appropriate
for defining the best fit distribution function in the
case of data from more than one gauging station
(Peel et al. 2001). But it was also used in the Litija
1 analyses to see if the plot results are in agree-
ment with the test results represented in Table 4.
The results of the AM analyses in the case when
the parameters were estimated with the ML were
better than those obtained by using the MOM or the
MLE. In our study, the skewness of data from Litija 1
station is 0.57 (period 1953–2010) and the sample
size is relatively large. The ML gives better results
than the MOM. These results are in agreement with
the findings of Sankarasubramanian and Srinivasan
(1999). They found that the conventional moments
are preferable at lower skewness, particularly for
smaller samples. L-moments have the advantage at
higher skewness values for all sample sizes. Rosbjerg
et al. (1992) compared the MOM and the method of
PWM for estimation of partial duration series model
parameters. The MSE test was used and they found
that MOM gave better results than PWM. The gen-
eralized Pareto distribution was used for modelling
the magnitudes of exceedences.
Comparison between POT and AM series
The comparison between AM and POT series was
based on daily discharge values with included local
maximums for the period 1895–2010. Figure 7 shows
 Comparison between POT and AM methods 971

Fig. 6 Quantile functions of the AM series when distribution parameters are estimated with (a) the method of moments,
(b) the method of L-moments, and (c) the maximum likelihood estimation.

Fig. 7 Comparison between the AM sample and the POT sample with, on average, five peaks per year for different time
periods (left to right): 1895–2010, 1895–1952, and 1953–2010.

all events above the threshold for the POT sample majority of additional peaks lie just above the thresh-
with an average of five peaks per year, and the AM old. The POT 5 truncation level was chosen as opti-
series sample (for different time periods). The POT 5 mal for the data from gauging station Litija 1.
sample is five times larger than the AM sample. The Figure 8 shows histograms of both AM and POT 5
972 Nejc Bezak et al.
Fig. 8 Histograms of the AM series and the POT 5 data for different time periods (left to right): 1895–2010, 1895–1952
and 1953–2010.
samples for different sample lengths. It can be seen
that the AM series data are more symmetrical than
the POT sample and therefore the MOM might have
an advantage over the ML.
One of the main disadvantages of the POT
method is the selection of the truncation level.
Tavares and da Silva (1983) suggested selecting as
low a threshold value as possible. In the case of the
Litija 1 data, the POT 5 sample was selected as opti-
mal. A higher threshold level could be selected, but the
POT 5 sample gave good results in the applied statis-
tical and graphical tests. The sample with an average
of five peaks per year above threshold is also used in
the Flood Estimation Handbook (Robson and Reed
1999). The best statistical test results were obtained
with the POT 8 sample, but this sample gave worse
results in the graphical tests than the POT 5 sample.
From Fig. 4, the interval of possible threshold
choices can be defined. Tavares and da Silva (1983)
found that a POT sample with an average of two
peaks per year gave better results than the AM
method. Cunnane (1973) reported that an average of
at least 1.65 peaks per year should be selected in the
POT method to gain advantage over the AM method.
Madsen (1996) found that more than 0.91 peaks per
year should be selected for the POT model to have
advantage over the AM method when the parameters
are estimated with the MOM. For the observed data
from the Litija 1 gauging station (1953–2010), the
POT sample with an average of one peak per year
gave better results than the AM series with the use of
the LP3 distribution function. With lowering of the
threshold value, the POT results became even better
than the AM series results. Önöz and Bayazit (2001)
also found that the POT model was more efficient
than the AM series method.
The results of the AM and POT methods for the
cases that were selected as optimal are shown in
Fig. 9. For smaller return periods, the estimated dis-
charge values with AM series method (MOM, ML
and MLE) were slightly higher than in the case of the
POT 5 sample. For larger return periods, the POT
method gave higher estimated discharge values. The
difference between the AM method and the POT
method increases with increasing return period
values. Bačová-Mitková and Onderka (2010) found
that the POT method did not always give higher
estimated discharge values. Önöz and Bayazit
(2001) applied the POT method and the AM series
method to two data sets. The estimates of the T-year
event from the POT sample were smaller than those
obtained from the AM series. In our study, the POT
 Comparison between POT and AM methods 973

Fig. 9 Results of the AM and POT flood frequency analyses.

