Prediction of floods in ungauged basins - Runoff Prediction in Ungauged Basins: Synthesis across Processes, Places and Scales - page 202

standardised by their standard deviation or transformed in
another way to make them comparable. The ROI approach
is used in the UK Flood Estimation Handbook (IH, 1999 ).
It is possible to assign weights to the catchment character-
istics to give preference to some of them on the basis of a
prior understanding of what should be the important hydro-
logical controls (Kjeldsen and Jones, 2007 , 2009 ). Typical
characteristics used for the grouping are mean annual
rainfall, catchment area, and a baseflow index derived
from HOST soil data (hydrology of soil type classification
in UK).
In a case study in Arkansas, USA, Tasker et al.( 1996 )
found that the grouping by the ROI method produced
lower root mean square errors of the 50-year flood in
ungauged basins than other grouping methods. However,
Merz and Blöschl ( 2005 ) suggested that the ROI method
may not perform as well as other grouping methods (e.g.,
spatial proximity) if the catchment characteristics do not
reflect the underlying flood generation processes well.
Clearly, the choice of donor catchments and the availabil-
ity and choice of suitable catchment characteristics are
essential (see also Section 10.3.2). An example of the
selection of a pooling group for the UK is shown in Figure
9.10 . The target catchment is the River Dee at Polhollick
and the catchment characteristics used are catchment area,
mean annual precipitation, index of flood attenuation from
lakes and reservoir and the proportion of catchment
covered by the 100-year flood maps of the Environment
Agency.
Sometimes the catchment/climate characteristics are cor-
related, in which case it may be useful to transform the
characteristics by canonical correlation analysis (CCA).
This method has been applied to finding ROI pooling
groups by Cavadias ( 1990 ) and Ouarda et al.( 2008 ), and
also for identifying non-overlapping groups of similar
basins (as in Di Prinzio et al., 2011 ). Regression trees
(Breiman et al., 1984 ; Laaha and Blöschl, 2006a ) are
another grouping method that divides a heterogeneous
domain into a number of homogeneous groups by maxi-
mising the homogeneity of floods and catchment charac-
teristics within each group simultaneously. Burn ( 1997 )
used a genetic algorithm and Shu and Burn ( 2004a ) applied
the method of fuzzy expert systems and genetic algorithms
for the delineation of homogeneous pooling groups.
Numerous studies have used classification methods based
on artificial neural networks (e.g., Jingyi and Hall, 2004 ;
Lin and Chen, 2006 ; Srinivas et al., 2008 ; Di Prinzio et al.,
2011 ; Ley et al., 2011 ).
Whatever the classification method is, the pooling
groups will never be fully homogeneous in terms of hydro-
logical response. A number of methods have been pro-
posed to test for the homogeneity of the groups, i.e., to
examine if the groups obtained from climate/catchment

Figure 9.10. The pooling group for estimating floods for the River Dee
at Polhollick (indicated with a cross) consists of data from 20 gauged
catchments (dots) considered to be hydrologically similar to the subject
site. The FSR regions (NERC, 1975 ) are shown in colour.

characteristics are statistically similar in terms of runoff
(e.g., whether they have the same CV or the same growth
curve). A widely used homogeneity test is that of Hosking
and Wallis ( 1993 ). Viglione et al.( 2007b ) compared the
power of several statistical homogeneity tests and Castel-
larin et al.( 2008 ) showed how the cross-correlation among
sites can affect the performance of these tests. As these
tests use flood peak data, they can only be performed for
gauged basins, so their value for ungauged basins is
limited.
There is a trade-off between group size and regional
homogeneity. To avoid a biased estimate, the selected
stations should ideally belong to a reasonably homoge-
neous group in terms of hydrological processes, which
includes the ungauged sites where flood estimates are
required. However, there is also a need for a sufficiently
large sample of gauged sites to properly define the param-
eters in the model. In a Monte Carlo study with prescribed
type and level of heterogeneity, Hosking and Wallis ( 1988 )
discuss this trade-off (also see IH, 1999 ).