Professional Documents
Culture Documents
0083 PDF C03
0083 PDF C03
M p, r, s = E [ x F ( 1 – F ) ] = ∫ [ x(F )] F (1 – F )
p r s p r s
dF (3.1.1)
0
In particular, the following two moments M1,0,s and M1,r,0 are often
considered:
M 1, 0, s = α s = ∫ x(F )(1 – F )
s
dF (3.1.2)
0
M 1, r, 0 = β r = ∫ x ( F )F
r
dF (3.1.3)
0
where p, r, and s are real numbers. When r and s are equal to zero
and p is a non-negative number, Mp,o,o represents the conventional
moment of order p about the origin, µ′p . When p = 1 and either r or
s is equal to zero, then M1,r,0 = βr and M1,0,s = αs are linear in x and of
sufficient generality for parameter estimation (Hosking, 1986a). As x
takes only the power of one, simpler relationships are obtained
between the parameters of the distributions and probability weighted
N
a S = α̂ s = M̂ 1, 0, s = ---- ∑
1 N – i N – 1
x/ (3.1.4)
N i = 1 s i s
N
b r = β̂ r = M̂ 1, r, 0 = ---- ∑
1 i – 1 N – 1
x/ (3.1.5)
N i = 1 r i r
N
1
a S = α̂ s = M̂ 1, 0, s = ---- ∑ ( 1– F i ) x i
s
(3.1.6)
Ni = 1
N
ˆ r = M̂ 1, r, 0 = ---1- ∑ F i r x i
br = β (3.1.7)
Ni = 1
s r
∑ i ( – 1 ) ∑ i
s r
αs = βi , βr = ( –1 ) αi
i i
(3.1.8)
i=0 i=0
In particular:
α0 = β0 , β0 = a0
α1 = β0 – β1 , β1 = α0 – α1
α2 = β0 – 2β1 + β2 , β2 = α0 – 2α1 + α2
α3 = β0 – 3β1 + 3β2 – β3 , β3 = α0 – 3α1 + 3α2 – α3
r r
λr + 1 = ( – 1 ) ∑ p r, k α k = ∑ p r, k β k
r * *
(3.1.9)
k=0 k=0
where
p r, k = ( – 1 )
* r–k r r + k . (3.1.10)
k k
In particular,
λ1 = α0 = β0
λ2 = α0 – 2α1 = 2β1 – β0
λ3 = α0 – 6α1 + 6α2 = 6β2 – 6β1 + β0
λ4 = α0 – 12α1 + 30α2 – 20α3 = 20β3 – 30β2 + 12β1 – β0
τ = λ2 / λ1 (3.1.11)
τr = λr / λ2 , r ≥ 3 (3.1.12)
EXAMPLE 3.1.1
Middle Wabash
1. Uniform: τ3 = 0 , τ4 = 0
2. Exponential: τ3 = 1/3 , τ4 = 1/6
3. Gumbel (EV1(2)): τ3 = 0.1699 , τ4 = 0.1504
τ4 = τ3 (1 + 5τ3)/(5 + τ3)
7. Generalized Logistic:
τ4 = (1 + 5 τ 3 )/6
2
or τ4 = 0.16667 + 0.83333 τ 3
2
= ( t 4 – τ4 )/σ 4
DIST DIST
Z
Hosking and Wallis (1991) give two statistics which are used to test
regional homogeneity. The first statistic is a discordancy measure,
intended to identify those sites that are grossly discordant with the
group as a whole. The discordancy measure D estimates how far a
(i) (i) (i) T
given site is from the center of the group. If u i = [ t , t 3 , t 4 ] is the
vector containing the t, t3 and t4 values for site (i), then the group
average for NS sites is given by Eq. 3.3.1.
NS
1
u = ------- ∑ u i (3.3.1)
NS i = 1
NS
∑ ( ui – u ) ( ui – u )
T
S = ( NS– 1 )
–1
(3.3.2)
i=1
1 T –1
D i = --- ( u i – u ) S ( u i – u ) (3.3.3)
3
NS NS
(i)
∑ N i(t ∑ Ni
2
V1 = – t) / (3.3.4)
i=1 i=1
NS NS
t = ∑ N i t / ∑ N i
(i)
(3.3.5)
i=1 i=1
2. Based on LCv and LCs, the weighted average distance from the
site to the group weighted mean on a t vs. t3 graph is computed.
NS NS
(i) (i) 2 1⁄2
∑ N i {( t ∑ Ni
2
V2 = – t) +(t 3 – t 3 ) } / (3.3.6)
i=1 i=1
NS NS
(i) (i) 1⁄2
∑ N i ( t 3 – t 3) +(t 4 -t 4 ) } ∑ Ni
2 2
V3 = / (3.3.7)
i=1 i=1
H i = ( V i – µ v )/ σ v (3.3.8)
In order to illustrate the use of concepts and tests discussed above the
Wabash river basin data are presented as a case study in regionaliza-
tion.
As one might expect for such a large region, the results in Figures
3.4.1 and 3.4.2 suggest that the region as a whole is heterogeneous.
Data points are widely scattered in Figures 3.4.1 and 3.4.2. The points
on the LCs – LCk diagram (Figure 3.4.2) extend over the region covered
by several three-parameter distributions instead of clustering around
a particular distribution. This conclusion is also supported by the
results obtained from the regional homogeneity tests. The unbiased
estimators of PWMs in Eqs. 3.1.4 and 3.1.5 are used in these tests.