model gave higher estimated discharge values than
the AM model with the LP3 (MOM and ML) and P3
(MLE) distributions for larger return periods. The
difference becomes even more significant with the
samples for 1895–2010 and 1895–1952. In these
cases, the difference between the AM sample with
the use of LP3 distribution (ML) and the POT 5
model is larger than 500 m3 s-1, and some other
distributions gave higher estimated discharge values
than the LP3 distribution (Table 5).
The estimated discharge values with return per-
iods of 10 and 100 years for the AM and POT series
are shown in Tables 5 and 6 (for different time
periods). The ML gave higher estimated discharge
values than the MOM for all six probability distribu-
tions considered (sample 1953–2010). The difference
between MOM and MLE, and MLE and the ML, is
not so unambiguous (sample 1953–2010). Lowering
of the threshold value does not mean higher esti-
mated discharge values (Table 6). The dispersion

Table 5 Estimated discharge values (m3 s-1) for the AM series for all three parameter estimation techniques for different
time periods.
Distribution 1895–2010 1895–1952 1953–2010
LM LM MOM L-moments MLE

Q10 Q100 Q10 Q100 Q10 Q100 Q10 Q100 Q10 Q100

LN 1815 2615 1857 2727 1764 2458 1784 2509 1764 2458
Gumbel 1780 2555 1799 2592 1741 2453 1770 2523 1776 2534
P3 1763 2286 1767 2215 1750 2300 1768 2368 1773 2438
LP3 1774 2261 1781 2129 1758 2398 1775 2423 1755 2378
GEV 1770 2260 1778 2167 1754 2304 1768 2376 1748 2346
GL 1739 2411 1750 2327 1709 2352 1736 2513 1783 2815

Table 6 Estimated discharge values (m3 s-1) at different threshold values (m3 s-1) for the POT series for different time
periods.
1895–2010 1895–1952 1953–2010

