You are on page 1of 10

494 JOURNAL OF APPLIED METEOROLOGY AND CLIMATOLOGY VOLUME 46

Maximum Monthly Rainfall Analysis Using L-Moments for an Arid Region in Isfahan
Province, Iran
S. SAEID ESLAMIAN*
Department of Civil and Environmental Engineering, Princeton University, Princeton, New Jersey

HUSSEIN FEIZI
Isfahan University of Technology, Isfahan, Iran

(Manuscript received 26 September 2004, in final form 11 July 2006)

ABSTRACT

Developing methods that can give a suitable prediction of hydrologic events is always interesting for both
hydrologists and statisticians, because of its importance in designing hydraulic structures and water resource
management. Because of the computer revolution in statistical computation and lack of robustness in at-site
frequency analysis, since early 1990 the application of regional frequency analysis based on L-moments has
been considered more for flood analysis. In this study, the above-mentioned method has been used for the
selection of parent distributions to fit maximum monthly rainfall data of 18 sites in the Zayandehrood basin,
Iran, and as a consequence the generalized extreme-value and Pearson type-III distributions have been
selected and model parameters have been estimated. The obtained extreme rainfall values can be used for
meteorological drought management in the arid zone.

1. Introduction and its arbitrary rules; problems with small samples;


and, last, the lack of power inherent in goodness-of-fit
An essential step in the design of any containment
tests (Wallis 1988).
structure and in water resource management is estimat-
Another problem with the at-site procedures that is
ing the probabilities of occurrence for the events that
often unclear, in particular at sites with short records, is
the structure is designed to alleviate or that will play an
that some of the sample statistics are bounded by in-
important role in our decisions. The statistical method
fection of sample sizes (Dalen 1987). It is the result of
for making such probability predictions has been revo-
such estimator limitations as skew and kurtosis.
lutionized by computers, and it seems that the current
In fact, the purpose of regional analysis based on
method based on regional analysis is much better than
L-moments is to improve the results and predictions of
the conventional at-site procedure.
frequency analysis by improving the goodness-of-fit
Many problems exist with the at-site procedure, and
tests and the estimates of the distribution parameters
the cumulative effect of these problems is that this ap-
used in single-site analysis. This is in contrast to a tra-
proach can no longer be condoned for real applications.
ditional at-site frequency analysis using the method of
These problems include the following: violation of the
moments to find the data distribution and to estimate
assumption of independence between the event values;
parameters. This study applies a regional frequency
the quantile function depends upon the method of fit,
analysis approach based on L-moments for the first
time on Iranian rainfall data. In a regional frequency
* Current affiliation: Department of Water Engineering, analysis, regional information is used to increase the
Ishafan University of Technology, Isfahan, Iran. reliability of rainfall intensity–duration–frequency esti-
mates at any particular site (Hosking and Wallis 1991).
One limitation is that the procedure assumes that sites
Corresponding author address: S. Saeid Eslamian, Associate
Professor, Department of Water Engineering, Isfahan University from a homogeneous region have the same frequency
of Technology, Isfahan, Iran. distribution apart from a site-specific scaling factor.
E-mail: saeid@cc.iut.ac.ir The L-moments are defined as expectations of cer-

DOI: 10.1175/JAM2465.1

© 2007 American Meteorological Society

JAM2465
APRIL 2007 ESLAMIAN AND FEIZI 495

TABLE 1. Main characteristics of rainfall-gauging sites under investigation and data series analyzed.

