Prediction of runoff hydrographs in ungauged basins - Runoff Prediction in Ungauged Basins: Synthesis across Processes, Places and Scales - page 256

in Section 11.20 for 198 stream gauges (and numerous
subcatchments) in Sweden. They defined hydrological
response units on the basis of land use and soil type. In a
stepwise procedure they calibrated 15 parameters for each
land use and soil type, and another 10 global parameters.
They calibrated parameters related to soil parameters first,
after which another group of parameters (e.g., river
routing) was calibrated. The calibration proceeded from
upstream to downstream catchments. Internal model vari-
ables (such as flow components) were checked for plausi-
bility. A similar stepwise approach was proposed by
Blöschl ( 2008 ) and Blöschl et al.( 2008 ), who first identi-
fied the parameters related to the seasonal scale (e.g.,
related to evaporation and groundwater) and then the par-
ameters related to the event scale by stratifying events by
type (e.g., synoptic, convective and snowmelt events).
Their parameter identification approach was supported by
groundwater level and flood inundation data. Hundecha
et al.( 2008 ) proposed regional estimation of parameters
based on kriging in the

and other data, which usually improves the internal
(ungauged) runoff predictive performance over the use
of a-priori data (e.g., Bandaragoda et al., 2004 ; Pokhrel
et al., 2008 ).
In the Distributed Model Intercomparison Project
(DMIP; Reed et al., 2004 ) 12 distributed models were
compared in three basins in the Oklahoma region, to
assess (among other things) how well the distributed
models perform at internal (ungauged) nodes. They
found that, on average, the performance was poorer in
the ungauged internal catchments than at the catchment
outlet and that some element of calibration improved the
performance over the use of a-priori parameters. In a
follow-up project (Smith et al., 2012 ), 16 models were
compared. Figure 10.28 shows some of the results of
that inter-comparison. The median correlation coeffi-
cients for the calibration period (solid line in Figure
10.28 ) at the gauged locations are around 0.7. As one
moves to ungauged internal nodes, the performance
drops to about 0.5, although there is a lot of variation
between catchments and between models. Similarly, the
validation performance (dashed line) drops from about
0.6toabout0.4.Thereisinfactatrendfromlargeto
small basins, indicating that the predictive uncertainty
increases from large gauged basins to small ungauged
basins. It is interesting that all models gave relatively
poor performance at basin eight (Blue River at Conner-
ville), which is due to the complex hydrogeology of the
basin (Halihan et al., 2009 ) that none of the models were
able to capture. This points to a need for understanding
the hydrological processes in a catchment prior to runoff
modelling, based on any available information (in this
case, in particular, hydrogeology) beyond the runoff
hydrographs, to ensure that the model is right for the
right reasons.

space of the catchment

characteristics.
In these studies, the HRUs are usually defined at the
element scale, but there is a lot of variability within each
model element. To account for this subgrid variability,
Samaniego et al.( 2010a , b ) proposed a multiscale par-
ameter regionalisation method where parameters at a
large grid scale are related to parameters at a finer scale
by upscaling operators such as the harmonic mean. The
fine-scale parameters are then related to fine-scale catch-
ment characteristics in a similar way as in other studies
(e.g., Hundecha and Bárdossy, 2004 ; Götzinger and
Bárdossy, 2007 ; Hartmann and Bárdossy, 2005 ). An
application of this method to the Upper Neckar basin
in Figure 10.27 shows the porosities of the top soil layer
at various scales (from 1 to 8 km) estimated by both
their new (multiscale) method and the traditional (stand-
ard) method. For comparison, porosity at a scale of 100
m is shown. Figure 10.27 indicates that the multiscale
method preserves the fine-scale spatial pattern signifi-
cantly better than the standard methods as the grid size
increases. The non-linear aggregation effects are more
realistically represented than in the standard method,
which makes the method less dependent on model ele-
ment size.
When using distributed models in the downscaling
procedure, the total number of calibration parameters
decreases as compared to the product of model elements
and model parameters per element. For example,
Samaniego et al. (2011) use 28 parameters per element,
which gives a total of 28 000 parameters for 1000 cells.
With the downscaling procedure, only 62 coefficients
need to be calibrated. The smaller number of parameters
allows their identification more robustly from runoff

10.4.5 Constraining model parameters by dynamic
proxy data and runoff
Model parameters in ungauged basins can be obtained
from catchment characteristics a priori ( Section 10.4.3 ),
they can be transferred from calibrated parameters in
neighbouring catchments ( Section 10.4.4 ), or they can be
estimated from dynamic data in the ungauged catchments
of interest, such as soil moisture or regionalised runoff.
The latter is covered in this section. These three paths to
estimating model parameters are not mutually exclusive.
All combinations are possible, depending on data avail-
ability, and have been analysed in the literature on
ungauged basins. In fact, the usual approach is to use the
dynamic data to constrain the parameters beyond what is
estimated from alternative sources (a priori or transferred
from gauged catchments). For spatially distributed models,