Threshold Q10 Q100 Threshold Q10 Q100 Threshold Q10 Q100

1060 (POT 1) 1748 2467 1088 (POT 1) 1743 2426 1021 (POT 1) 1758 2527
683 (POT 3) 1800 2583 707 (POT 3) 1852 2655 785 (POT 2) 1755 2530
512 (POT 5) 1807 2597 529 (POT 5) 1856 2665 662 (POT 3) 1734 2486
629 (x0 = μx + kσx) 1708 2435
579 (POT 4) 1706 2431
508 (POT 5) 1723 2463
442 (POT 6) 1738 2491
392 (POT 7) 1744 2501
347 (POT 8) 1746 2504
974 Nejc Bezak et al.
(range) of estimated discharge values for the POT
model is smaller than in the AM method (Tables 5
and 6). However, the differences between the two
methods are not very large, especially if we look at
them in the light of accuracy of the data (measuring
errors, rating curve extrapolation), which can lead to
errors of up to 10%, or even 15%. The ML results
were in a narrower range than results obtained with
the MOM and MLE. We can conclude that, in our
case, the results depend mostly on the distribution
selected, especially for high return periods. Similar
conclusions can be made for the other two samples
(1895–2010 and 1895–1952), where only the ML
was used and three threshold values were considered
(POT). For these two samples, the differences
between the POT and AM methods are a bit larger,
and the LN distribution gave the highest estimated
discharge values.
CONCLUSIONS
This article presents analyses of high flows in rela-
tion to the data from the Litija 1 gauging station on
the Sava River for the period 1895–2010. Flood
frequency analyses were carried out with the AM
and partial duration series (POT).
One of the main disadvantages concerning the
POT method is the selection of the threshold value.
Several different truncation levels were applied to the
Litija 1 data (sample 1953–2010). Tests were applied
to find which threshold value gave the best fit to data.
The POT 5 sample was selected as optimal. This
truncation level is also used in the Flood Estimation
Handbook (Robson and Reed 1999). Since the
threshold selection is a subjective process, one
could easily choose any higher or even lower thresh-
old value, but the chosen base level (POT 5) pro-
vided enough advantage regarding sample size, and
the results of statistical tests were considerably better
than in the case of the AM series. The POT model
used the exponential and Pareto distributions for
modelling the magnitudes of exceedences, and the
Poisson, binomial and negative binomial distribu-
tions for the annual number of events above the
threshold. Independence criteria that were suggested
by USWRC (1982) were applied to the Litija 1 data.
Four tests were used to check the adequacy of the
Poisson and exponential distributions. For modelling
the annual number of exceedences above the thresh-
old, the Poisson distribution gave better results than
the binomial distribution. The estimated discharge
values were almost identical using combinations of
the Poisson and exponential distributions, and the
binomial and exponential distributions. Similar con-
clusion were drawn by Cunnane (1979) and Önöz
and Bayazit (2001).
Different parameter estimation techniques were
used and some frequently used probability distribu-
tion functions were applied to the AM data. The LP3
distribution gave the best fit to the AM data (sample
1953–2010) using the MOM and the ML. For the
MLE, the best results were obtained with the P3
distribution function. The GL distribution gave a
poorer fit to the AM sample. The LP3 distribution
could not be rejected by the K-S, A-D and PPCC
tests with a significance level of 0.05. The three-
parameter log-Pearson distribution is often used in
FFA and was also suggested by the Water Resources
Council (USWRC 1982). The ML generally showed
better results than the MOM or the MLE (sam-
ple 1953–2010). This conclusion is in agreement
with the findings by other researchers
(Sankarasubramanian and Srinivasan 1999, Hosking
and Wallis 2005). The difference between the MOM
and MLE is not so unambiguous; the MLE gave
better results than the MOM in some distribution
functions (P3, Gumbel) and, in the case of the
Gumbel distribution function, even better results
than the ML.
The POT method gave better results than the AM
series method. Cunnane (1973), Tavares and da Silva
(1983), Madsen (1996) and also Bhunya et al. (2012)
found that the POT method with an average of two
events per years had an advantage over the AM
method. The same can be said for any lower threshold
value. In the Litija 1 analyses (sample 1953–2010),
even the POT sample with an average of one peak
per year above the threshold gave better results than
the best of the AM combinations (LP3 and ML). For
lower truncation levels, the test results were even bet-
ter. Tavares and da Silva (1983) state that the threshold
value should be selected as low as possible; however,
the model assumptions must remain valid. The POT 5
sample was composed of the Poisson distribution for