Data First Fifth


Alt Lon Lat Record collection L-moment L-C␷ L-Cs L-Ck L-moment
No. Site name (m) (E) (N) length (yr) period (mm) (␶) (␶3) (␶4) ratio (␶5)
1 Isfahan 1590 51°40⬘ 32°37⬘ 49 1953–2001 39.84 0.228 0.072 0.056 0.051
2 Pol-Kalleh 1720 51°14⬘ 32°23⬘ 45 1957–2001 50.04 0.231 0.0196 0.051 ⫺0.011
3 Pole-Mazraeh 1650 51°28⬘ 32°22⬘ 34 1968–2001 40.53 0.218 0.171 0.127 0.042
4 Garmaseh 1610 51°31⬘ 32°32⬘ 34 1968–2001 47.32 0.208 0.178 0.138 0.100
5 Khajoo 1585 51°41⬘ 32°37⬘ 34 1968–2001 36.70 0.248 0.252 0.160 0.070
6 Ziar 1530 51°56⬘ 32°32⬘ 31 1971–2001 33.36 0.229 0.230 0.222 0.087
7 Damaneh 2300 50°29⬘ 33°01⬘ 34 1968–2001 91.11 0.216 0.121 0.122 0.027
8 Eskandari 2130 50°25⬘ 32°48⬘ 26 1976–2001 106.58 0.178 0.082 0.155 0.064
9 Paiabe-Sad 1960 50°47⬘ 32°43⬘ 32 1970–2001 72.96 0.213 0.159 0.289 0.162
10 Pole-Zamankhan 1860 50°54⬘ 32°29⬘ 33 1969–2001 102.74 0.189 0.148 0.229 0.736
11 Tiran 1840 51°09⬘ 32°42⬘ 33 1969–2001 50.82 0.201 0.105 0.186 0.082
12 Khomeini-Shahr 1600 51°32⬘ 32°41⬘ 12 1990–2001 46.11 0.246 0.254 0.228 0.524
13 Meimeh 2000 51°11⬘ 33°26⬘ 26 1976–2001 30.88 0.220 0.100 0.113 0.086
14 Shahrokh 2080 50°07⬘ 32°40⬘ 27 1975–2001 112.56 0.211 0.158 0.112 0.022
15 Koohpaieh 1800 52°26⬘ 32°43⬘ 27 1975–2001 32.29 0.194 0.060 0.128 0.013
16 Firoozabad 2250 50°56⬘ 31°35⬘ 26 1976–2001 64.50 0.213 0.135 0.187 0.117
17 Maqsoodbeig 1980 51°59⬘ 31°50⬘ 26 1976–2001 36.50 0.232 0.199 0.245 0.125
18 Mahiar 1650 51°48⬘ 32°16⬘ 26 1976–2001 46.23 0.274 ⫺0.004 0.141 ⫺0.019
Regional average of L-moments ratio 0.2192 0.1454 0.1522 —

tain linear combinations of order statistics (Hosking logic events, such as rainfall, resulting from a lack of
1990). They are analogous to conventional moments historical data and old gauging stations.
with measures of location (mean), scale (standard de- The procedure of this study is based on the use of
viation), and shape (skewness and kurtosis). Applica- L-moments in a regional frequency analysis for estima-
tion of these moments in hydrology and frequency tion of maximum monthly rainfalls in Zayandehrood
analysis of hydrologic events began in the late 1970s basin. The data used have been collected from 18 rain-
(Greenwood and et al. 1979). Because L-moments are gauging stations, the records of which are published by
linear combinations of ranked observations and do not the Iranian National Meteorology Department. Alti-
involve squaring or cubing the observations, as is done tude and latitude have been assumed as initial statistics
for the conventional method of moments estimators, of hydrologic homogeneity, and station selection crite-
they are generally more robust and less sensitive to ria were based on these characteristics. All sites used in
outliers. this procedure are located between 31°50⬘ and 33°25⬘N
Zayandehrood basin is one of the large catchments in latitudes, and 1530 and 2300 m altitudes (MSL).
the central region of Iran, with the Qom desert in the Records used for the analysis have ended the same
north, the Zagros Mountain Range in the west and the year, and there are no gaps in the records (Table 1). All
Iran central desert in the east and south. The Zagros of the data were tested using the method of Wald and
Mountain Range prevents humid air masses from mov- Wolfowitz (1943) for possible correlation. There was
ing eastward to reach this region; therefore, this basin not any at-site and regional correlations between data.
receives no precipitation for at least 9 months each
year. Although it is known as an arid basin, the exis- 2. Method
tence of the Zayandehrood River and coastal farmlands
a. L-moments: Definition
in this basin cause this arid region to be one of the
important agricultural regions of central Iran. Because The L-moments are the summary statistics for prob-
rainfall occurrence and its quantity during specific du- ability distributions and data samples and are analo-
rations, such as daily, monthly, and seasonal depths, gous to ordinary moments (Hosking 1990). They pro-
play an important role in water resources planning and vide measures of location, dispersion, skewness, kurto-
crop water management, estimation of these values has sis, and other aspects of the shape of probability
been of particular interest for engineers in this region. distributions or data samples. Using the uniform distri-
They have faced the major problem of small data bution function as its foundation and based on shifted
samples derived from short record lengths of hydro- Legendre polynomials, each statistical L-moment is
496 JOURNAL OF APPLIED METEOROLOGY AND CLIMATOLOGY VOLUME 46

