Geography Reference
In-Depth Information
flood series data (Ishak et al.,
2010
). Large parts of Australia
were affected by severe drought starting in the 1990s, with
the consequence that the post-1990 annual maximum flood
series data for many stations were dominated by within-bank
flows. It is not confirmed whether the detected decreasing
trend in annual maximum flood series data for these stations
is a part of long-term climate variability or due to climate
change. Hence, it was decided to exclude those stations when
developing the current RFFAmethods. However, the impact
of climate variability and change on regional floods is being
further examined in order to develop some adjustment
factors to be applied to the regional flood estimates obtained
from the historic flood records.
the fixed region approach (Hackelbusch et al.,
2009
; Had-
dad and Rahman,
2012
).
It was found that QRT and PRT methods performed very
similarly for various Australian states. It is noted that the PRT
method offers several practical advantages over the QRT: (i)
PRT flood quantiles increase smoothly with increasing aver-
age recurrence intervals (ARIs), i.e., return periods; (ii) flood
quantiles of any ARI (in the range of 2 to 100 years) can be
estimated from the regional LP3 distribution; and (iii) it is
straightforward to combine any at-site flood information
with regional estimates using the approach described by
Micevski and Kuczera (
2009
) to produce more accurate
quantile estimates. For these reasons, the PRT has been
recommended for general application in Australia. However,
for the semi-arid and arid regions, due to data limitations, a
simple index-flood method (similar to Farquharson et al.,
1992
) has been adopted.
Method
A number of RFFA models were developed and tested
using the national database of 676 stations. These include
the probabilistic rational method (PRM) (
Australian Rain-
niques (such as those discussed in
Section 9.3.1
): quantile
regression technique (QRT) based on ordinary least
squares (QRT-OLS) and on generalised least squares
(QRT-GLS) (Tasker and Stedinger,
1989
; Reis et al.,
2005
; Gruber et al.,
2007
) and parameter regression
technique (PRT) based on GLS regression (PRT-GLS).
In the PRT, prediction equations were developed for the
first three moments of the LP3 distribution. Application of
the Hosking and Wallis (
1993
) test (even after considering
a number of possible alternative sub-regions) indicated that
Australian annual maximum flood data exhibited a high
degree of heterogeneity with an H statistic much larger
than 1. Hence, the index flood method was not considered
for general application in Australia.
Rahman et al.(
2011a
,
b
) compared QRT against the PRM,
which is the mainstay of RFFA methods currently in use
(Australian Rainfall and Runoff, 1987). They found that
QRT outperformed PRM. The choice of regression estima-
tionmethodwas then evaluated, withHaddad et al.(
2011a
,
b
,
c
) and Haddad and Rahman (
2011
) concluding that the QRT-
GLS method outperformed the QRT-OLS method.
The next phase of the assessment involved evaluating
the QRT-GLS and PRT-GLS methods using the fixed
region and region-of-influence (ROI) approaches (Burn,
1990a
,
b
; Zrinji and Burn,
1994
). A significant innovation
over previous ROI applications was the use of GLS pre-
dictive error to guide the selection of the stations included
in the ROI. The selected ROI contained the nearest N
stations, where N was selected so as to minimise the
predictive error, which accounts for both model and par-
ameter uncertainty. This strategy seeks to minimise the
heterogeneity unaccounted for by the regression predictors.
It was found that the ROI approach clearly outperformed
Results
Summary results of the application of the Bayesian-GLS
regression for the fixed region and ROI approach for the
state of New South Wales are presented below. The selec-
tion of the basin attributes took advantage of the ability of
GLS regression to differentiate between sampling and
model errors (for details, see Hackelbusch et al.,
2009
and Haddad and Rahman,
2012
). The basin attributes
reported below were the ones found to minimise model
error. Examples of the prediction equations are:
QRT (Fixed region: NSW):
ln
ð
Q
2
Þ¼
4
:
06
þ
1
:
26
:
z
ð
area
Þþ
2
:
42
:
z
ð
I
tc
,
2
Þ
ln
ð
Q
5
Þ¼
5
:
11
þ
1
:
19
:
z
ð
area
Þþ
2
:
08
:
z
ð
I
tc
,
5
Þ
ln
ð
Q
50
Þ¼
6
:
55
þ
1
:
01
:
z
ð
area
Þþ
1
:
73
:
z
ð
I
tc
,
50
Þ
ln
ð
Q
100
Þ¼
6
:
47
þ
0
:
97
:
z
ð
area
Þþ
1
:
50
:
z
ð
I
tc
,
100
Þ
PRT (Fixed region: NSW):
mean
¼
4
:
09
þ
0
:
67
:
z
ð
area
Þþ
2
:
31
:
z
ð
I
12
,
2
Þ
stdev
¼
1
:
22
−
0
:
59
:
z
ð
rain
Þ−
0
:
13
:
z
ð
S1085
Þ
skew
¼ −
0
:
42
−
0
:
10
:
z
ð
area
Þ−
0
:
10
:
z
ð
forest
Þ
where the explanatory variables are transformed as
n
X
n
1
z
ð
x
i
Þ¼
ln
ð
x
i
Þ−
ln
ð
x
i
),
i
¼
1
area is basin area in km
2
, I
12,2
is the design rainfall inten-
sity for 12 hours duration and 2 years average recurrence
Search WWH ::
Custom Search