modelling the annual number of exceedences and the
exponential distribution for modelling the magnitudes
of events above the threshold. Both distributions were
tested and could not be rejected with a significance
level of 0.05. The POT method gave higher estimated
discharge values than the AM series method (for all
three periods: 1953–2010; 1895–2010; 1895–1952);
however, some other research has found the opposite
(Önöz and Bayazit 2001, Bačová-Mitková and
Onderka 2010). Some research states that the main
 Comparison between POT and AM methods
advantage of the POT method is for smaller sample
sizes (Robson and Reed 1999, Bačová-Mitková and
Onderka 2010), but the results of our analyses show
that the POT method can also be used when longer
samples are analysed. The POT method is a little more
complex than the AM series method, but the results of
statistical and graphical tests show that the POT
method gives better fits to data.
The POT method is a good alternative and
improvement on the AM series method and can also
be used in the case of longer data series. If one wants
to make a quick estimate of design discharge, we
suggest using the ML as the parameter estimation
technique and three or four different distribution
functions (e.g. LP3, GEV, P3 and GL), but if more
serious work is to be done then more detailed ana-
lyses should be carried out, including partial duration
series analyses for a few different thresholds values.
In both cases some statistical or (and) graphical tests
should be added to analyses. For further conclusions
about the adequacy of the separate aspects of FFA,
similar analyses should also be performed for other
gauging stations in Slovenia.
Acknowledgements We wish to thank the
Environmental Agency of the Republic of Slovenia
for data provision. The critical and useful comments
of two anonymous reviewers improved this work, for
which the authors are very grateful.
Funding The results of the study are part of the
Slovenian national research project J2-4096. The
study was partially supported by the Slovenian
Research Agency (ARRS).
REFERENCES
Adamson, P.T. and Zucchini, W., 1984. On the application of a
censored log-normal distribution to partial duration series of
storms. Water SA, 10 (3), 136–146.
Ahilan, S., O’Sullivan, J.J., and Bruen, M., 2012. Influences on flood
frequency distributions in Irish river catchments. Hydrology
and Earth System Sciences, 16 (4), 1137–1150. doi:10.5194/
hess-16-1137-2012.
Ahmad, M.I., Sinclair, C.D., and Spurr, B.D., 1988. Assessment of
flood frequency models using empirical distribution function
statistics. Water Resources Research, 24 (8), 1323–1328.
doi:10.1029/WR024i008p01323.
Alexandersson, H., 1986. A homogeneity test applied to precipitation
data. Journal of Climatology, 6, 661–675. doi:10.1002/
joc.3370060607.
Ashkar, F. and Rousselle, J., 1983a. Some remarks on the truncation
used in partial flood series models. Water Resources Research,
19 (2), 477–480. doi:10.1029/WR019i002p00477.
Ashkar, F. and Rousselle, J., 1983b. The effect of certain restrictions
imposed on the interarrival times of flood events on the
975
Poisson distribution used for modeling flood counts. Water
Resources Research, 19 (2), 481–485. doi:10.1029/
WR019i002p00481.
Ashkar, F. and Rousselle, J., 1987. Partial duration series modeling
under the assumption of a Poissonian flood count. Journal of
Hydrology, 90 (1–2), 135–144.
Bačová-Mitková, V. and Onderka, M., 2010. Analysis of extreme
hydrological events on the Danube using the peak over thresh-
old method. Journal of Hydrology and Hydromechanics, 58
(2), 88–101. doi:10.2478/v10098-010-0009-x.
Begueria, S., 2005. Uncertainties in partial duration series modelling
of extremes related to the choice of the threshold value.
Journal of Hydrology, 303 (1–4), 215–230.
Bhunya, P.K., et al., 2012. Flood analysis using generalized logistic
models in partial duration series. Journal of Hydrology, 420–
421, 59–71. doi:10.1016/j.jhydrol.2011.11.037.
Bonacci, O. and Ljubenkov, I., 2008. Changes in flow conveyance
and implication for flood protection, Sava River, Zagreb.
Hydrological Processes, 22 (8), 1189–1196. doi:10.1002/
hyp.6688.
Brilly, M. and Polic, M., 2005. Public perception of flood risks,
flood forecasting and mitigation. Natural Hazards and Earth
System Sciences, 5 (3), 345–355. doi:10.5194/nhess-5-345-
2005.
Chandra, M., et al., 1981. Kolmogorov statistics for tests of fit for
the extreme-value and Weibull distributions. Journal of the
American Statistical Association, 76 (375), 729–731.
Chowdhury, J.U., Stedinger, J.R., and Lu, L., 1991. Goodness-of-fit
tests for regional generalized extreme value flood distributions.
Water Resources Research, 27 (7), 1765–1776. doi:10.1029/
91WR00077.
Crutcher, H.L., 1975. A note on the possible misuse of the
Kolmogorov-Smirnov test. Journal of Applied Meteorology,
