PUB in practice: case studies - Runoff Prediction in Ungauged Basins: Synthesis across Processes, Places and Scales - page 313

Table 11.10. RRMSE values from leave-one-out
cross-validation for NSW

An important feature of the new RFFA methods is their
ease of application. Web-based software will allow users to
estimate flood quantiles at any location. It will also facili-
tate the management of future updates. In summary, the
new RFFA methods represent a significant improvement in
terms of data coverage, accuracy and ease of application
over the current Australian methods.

Model RRMSE(%)

PRT QRT

Fixed region ROI Fixed region ROI

Q 2 73 62 68 59

Q 5 65 54 70 59

11.12 UNDERSTANDING FLOW PATHS
FOR HYDROGRAPH
PREDICTION IN AN ANDEAN
CATCHMENT, CHILE
t. blume

Q 10 67 56 74 55

Q 20 72 57 83 53

Q 50 81 70 100 67

Q 100 90 75 100 72

The QQ-plots of the standardised residuals vs. normal
score (z score) for the fixed region (based on leave-one-out
cross-validation) and ROI were examined. Figure 11.43
presents results for the LP3 skew model, which show that
all the points closely follow a straight line. Overall, it was
found that the diagnostics did not reject the key GLS
regression assumptions of normality and homoscedasticity
of variance.
Table 11.10 presents the RRMSE values for the PRT
and QRT models with both the fixed region and ROI. In
terms of RRMSE, the ROI approach produces smaller
values than the fixed region approach for all the ARIs.
These statistics reveal that there are only modest differ-
ences between the performance of QRT and PRT. In view
of the previously noted practical advantages of PRT over
QRT, the PRT-ROI method has been selected for general
application to Australia, except for the arid and semi-arid
regions of interior Australia, where there is a very limited
availability of gauged data.

The issue from societal and hydrological perspectives
This study focuses on a small catchment in the Andes of
southern Chile. For many decades, extensive land use
changes have taken place in central and southern Chile,
leading to conversion of vast areas of farmland and native
forest to forest plantations of exotic species such as Eucalyp-
tus and Pinus radiata (Monterey Pine). Governmental sup-
port through subsidies caused an increase in the area under
plantation from 330 000 ha in 1974 to 1.5 million ha in 1992
and to 2.1 million ha in 2006. Land use changes of this spatial
extent cannot remain without consequences and are likely to
affect biodiversity, water and nutrient budgets as well as
erosion and sediment transport. In recent years tourism and
recreational land use (such as hiking and winter sports) have
been gaining more importance and thus a new kind of pres-
sure is exerted, especially in protected areas such as national
parks. On the other hand, southern Chile offers the rare
possibility to study hydrological and ecological systems that
have not experienced anthropogenic intervention.
Understanding undisturbed systems therefore becomes
increasingly important as many areas of the world are
subject to fast changes, either in land use or in climate or
even both. In contrast to systems with rapidly changing
land use or climate, anthropogenically undisturbed systems
of southern Chile are much more likely to be either close to
or at steady state. Understanding the processes and process
interactions in this steady state will help us to also improve
our understanding of disturbed systems, the shift in pro-
cesses that followed the disturbance, and maybe even the
future process evolution towards a new steady state
(Blume, 2008 ). However, the water authorities of most
countries focus their efforts of data collection on larger
and economically more relevant rivers; so when studying
undisturbed systems one can rarely build on existing data
sets and is thus faced with the challenge of working in an
ungauged or data-scarce catchment.

Discussion
As part of the revision of the flood estimation guideline
Australian Rainfall and Runoff, a major review of RFFA
methods was undertaken. The first step involved an exten-
sive data collation task with the assistance of state water
authorities. This produced a high-quality database of 676
stations, which was used to evaluate a number of RFFA
methods. In all the cases, leave-one-out cross-validation
was conducted to estimate the relative accuracy of a RFFA
model. It was found that Australia needs to be divided into
six regions based on climate, topography and data avail-
ability. For the four regions with adequate station cover-
age, a GLS regression-based approach linked with a ROI
approach that minimised predictive error performed best.
For the remaining two regions that coincided with the data-
poor arid regions, it was necessary to adopt the parsimoni-
ous index flood approach.