computed linearly (hence the L) giving a more robust Hosking and Wallis (1991) provide two statistics for a
estimate for a given amount of data than other meth- test of homogeneity. The first is a measure of dissimi-
ods. larity. This statistic can determine the sites that are
For the random variables X1, . . . , Xn of sample size disharmonious from a group of other sites. Estimates
n drawn from the distribution of a random variable X for this measure determine the status of distance of a
with the mean m and variance of s2, let X1:1 ⱕ · · · ⱕ site from the center of a group. If the vector ui ⫽ [t (i) t (i)
3
X1:n be the order statistics such that the L-moments of t (i) T
4 ] includes values of sample L-moments ratios t, t3,
X are defined by and t4 for site (i) (the superscript T denotes the trans-

兺 共⫺1兲 冉 冊
position of a vector or matrix), the average of the group
r⫺1
r⫺1 for Ns sites can be written as
␭r ⬅ r⫺1 k
E关Xr⫺k:r兴, r ⫽ 1, 2, . . . ,
k⫽0 k
Ns

兺u,
1
共1兲 u⫽ i 共8兲
Ns i⫽1
where r is the rth L-moment of a distribution and
E[Xi:r] is the expected value of the ith smallest obser- and the matrix of covariance of sample is
vation in a sample of size r. Ns
The first four L-moments of a random variable X can
be written as
S ⫽ 共Ns ⫺ 1兲⫺1 兺 共u ⫺ u兲共u ⫺ u兲 ;
i⫽1
i i
T
共9兲

␭1 ⫽ E 关X兴, 共2兲 then the measure of dissimilarity can be written as


␭2 ⫽ 共1Ⲑ2兲E 关X2:2 ⫺ X1:2兴, 共3兲
Di ⫽ 共1Ⲑ3兲共ui ⫺ u兲TS⫺1共ui ⫺ u兲. 共10兲
␭3 ⫽ 共1Ⲑ3兲E 关X3:3 ⫺ 2X2:3 ⫹ X1:3兴, and 共4兲
If Di ⱖ 3, the site i is disharmonious.
␭4 ⫽ 共1Ⲑ4兲E 关X4:4 ⫺ 3X3:4 ⫹ 3X2:4 ⫺ X1:4兴. 共5兲 The second statistic is a measure of heterogeneity.
This statistic estimates the degree of heterogeneity in a
Hosking (1990) demonstrated the utility of estima-
group of sites and evaluates whether these sites are
tors based on the L-moment ratios in hydrological ex-
reasonably homogeneous. In fact, this statistic com-
treme analysis. The second moment is often scaled by
pares the variations of L-moments of sites for a group
the mean so that a coefficient of variability is deter-
of these sites with the expected values of variations for
mined:
a homogeneous region.
␶ ⫽ L-C␷ ⫽ ␭2Ⲑ␭1, 共6兲 Three measures of variability, V1, V2 and V3, are
available. The weighted standard deviation of t, based
where ␭1 is the measure of location. Similar to the defi-
on L-C␷(t) is
nitions and the meaning of the ratios between ordinary
moments, the coefficients of L-kurtosis and L-skewness Ns Ns

are defined as V1 ⫽ 兺
i⫽1
Ni关t共i兲 ⫺ t兴2Ⲑ 兺N,
i⫽1
i 共11兲
␶r ⫽ ␭r Ⲑ␭2, r ⱖ 3, 共7兲
where Ns is number of sites, Ni is length of records at
where ␶3 is the measure of skewness (L-Cs) and ␶4 is the each site, and t is average of t(i) values computed by
measure of kurtosis (L-Ck). Unlike standard moments,
␶3 and ␶4 are constrained to be between ⫺1 and ⫹1 and
␶4 is constrained by ␶3 to be no lower than ⫺0.25. Be-
cause precipitation is nonnegative, ␶ is also constrained
t⫽ 冋兺 册冒冋兺 册
Ns

i⫽1
Nit共i兲
Ns

i⫽1
Ni . 共12兲

to the range from 0 to 1. The weighted average distance between the site and
weighted average of the group based on L-C␷ and L-Cs
b. L-moment: Applications is computed by
1) HOMOGENEITY TEST Ns Ns

In a regional frequency analysis, it is necessary that V2 ⫽ 兺


i⫽1
Ni 兵 关t 共i兲 ⫺ t兴2 ⫹ 关t 共3i兲 ⫺ t3兴2其1Ⲑ2Ⲑ 兺N.
i⫽1
i 共13兲
all of the sites located in one region have data with the
same probability distribution. For this reason, the ho- Weighted average distance between the site and weighted
mogeneity of the region should be tested before per- average of the group in a diagram of t3 versus t4, based
forming the regional analysis. on L-skewness and L-kurtosis, can be written as
APRIL 2007 ESLAMIAN AND FEIZI 497

TABLE 2. Dissimilarity measures for 18 Zayandehrood sites with TABLE 4. Results of testing hypothesis for definition of
monthly rainfall data. homogeneous regions (18 sites).