Values of the site discordancy measure Di, the heterogeneity measure
H, and the goodness-of-fit measure (ZDIST) were computed for the whole
region by using the FORTRAN computer program developed by Hosk-
ing (1991b). Of 93 sites, 12 were found to be discordant with the region
as a whole. The heterogeneity measures H1, H2 and H3 computed before
and after removing the discordant stations are given in the first row
of Table 3.4.1. These results indicate heterogeneity with respect to
both LCv values and the average distance within the LCv , LCs diagram
(Figure 3.4.1), and possible heterogeneity with respect to average dis-
tance within the LCs, LCk diagram (Figure 3.4.2). Hosking (1991b)
recommends the use of H1 to check homogeneity since both H2 and H3
can give false indication of homogeneity, especially in regions with a
small number of sites. A value of H1 = 5.99 in Table 3.4.1, which is
obtained before the discordant sites are removed, indicates a very high
degree of heterogeneity. In addition, the value of H2 is 2.69, supporting
this conclusion. The situation improves when discordant stations are
removed; but H1 is 3.15 even after removing the discordant stations,
suggesting that the region is heterogeneous. However, since the region
is homogeneous with respect to H3, the parameter estimates may be
obtained by using the regional average L-skewness along with the at-
site mean and LCv similar to the GEV(2) method of Lettenmaier et al.
(1987).
Figure 3.4.2. The LCs – LCk moment ratio diagram for 93 stations in the Wabash
river basin.
Discordant
All stations Stations Removed
Region NS H1 H2 H3 NS1 H1 H2 H3
ALL 93 5.99c 2.69c 1.25b 81 3.15c 0.10a –1.40a
UWAB 25 5.20c 4.03c 2.51c 24 4.97c 3.58c 2.06c
MWAB 15 1.85b 0.89a 0.10a 15 1.85b 0.89a 0.10a
WFWR 29 0.17a –1.56a –2.41a 28 –0.28a –2.05a –2.74a
EFWR 24 3.38c 1.79b 2.21c 22 1.24b 0.28a 1.10b
a = homogeneous, b = possibly heterogeneous, c = heterogeneous.
The main conclusion from the above analysis is that the Wabash
River basin is statistically heterogeneous. For the region as a whole,
the first row in Table 3.4.2 gives the goodness-of-fit measure for dif-
ferent distributions. The generalized extreme value distribution (GEV)
seems to be the best choice if the region is to be considered as a single
unit, although the ZDIST value is quite large for large portions of the
basin.
The stations were divided into four subregions according to their
geographical location: the upper Wabash basin (UWAB, 25 stations),
the middle Wabash basin (MWAB, 15 stations), the east fork White
River basin (EFWR, 24 stations), and the west fork White River basin
(WFWR, 29 stations). These subregions are shown in Figure 2.1.1. In
Table 3.4.1 the heterogeneity measures are given for the four sub-
basins before and after removing the discordant sites. The results in
Table 3.4.1 indicate that the upper Wabash basin is heterogeneous and
may have to be further subdivided in order to obtain reliable homoge-
neous regions. Both the middle Wabash and east fork White River
basin are possibly heterogeneous. The west fork White River basin, in
spite of being the largest (29 stations), is the most homogeneous, with
negative values of H indicating less spread than expected in a region
of such size. The goodness-of-fit measure ZDIST in Table 3.4.2 indicates
that the GEV distribution may be a good regional fit for the west fork
White river and upper Wabash regions. The LN(3) distribution is
suitable for the west fork White river, while the GLOG seems to be
suitable for the middle Wabash river and east fork White river regions.
However, the only reliable conclusion applies to the west fork White
river region (GEV and LN(3)) since non-homogeneous regions, by def-
inition, do not have a single parent distribution.
number of sites in the region. The results from this method are given
in Table 3.4.3. Considering the fact that both methods may give mis-
leading results about homogeneity in small regions, it can be concluded
that the results from Wiltshire’s method are in agreement with those
from the L-moments analysis.
χ 0.90 ( n – 1 ) χ 0.95 ( NS – 1 )
2 2
Region NS S Comments
ALL 93 155.97 109.76 115.39 heterogeneous
UWAB 25 85.27 33.20 36.42 heterogeneous
MWAB 15 17.57 21.06 23.68 homogeneous
WFWR 19 24.42 37.92 41.34 homogeneous
EFWR 24 39.88 32.01 35.17 heterogeneous
seen in this study, these regions are not homogeneous. The fact that
regression equations developed to estimate flows are quite different
for each of these regions also reflects the lack of homogeneity in the
USGS classification of watersheds.
Table 3.4.4. Estimated Quantiles for Station 95 in the WFWR Region for Different
At-Site and Regional Distributions
Period (years)
Observed
(approx.)
Return
At-site Regional
GEV LN(3) P(3) GLOG GEV LN(3) P(3) GLOG
10 65,640 64,879 65,280– 65,789+ 63,278 66,368+ 66,848 67,443 64,676–
20 76,846 77,448+ 77,586 77,545 76,242– 79,912+ 80,109 80,082 78,598+
50 92,259 94,855 94,150+ 92,613+ 96,122 98,836 98,063+ 96,334+ 100,076
100 112,752 108,795– 107,068 103,804 113,853+ 114,123+ 112,138– 108,433 119,348
150 122,608 117,298– 114,816 110,297 125,517+ 123,504+ 120,606– 115,463 132,074
200 127,536 123,497– 120,407 114,887 134,441+ 130,368+ 126,729– 120,437 141,836
Underlined are closest to observed; + and – indicate over or under estimated, respectively.
Figure 3.4.5. The LCs – LCk moment ratio diagram for the WEST region.
Figure 3.4.7. The LCs – LCk moment ratio diagram for the UPPER region.