Geography Reference
In-Depth Information
rainfall input data are obtained from the Silo Data Drill
(Jeffrey et al.,
2001
), a data set gridded at a 0.05
(
Cross-verification of the spatialised models is achieved
by modelling each catchment using its local climate data
but with parameters taken from the nearest-neighbouring
gauged catchment. The distance to the nearest neighbour
(measured from centroid to centroid) ranges from 7 km to
278 km. This cross-verification procedure thus gives an
indication of the likely quality of ungauged basin predic-
tions that could be achieved if proximity was used as the
sole regionalisation criterion.
An alternative option for prediction in ungauged basins
is to use a model designed for continental-scale applica-
tions. In general, this requires a model whose design
incorporates many of the processes that control spatial
variability in runoff (e.g., land use, vegetation density).
The use of a single set of parameters to describe the
hydrological fluxes across the entire domain provides a
consistent modelling strategy across the continent, but also
makes greater demands on underlying data than conven-
tional runoff models.
Although a combination of daily efficiency and bias is
used for model calibration (for the lumped models), the
criterion used to assess model performance is the monthly
efficiency. This criterion is chosen because only monthly
5 km)
spacing. The Data Drill rainfall data are interpolated from
point observations of daily rainfall. Areal potential evapor-
ation data are also derived from the Data Drill.
~
Method
Models
Five lumped, conceptual, daily rainfall
runoff models are
calibrated separately on each of the 408 catchments:
AWBM (Boughton,
2004
), IHACRES (Croke et al.,
2006
), Sacramento (Burnash et al.,
1973
), Simhyd (Chiew
et al.,
2002
) and SMAR-G (Goswami et al.,
2002
). In this
study six parameters are optimised for AWBM, seven for
IHACRES, thirteen for Sacramento, six for Simhyd and
eight for SMAR-G (see
Section 10.4
).
Each model is operated using the gridded rainfall and
potential evaporation data in 0.05
×0.05
grid cells
across each catchment. For calibration, the observed
runoff at the catchment outlet is compared with a
spatial average of the modelled runoff in each grid cell
within the catchment. The lumped models are thus
spatialised to the extent that each grid cell within a
catchment has different climate input. However, the
same set of model parameters is used for all grid cells
within a catchment.
In addition, three continental-scale models are also
assessed. These are WaterDyn, which is implemented
in AWAP (
Raupach et al., 2009
), AWRA-L version 0.5
(van Dijk,
2010
)andCABLE(
Kowalczyk et al., 2006
).
Typically, such models use a single set of parameters to
describe the hydrological fluxes across the entire
domain. These models are also usually uncalibrated or
are calibrated using only a limited set of observations.
They often have modelling objectives other than or in
addition to runoff only.
-
-
but not daily
predictions are available (in this study) for
AWAP and CABLE. Analysis of daily efficiency statistics
for the six remaining models (not reported here) shows
strong correlation between trends in daily and monthly
efficiency. Prediction bias has also been assessed, and
although the results are not presented here, they reinforce
the conclusions that can be drawn from analysing monthly
and daily efficiency.
-
Model averaging
A complementary modelling approach that has potential to
reduce uncertainty in ungauged catchment predictions is
the use of ensemble techniques, whereby predictions from
different sources are pooled to produce a consensus pre-
diction (e.g., Viney et al.,
2009a
). Ensembles may be
constructed from different
Assessment in ungauged basins
One common regionalisation method is to transfer cali-
brated model parameters from a nearby gauged catchment
(e.g., Oudin et al.,
2008
; Chiew et al.,
2009
). This is
typically done using relatively parsimonious lumped catch-
ment models. The key assumption implicit in this approach
is that catchments in close proximity are likely to share
similar soils, topography, land cover and climate and that
they therefore have similar hydrological response charac-
teristics. Model parameters calibrated for one catchment
are therefore likely to predict runoff reasonably well for a
nearby catchment. To generate a continent-wide represen-
tation of runoff generation, runoff in each part of the
continent is modelled using parameters from the nearest
gauged catchment (see
Section 10.4.4
).
realisations of
the same
rainfall
runoff model (a single-model ensemble) or from
several structurally different models (a multi-model ensem-
ble). Several researchers have reported that the optimal
number of members for both single-model and multi-
model ensembles is about five.
Two ensemble or averaging schemes are also investi-
gated. One is a multi-model average in which a daily runoff
series is constructed by averaging the predicted flow of the
constituent models. Two such averages are assessed: one
using just the five spatialised models, and one using all eight
models. In each case the model average is unweighted. The
second ensemble scheme is a multi-donor averaging scheme
(an example of the single-model approach). This is applic-
able only to the spatialised models and involves averaging
-
Search WWH ::
Custom Search