No. Site name Record length (yr) D(i) No. of Test Test Test
No. of stations statistic statistic statistic
1 Isfahan 49 2.49
simulations in region H1 H2 H3
2 Pol-Kalleh 45 1.88
3 Pole-Mazraeh 34 0.30 500 18 ⫺1.65 ⫺1.77 ⫺1.02
4 Garmaseh 34 0.88
5 Khajoo 34 1.05
6 Ziar 31 0.64 a region is reasonably homogeneous if Hi ⬍ 1 and that
7 Damaneh 34 0.19 a region is fairly homogeneous if 1 ⱕ Hi ⱕ 2. If Hi ⬎ 2,
8 Eskandari 26 1.31
9 Paiabe-Sad 32 1.65
then the region is absolutely heterogeneous.
10 Pole-Zamankhan 33 1.33
11 Tiran 33 0.26 2) GOODNESS-OF-FIT TEST
12 Khomeini-Shahr 12 1.21
Hosking and Wallis (1991) presented a measure of
13 Meimeh 26 0.76
14 Shahrokh 27 0.60 goodness of fit based on sample average regional kur-
15 Koohpaieh 27 0.54 tosis tr. This statistic is more applicable for three-
16 Firoozabad 26 0.15 parameter distributions, because all of the three-
17 Maqsoodbeig 26 0.98 parameter distributions fitted to data have the same t3
18 Mahiar 26 1.75
in L-Cs versus the L-Ck diagram. Quality of fitness can
be adjusted by the measure of difference between tr and
the value of ␶ DIST
4 upon the fitted distribution. The
Ns Ns
ZDIST statistic that is a measure of goodness of fit can
V3 ⫽ 兺 N 兵 关t
i⫽1
i
共i兲
3 ⫺ t3兴2 ⫹ 关t共4i兲 ⫺ t4兴2其1Ⲑ2Ⲑ 兺N.
i⫽1
i 共14兲 be written as

By simulation of a large number of sites by fitting a ZDIST ⫽ 共t4 ⫺ ␶DIST


4 兲Ⲑ␴4, 共16兲
four-parameter kappa distribution (Hosking and Wallis where ␴4 is the standard deviation of t4. The value of ␴4
1988), we can compute the heterogeneity measure by can be computed by simulation after fitting a kappa
distribution to the data (Hosking and Wallis 1988). If
Hi ⫽ 共Vi ⫺ ␮␷兲Ⲑ␴␷, 共15兲
|ZDIST| ⱕ 1.64, we can say the fitted distribution is rea-
where ␮␷ and ␴␷ are the mean and the standard devia- sonable.
tion of simulated data, respectively.
The four-parameter kappa distribution has several 3) PARAMETER ESTIMATION

useful features. As a generalization of the generalized A common problem in statistics and its application in
logistic, generalized extreme-value, and generalized hydrology, such as frequency analysis, is the estimation,
Pareto distributions, it is a candidate for being fitted to from a random sample of size n, of a probability distri-
data when these three-parameter distributions give an bution whose specification involves a finite number p of
inadequate fit, or when the experimentalist does not the unknown parameters. Analogous to the ordinary
want to be committed to the use of a particular three- method of moments, the method of L-moments obtains
parameter distribution. Hosking and Wallis (1993) used parameters by equating the first p sample L-moments
the four-parameter kappa distribution to generate arti- to the corresponding population quantiles.
ficial data for assessing the goodness of fit for different The exact distribution of parameter estimators ob-
distributions. Hosking and Wallis (1991) suggested that tained by this method is difficult to derive in general.
Asymptotic distributions can be found by treating the
estimators as a function of sample L-moments and ap-
TABLE 3. Dissimilarity measures for six Zayandehrood sites with
plying Taylor series methods.
monthly rainfall data.