14 (8), 1600–1603. doi:10.1175/1520-0450(1975)014<1600:
ANOTPM>2.0.CO;2.
Cunnane, C., 1973. A particular comparison of annual maxima and
partial duration series methods of flood frequency prediction.
Journal of Hydrology, 18 (3–4), 257–271.
Cunnane, C., 1979. A note on the Poisson assumption in partial
duration series models. Water Resources Research, 15 (2),
489–494. doi:10.1029/WR015i002p00489.
D’Agostino, R.B. and Stephens, M.A., 1986. Goodness-of-fit techni-
ques. New York: M. Dekker.
Douglas, E.M., et al., 2000. Trends in floods and low flows in the
United States: impact of spatial correlation. Journal of
Hydrology, 240 (1–2), 90–105.
Frantar, P. and Hrvatin, M., 2005. Pretocni rezimi v Sloveniji med
letoma 1971 in 2000. Geografski vestnik, 77, 115–127.
Frantar, P. and Hrvatin, M., 2008. Discharge regimes. In: P. Frantar,
ed. Water balance of Slovenia 1971–2000. Ljubljana: MOP
ARSO, 43–50.
Haan, C.T., 2002. Statistical methods in hydrology. Ames, IA: Iowa
State Press.
Haktanir, T., 1991. Practical computation of gamma frequency fac-
tors. Hydrological Sciences Journal-Journal Des Sciences
Hydrologiques, 36 (6), 599–610. doi:10.1080/
02626669109492546.
Haktanir, T. and Horlacher, H.B., 1993. Evaluation of various dis-
tributions for flood frequency-analysis. Hydrological Sciences
Journal, 38 (1), 15–32. doi:10.1080/02626669309492637.
Hosking, J.R.M., 1990. L-moments: analysis and estimation of dis-
tributions using linear combinations of order statistic. Journal
of the Royal Statistical Society Series B-Methodological, 52
(1), 105–124.
Hosking, J.R.M. and Wallis, J.R., 2005. Regional frequency analysis:
an approach based on L-moments. Cambridge University
Press.
976 Nejc Bezak et al.
Irvine, K.N. and Waylen, P.R., 1986. Partial series analysis of high
flows in Canadian rivers. Canadian Water Resources Journal,
11 (2), 83–91. doi:10.4296/cwrj1102083.
Karim, M.A. and Chowdhury, J.H., 1995. A comparison of four
distributions used in flood frequency analysis in Bangladesh.
Hydrological Sciences Journal, 40 (1), 55–66. doi:10.1080/
02626669509491390.
Karmakar, S. and Simonovic, S.P., 2008. Bivariate flood frequency
analysis: part 1. Determination of marginals by parametric and
nonparametric techniques. Journal of Flood Risk Management,
1 (4), 190–200. doi:10.1111/j.1753-318X.2008.00022.x.
Karmakar, S. and Simonovic, S.P., 2009. Bivariate flood frequency
analysis. Part 2: a copula-based approach with mixed marginal
distributions. Journal of Flood Risk Management, 2 (1), 32–44.
doi:10.1111/j.1753-318X.2009.01020.x.
Kendall, M.G., 1975. Multivariate analysis. London: Griffin.
Khaliq, M.N. and Ouarda, T.B.M.J., 2007. On the critical values of
the standard normal homogeneity test (SNHT). International
Journal of Climatology, 27 (5), 681–687. doi:10.1002/
joc.1438.
Khaliq, M.N., et al., 2006. Frequency analysis of a sequence of
dependent and/or non-stationary hydro-meteorological obser-
vations: a review. Journal of Hydrology, 329 (3–4), 534–552.
Kim, S., et al., 2008. Derivation of the probability plot correlation
coefficient test statistics for the generalized logistic and the
generalized Pareto distributions. Proceedings of the World
Environmental and Water Resources, Congress, 12–16 May.
Ahupua’a: American Society of Civil Engineers.
Kjeldsen, T.R., Lundorf, A., and Rosbjerg, D., 2000. Use of a two-
component exponential distribution in partial duration model-
ling of hydrological droughts in Zimbabwean rivers.
Hydrological Sciences Journal, 45 (2), 285–298. doi:10.1080/
02626660009492325.
Kobold, M., 2011. Comparison of floods in September 2010 with
registered historic flood events. Ujma, 25, 48–56 (in Slovenian).
Kottegoda, N.T. and Rosso, R., 2008. Applied statistics for civil and
environmental engineers. Oxford: Blackwell.
Lang, M., et al., 1997. Échantillonnage par valeurs supérieures à un
seuil : modélisation des occurrences par la méthode du renou-
vellement. Revue des Sciences de l’Eau, 10, 279–320.
doi:10.7202/705281ar.
Lang, M., et al., 1999. Towards operational guidelines for over-thresh-
old modeling. Journal of Hydrology, 225 (3–4), 103–117.
Langbein, W.B., 1949. Annual floods and the partial-duration flood
series. Transactions, American Geophysical Union, 30 (6),
879–881. doi:10.1029/TR030i006p00879.
Lilliefors, H.W., 1967. On the Kolmogorov-Smirnov test for normal-
ity with mean and variance unknown. Journal of the American
Statistical Association, 62 (318), 399–402. doi:10.1080/
01621459.1967.10482916.
Madsen, H., 1996. At-site and regional modelling of extreme hydrologic
events. Thesis (PhD). Technical University of Denmark, Lyngby.
Madsen, H., Rasmussen, P.F., and Rosbjerg, D., 1997. Comparison
of annual maximum series and partial duration series methods
for modeling extreme hydrologic events. 1: at-site modeling.
Water Resources Research, 33 (4), 747–757. doi:10.1029/