No. Site name Record length (yr) D(i) TABLE 5. Results of testing hypothesis for definition of
1 Pole-Mazraeh 34 1.35 homogeneous regions (six sites).
2 Ziar 31 0.53
No. of Test Test Test
3 Eskandari 26 1.37
No. of stations statistic statistic statistic
4 Khomeini-Shahr 12 1.81
simulations in region H1 H2 H3
5 Meimeh 26 1.31
6 Firoozabad 26 0.63 500 6 ⫺1.06 ⫺1.11 ⫺1.24
498 JOURNAL OF APPLIED METEOROLOGY AND CLIMATOLOGY VOLUME 46

TABLE 7. Result of goodness-of-fit test for distributions fitted


(six sites).

Generalized Pareto GEV Generalized logistic


⫺3.25 ⫺0.69 0.5

3. Results and discussion


Maximum monthly rainfall records (the wettest month
in each year), with more than 12 data points, related to
18 gauging stations in Zayandehrood basin, have been
analyzed in the context of a regional analysis. Table 1
reports the values of some geophysical parameters and
record length for these sites, as well as the L-moments
FIG. 1. L-moments ratio diagram: L-C␷ vs L-Cs. and L-moment ratios for all stations.
Table 2 shows the results of applying the L-moment
test for the measure of dissimilarity. It seems that our
Hosking et al. (1985) and Hosking and Wallis (1987) grouping does not look particularly intuitive, because
found that asymptotic approximations are usually reli- there are no sites with D(i) greater than 3 (the maxi-
able for samples of 50 or more. Therefore, for sample mum value of this measure in this group is 2.49). Be-
sizes less than 50 we must use the regional analysis by cause of these results, we can say that this region does
the method of L-moments. not have any dissimilar stations. However, as shown in
For estimating parameters from a given distribution Table 2, there are some sites with the values greater
in the regional method, first the shape parameter of the than D(i), such as Isfahan, that may affect the other
distribution should be estimated from regional results of the procedure if the homogeneity assumption
weighted L-moments. In the regional estimation of pa- was not correct. For this reason, we have taken a ran-
rameters, all of the sites that are in one region should dom sample from this region and have calculated D(i)
have the same shape parameter value. After the esti- and other statistics necessary for the procedure. Table 3
mation of the shape parameter, the other parameters of shows the D(i) values for this sample region. As shown
the distribution will be estimated for each site. As will in this table, the values of D(i) for the sample region
be shown, because of the goodness-of-fit test, the gen- sites are closer to D(i) for the same sites in the region
eralized extreme-value (GEV) distribution should fit with 18 sites.
the data of the region. To test hypothesis for the definition of homogeneous
The probability density function of this distribution is regions, Monte Carlo simulation was performed with
the four-parameter kappa parent distribution, with re-
f 共x兲 ⫽ ␣⫺1兵1 ⫺ k 关共k ⫺ u兲Ⲑ␣兴 其共1Ⲑk兲⫺1

⫻ exp具⫺兵1 ⫺ k[共x ⫺ ␣兲/␣]其1Ⲑk典. 共17兲


The parameter estimators of this distribution are
z ⫽ 2Ⲑ共3 ⫹ t3兲 ⫺ log2 Ⲑ log3, 共18兲

k̂ ⬇ 7.8590z ⫹ 2.9554z2, 共19兲

␣ˆ ⫽ l2k̂共1 ⫺ 2⫺k̂兲 ⌫共1 ⫹ k̂兲, and 共20兲

û ⫽ l1 ⫹ ␣ˆ 关⌫共1 ⫹ k̂兲 ⫺ 1兴Ⲑk̂. 共21兲

TABLE 6. Result of goodness-of-fit test for distributions fitted


(18 sites).

Generalized Pareto GEV Generalized logistic


⫺6.26 ⫺0.79 1.78
FIG. 2. L-moments ratio diagram: L-Cs vs L-Ck.
APRIL 2007 ESLAMIAN AND FEIZI 499

TABLE 8. Estimated parameters for GEV distribution using sults shown in Table 3 for the 18-site region and in
at-site and regional procedures (18 sites). Table 4 for the sample region.
At site Regional Using a four-parameter kappa parent distribution
and deriving Hi for moments, we obtained values of
Site name u ␣ k U ␣ K
Hi ⬍ 1, indicating that all values of Hi are less than 1; on
Isfahan 33.63 14.28 0.180 33.21 12.23 0.041 the other hand, this region is a homogeneous region
Pol-Kalleh 40.11 16.04 ⫺0.040 41.44 15.98 0.041
(Table 5). The negative value indicates that the data
Pole-Mazraeh 33.14 12.73 ⫺0.002 33.56 12.95 0.041
Garmaseh 39.04 14.03 ⫺0.012 39.18 15.12 0.041 have dispersion that is less than the amount we expect
Khajoo 28.42 11.56 ⫺0.123 30.39 11.74 0.041 for a homogeneous region (Rao and Hamed 1999). The
Ziar 26.56 10.04 ⫺0.091 27.62 10.65 0.041 important point in using this homogeneity test is that
Damaneh 75.75 30.34 0.077 75.45 29.11 0.041 the primary measure for this test is the value of H1. This
Eskandari 92.68 30.66 0.141 88.26 34.05 0.041
is because, for small regions, H2 and H3 values may
Paiabe-Sad 60.17 22.77 0.016 60.42 23.31 0.041
Pole-Zamankhan 86.98 28.94 0.034 85.08 32.82 0.041 result in incorrect and virtual status of homogeneity.
Tiran 43.03 16.10 0.103 42.08 16.23 0.041 Furthermore, the L-moments ratio diagram (L-C␷ vs
Khomeini-Shahr 35.77 14.37 ⫺0.126 38.18 14.73 0.041 L-Cs) confirms this homogeneity (Fig. 1). However, it
Meimeh 25.74 10.78 0.112 25.57 9.86 0.041 seems that the Mahiar station has some differences with
Shahrokh 93.07 34.80 0.017 93.21 35.96 0.041
the other sites of the group because it is shown as an
Koohpaieh 27.86 10.41 0.178 26.74 10.31 0.041
Firoozabad 53.53 20.85 0.054 53.41 20.60 0.041 outlier event in this figure.
Maqsoodbeig 29.41 11.37 ⫺0.044 30.22 11.66 0.041 For making a decision about the primary distribu-
Mahiar 41.70 13.04 0.291 38.28 14.77 0.041 tions of this region, the goodness-of-fit test is applied as
mentioned above and then the GEV distribution is se-
lected through the candidate distributions, based on
TABLE 9. Estimated parameters for GEV distribution using their |ZDIST| ⬍ 1.64 (Table 6). For the sample region, it
regional procedure (six sites). also shows that the GEV distribution is a good choice
for this region, although the generalized logistic distri-
Site name U ␣ K bution has the same conditions (Table 7 and Fig. 2).
Pole-Mazraeh 33.59 12.83 0.038 The final step for a frequency analysis of hydrologic
Ziar 27.65 10.56 0.038 events is the parameter estimation of the selected dis-
Eskandari 88.34 34.74 0.038
tributions and, ultimately, the estimation of their quan-
Khomeini-Shahr 38.22 14.59 0.038
Meimeh 25.60 9.77 0.038 tiles. For this reason, we used both the at-site and re-
Firoozabad 53.46 20.42 0.038 gional procedures. Tables 8 and 9 report the results of
estimating the parameters of GEV for the region and

TABLE 10. Estimated quantiles for GEV distribution using at-site and regional procedures (18 sites). Here, T is return period.

At site T (yr) Regional T (yr)


Site name 10 20 50 100 200 10 20 50 100 200
Isfahan 61.50 68.59 76.64 81.83 86.66 60.35 68.57 78.86 86.32 93.54
Pol-Kalleh 77.91 90.74 107.91 121.21 134.84 75.80 86.13 99.06 108.42 117.49
Pole-Mazraeh 61.88 71.10 83.06 92.04 101.00 61.39 69.76 80.23 87.82 95.16
Garmaseh 71.07 81.50 95.13 105.44 115.81 71.68 81.45 93.67 102.53 111.10
Khajoo 58.44 69.95 86.44 100.11 114.95 55.59 63.17 72.65 79.52 86.17
Ziar 51.67 60.84 73.65 83.99 94.97 50.53 57.42 66.03 72.28 78.32
Damaneh 138.45 156.30 187.00 193.26 207.66 138.01 156.82 180.36 197.41 213.91
Eskandari 151.78 167.04 184.61 196.34 206.92 161.44 183.45 210.98 230.92 250.23
Paiabe-Sad 110.50 126.22 146.28 161.13 175.75 110.52 125.59 144.44 158.09 171.31
Pol-Zamankhan 149.67 168.71 192.67 210.13 227.11 155.56 176.84 203.38 222.61 241.22
Tiran 75.36 84.20 94.72 101.96 108.67 76.98 87.47 100.60 110.11 119.32
Khomeini-Shahr 73.19 87.58 108.26 125.44 144.12 69.84 79.36 91.28 99.90 108.26
Meimeh 47.20 52.99 59.83 64.50 68.80 46.78 53.16 61.14 66.92 72.51
Shahrokh 169.83 193.74 224.23 246.74 268.89 170.50 193.74 222.81 243.88 264.27
Koohpaieh 47.17 51.88 57.13 60.54 63.54 48.91 55.58 63.92 69.96 75.81
Firoozabad 97.68 110.69 126.78 138.32 149.37 97.70 111.02 127.68 139.75 151.43
Maqsoodbeig 56.37 65.54 77.90 87.50 97.37 55.29 62.83 72.25 79.09 85.70
Mahiar 63.22 67.62 72.09 74.73 76.88 70.03 79.58 91.52 100.17 108.54
500 JOURNAL OF APPLIED METEOROLOGY AND CLIMATOLOGY VOLUME 46