96WR03848.
Madsen, H., et al., 1993. Application of the partial duration series
approach in the analysis of extreme rainfalls. In: Z.W.
Kundzewicz, et al., eds. Extreme hydrological events: precipi-
tation, floods and droughts. Wallingford: IAHS Press, IAHS
Publ. 213, 257–366.
Mikoš, M., et al., 2004. Floods and landslides in Slovenia. Acta
Hydrotechnica, 22 (37), 113–133.
Önöz, B. and Bayazit, M., 1995. Best-fit distributions of largest
available flood samples. Journal of Hydrology, 167 (1–4),
195–208.
Önöz, B. and Bayazit, M., 2001. Effect of the occurrence process of
the peaks over threshold on the flood estimates. Journal of
Hydrology, 244 (1–2), 86–96.
Peel, M.C., et al., 2001. The utility of L-moment ratio diagrams for
selecting a regional probability distribution. Hydrological
Sciences Journal, 46 (1), 147–155. doi:10.1080/
02626660109492806.
Pekarova, P., et al., 2008. Flood regime of rivers in the Danube river
basin: a case study of the Danube at Bratislava [online].
National report for the IHP UNESCO. Institute of Hydrology
SAS, Bratislava, Slovakia. Available from: http://
147.213.145.2/DanubeFlood/PDF/Report01_Flood.pdf
[Accessed 10 December 2012].
Robson, A.J. and Reed, D.W., 1999. Statistical procedures for flood
frequency estimation. Volume 3 of the Flood Estimation
Handbook. Wallingford: Centre for Ecology & Hydrology.
Rosbjerg, D., 1987. On the annual maximum distribution in depen-
dent partial duration series. Stochastic Hydrology and
Hydraulics, 1 (1), 3–16. doi:10.1007/BF01543906.
Rosbjerg, D., Madsen H., and Rasmussen P.F., 1992. Prediction in
partial duration series with generalized Pareto-distributed
exceedances. Water Resources Research, 28 (11), 3001–3010.
doi:10.1029/92WR01750.
Sahin, S. and Cigizoglu, H.K., 2010. Homogeneity analysis of
Turkish meteorological data set. Hydrological Processes, 24
(8), 981–992. doi:10.1002/hyp.7534.
Salas, J.D., et al., 2012. Quantifying the uncertainty of return period
and risk in hydrologic design. Journal of Hydrologic
Engineering. doi:10.1061/(ASCE)HE.1943–5584.0000613.
Sankarasubramanian, A. and Srinivasan, K., 1999. Investigation and
comparison of sampling properties of L-moments and conven-
tional moments. Journal of Hydrology, 218 (1–2), 13–34.
Seckin, N., Haktanir, T., and Yurtal, R., 2011. Flood frequency
analysis of Turkey using L-moments method. Hydrological
Processes, 25 (22), 3499–3505. doi:10.1002/hyp.8077.
Shane, R.M. and Lynn, W.R., 1964. Mathematical model for flood
risk evaluation. Journal of the Hydraulics Division.
Proceedings of the American Society of Civil Engineers, 90
(HY6), 1–20.
Silva, A.T., Portela, M.M., and Naghettini, M., 2012.
Nonstationarities in the occurrence rates of flood events in
Portuguese watersheds. Hydrology and Earth System
Sciences, 16 (1), 241–254. doi:10.5194/hess-16-241-2012.
Šraj, M., 2001. Watershed coding system of the republic of Slovenia.
Acta Hydrotechnica, 19 (30), 3–24.
Šraj, M., et al., 2008. The experimental watersheds in Slovenia. IOP
Conference Series, 4, 1–13.
Stephens, M.A., 1979. Tests of fit for the logistic distribution based
on the empirical distribution function. Biometrika, 66 (3), 591–
595. doi:10.1093/biomet/66.3.591.
Tavares, L.V. and Da Silva, J.E., 1983. Partial duration series method
revisited. Journal of Hydrology, 64 (1–4), 1–14.
Todorovic, P. and Zelenhasic, E., 1970. A stochastic model for flood
analysis. Water Resources Research, 6 (6), 1641–1648.
doi:10.1029/WR006i006p01641.
USWRC (United States. Interagency Advisory Committee on Water
Data. Hydrology Subcommittee), 1982. Guidelines for deter-
mining flood flow frequency. Reston, VA: US Department of
the Interior, Geological Survey, Office of Water Data
Coordination.
Van Montfort, M.A.J. and Witter, J.V., 1985. Testing exponentiality
against generalised Pareto distribution. Journal of Hydrology,
78 (3–4), 305–315.
Vogel, R.M., 1986. The probability plot correlation coefficient test
for the normal, lognormal, and Gumbel distributional hypoth-
eses. Water Resources Research, 22 (4), 587–590. doi:10.1029/
WR022i004p00587.
 Comparison between POT and AM methods 977

Vogel, R.W. and McMartin, D.E., 1991. Probability plot goodness-
of-fit and skewness estimation procedures for the Pearson type
3 distribution. Water Resources Research, 27 (12), 3149–3158.
doi:10.1029/91WR02116.
Von Neumann, J., 1941. Distribution of the ratio of the mean square
successive difference to the variance. The Annals of Mathematical
Statistics, 12, 367–395. doi:10.1214/aoms/1177731677.
WMO (World Meteorological Organization), 1989. Statistical distri-
butions for flood frequency analysis. Geneva: World
Meteorological Organization.
Zheng, J., 2002. Quantification of variability and uncertainty in
emission estimation: general methodology and software imple-
mentation. Unpublished dissertation. North Carolina State
University.