TABLE 11. Estimated quantiles for GEV distribution using


regional procedures (six sites). Here, T is return period.

T (yr)
Site name 10 20 50 100 200
Pole-Mazraeh 61.25 69.60 80.07 87.67 95.04
Ziar 50.41 57.29 65.90 72.16 78.22
Eskandari 161.07 183.03 210.55 230.54 249.92
Khomeini-Shahr 69.68 79.18 91.09 99.74 108.12
Meimeh 46.68 53.04 61.02 66.81 72.42
Firoozabad 97.47 110.76 127.42 139.51 151.24

FIG. 4. Quantile plot for site Eskandari.

sample region. As shown in these tables, the major dif-


ferences in these results are seen in the shape param-
eters. Conversely, the estimated parameters of location
and scale from the regional procedure do not have the
numerous differences with the estimated value from
at-site procedure using GEV distribution.
Comparing parameters estimated from the region
and the sample region indicates that the values have no
major differences and are very close to each other for
the location, scale, and shape parameters.
Sticking points about the predicted quantile from
comparing at-site and regional procedures, we can say
that the quantiles estimated from the at-site procedure
FIG. 5. Quantile plot for site Pole-Zamankhan.
for small return periods are similar to values predicted
from the regional method, particularly for sites with
record lengths of more than 30 yr. Through decreasing
the record length for some stations, the differences of
values from two methods increase, particularly for the
high return periods (Tables 10 and 11).
In the sites of Isfahan (Fig. 3), Eskandari (Fig. 4),
Pole-Zamankahn (Fig. 5), Tiran (Fig. 6), Meimeh (Fig.
7), Koohpaieh (Fig. 8), Maqsoodbeig (Fig. 9), and Ma-
hiar (Fig. 10), fitted distributions from both at-site and
regional methods have good harmony with the ob-
served data and their variations, but it seems that the
at-site estimates could predict quantiles for the high
return periods better than the regional estimates. Ac- FIG. 6. Quantile plot for site Tiran.

FIG. 3. Quantile plot for site Isfahan. FIG. 7. Quantile plot for site Meimeh.
APRIL 2007 ESLAMIAN AND FEIZI 501

FIG. 8. Quantile plot for site Koohpaieh. FIG. 12. Quantile plot for site Pole-Mazraeh.

FIG. 9. Quantile plot for site Maqsoodbeig. FIG. 13. Quantile plot for site Garmaseh.

FIG. 10. Quantile plot for site Mahiar. FIG. 14. Quantile plot for site Khajoo.

FIG. 11. Quantile plot for site Pol-Kalleh. FIG. 15. Quantile plot for site Ziar.
502 JOURNAL OF APPLIED METEOROLOGY AND CLIMATOLOGY VOLUME 46

FIG. 16. Quantile plot for site Khomeini-Shahr. FIG. 18. Quantile plot for site Paiabe-Sad.

cording to the above-mentioned sections, the distance regional and at-site methods are almost similar, and
of the regional estimation line from the observation one both methods have good harmony with the observed
has been enlarged by increasing the return periods, es- value for the low and high return periods.
pecially for the greater-than-50-yr return period.
Contrary to these sites, for some of the other sites in 4. Conclusions
this region, such as Pol-Kalleh (Fig. 11), Pole-Mazrae
(Fig. 12), Garmaseh (Fig. 13), Khajoo (Fig. 14), Ziar According to the results from applying L-moments to
(Fig. 15), and Khomeini-Shahr (Fig. 16), regional esti- Iranian rainfall data and associated ratio diagrams, it is
mates have better predictions of quantiles for the high found that the method presented is a useful and robust
return periods. For example, for the Khomeini-Shahr tool for confirming either similarities or differences in
site, which has the shortest record length in this region, the regional frequency analysis of the rainfall events.
the differences of regional and at-site estimates are Furthermore, the robustness of regional frequency
large. Furthermore, the regional estimates appear to be analysis based on L-moments for parameter estimation
more adequate than the at-site predictions. allows one to obtain relatively accurate quantiles from
An important point that has been shown in Fig. 10 is these distributions that will provide reliable estimates
the high differences between at-site and regional esti- of rainfall for most cases.
mates of quantiles for the return periods of more than It is found for stations having a record length of less
20 yr. Therefore, the regional rainfall predictions of than 30 yr (8 of 18 sites), based on at-site maximum
Mahiar are beyond the confidence intervals. According likelihood estimates, we obtain about 15%–20% under-
to this, in addition to the outlier manner of this site estimation for the greater-than-200-yr return period
shown in the L-C␷–L-Cs diagram, we can conclusively (Feizi 2003).
say this site is heterogeneous from the other sites of this It also seems that, for some cases, quantiles estimated
region. from the regional method could not give the reliable
For the remaining sites of this region, namely, Dam- prediction of such events, especially in samples having
aneh (Fig. 17), Paiabe-Sad (Fig. 18), Firoozabad (Fig. the L-moments ratio with a relatively large distance
19), and Shahrokh (Fig. 20), the estimates from both from the regional weighted mean of L-moments ratio.

FIG. 17. Quantile plot for site Damaneh. FIG. 19. Quantile plot for site Firoozabad.
APRIL 2007 ESLAMIAN AND FEIZI 503

REFERENCES

Dalen, J., 1987: Algebraic bounds on standardized sample mo-


ments. Stat. Probab. Lett., 5, 329–333.
Feizi, H., 2003: Analytical comparing of parameter estimation us-
ing L-moments and maximum likelihood estimates for some
station of Isfahan region. M. S. dissertation, Isfahan Univer-
sity of Technology, 115 pp.
Greenwood, J. A., J. Landwehr, N. C. Matalas, and J. R. Wallis,
1979: Probability weighted moments: Definition and relation
to parameters of several distribution expressible in inverse
form. Water Resour. Res., 15, 1049–1054.
Hosking, J. R. M., 1990: L-moments: Analysis and estimation of
FIG. 20. Quantile plot for site Shahrokh. distributions using linear combinations of order statistics.
Roy. Stat. Soc. London, 52, 105–124.
——, and J. R. Wallis, 1987: Parameter and quantile estimations
To overcome this problem, it might be necessary to
for the generalized Pareto distribution. Technometrics, 29,
accomplish some changes in values of homogeneity and 339–349.
dissimilarity statistics. ——, and ——, 1988: The 4-parameter kappa distribution. Re-
Another important outcome of this study is that the search Rep. RC 13412, IBM Research Division, Yorktown
homogeneity test and statistics are necessary for testing Heights, NY.
homogeneity and similarity of the sites in one region ——, and ——, 1991: Regional frequency analysis using L-
moments. Research Rep., Watson Research Center, IBM Re-
but it may not be enough to test fully for homogeneity.
search Division, Yorktown Heights, NY.
On the other hand, application of these statistics with- ——, and ——, 1993: Some statistical useful in regional frequency
out using the L-moments ratio diagrams, especially the analysis. Water Resour. Res., 29, 271–281.
L-C␷–L-Cs diagram, may result in incorrect homogene- ——, ——, and E. F. Wood, 1985: Estimation of generalized ex-
ity status. treme value distribution by the method of probability
The final point is that for regions with a large number weighted moments. Technometrics, 27, 251–261.
of sites, we usually have one distribution that passes the Rao, C. R., and K. H. Hamed, 1999: Flood Frequency Analysis.
CRC Press, 330 pp.
goodness-of-fit test; however, for the small region,
Wald, A., and J. Wolfowitz, 1943: An exact test for randomness in
there might be more than one distribution that could
the nonparametric case based on serial correlation. Ann.
pass the goodness-of-fit test. Math. Stat., 14, 378–388.
Wallis, J. R., 1988: Catastrophes, computing and containment:
Acknowledgments. We thank Ms. Tara Troy from Living with our restless habitat. Speculations Sci. Technol.,
Princeton University for her valuable editorial help. 11, 295–324.

You might also like