Outcomes of synthesis - Runoff Prediction in Ungauged Basins: Synthesis across Processes, Places and Scales - page 367

Austria. These differences highlight the importance of
data for predicting runoff in ungauged basins.

places, the notion of hydrological similarity has been used
as a central theme throughout the topic, to compare differ-
ent catchments and landscape units, and to learn from the
similarities and differences (just as medical practitioners
pool information from many patients to examine differ-
ences and similarities). All regionalisation methods are
based on the notion of hydrological similarity. This enables
both model structures and model parameters to be extrapo-
lated from gauged to ungauged catchments.
What makes two catchments similar, in terms of a particu-
lar runoff signature? In this topic we have looked at the
quantification of similarity in terms of runoff similarity,
climate similarity and catchment similarity, each of which
provides a particular view of hydrological similarity, and is
therefore only an approximation to what we are looking for.
One main underpinning of the concept of similarity is that the
current state of any catchment is the result of the co-evolution
of climate, soils, landscape and vegetation. Two catchments
can be considered similar if they have followed similar
trajectories of co-evolution, and are therefore functionally
similar. The different parts of the system are configured
similarly because they have undergone a similar history of
co-evolution, and therefore their hydrological response is
also likely to exhibit similar characteristics (e.g., runoff sig-
natures). Almost all regionalisation methods use

Runoff hydrographs: The L1 and L2 assessments give
lower performances than those in assessment of the Aus-
trian data ( Figure 12.4 ). This is because the performance
in Figure 12.4 has been obtained by a geostatistical
regionalisation method that takes the stream network
structure into account. The main difference between L2
and Figure 12.4 is in fact that the stream network struc-
ture is exploited by the geostatistical method, as the run-
off model performance for the Austrian region is similar
to those of the other studies in L2. This difference high-
lights the importance of including the stream network
structure in predictive models of the runoff hydrographs.

Low flows and floods: For these two signatures, the L1 and
L2 performances are in fact higher than those in Figure
12.4 , even though, again, there is a tendency for a lower
stream gauge density in L1 and L2 than in Austria. This
may be a surprising result but is in fact related to the method
by which the low flows and floods have been predicted. In
L1 and L2 targeted statistical methods were used to esti-
mate floods and low flows. In the assessment of the Aus-
trian data, however, the floods and low flows were
computed as the extremes of a regionalised continuous
runoff hydrograph. As discussed above, estimation of flow
extrema from a hydrograph model that has been optimised
for all values of the hydrograph tends to performworse than
using a different model that focuses on the extrema, and
Figures 12.3 and 12.4 suggest that this is actually the case.

' catchment

similarity to form catchment groupings, which then enables

model parameters and model structures to be related to
climate and/or catchment characteristics, leading to various
ways to generalise the results (regionalise them) to ungauged
catchments within the same homogeneous region.
There are two ways in which such similarity can be used.
The first is in a lumped way, to use some overall index (e.g.,
aridity index), to determine if catchments are similar in
some particular way (e.g., based on the process controls
on the partitioning of precipitation). The signatures or
model parameters can then be transferred across places.
The second is in a distributed way, such that landscape units
are considered similar if their local characteristics are simi-
lar, as is done in the topographic wetness index, or the
topography-based classification of runoff mechanisms
(Savenije, 2010 ), or in pedo-transfer functions (which are
si milarity relationships that relate texture to soil physical
characteristics and are used to generalise from the measure-
ment locations to the entire landscape). The L1 and L2
assessments were focused on catchment similarity in a
lumped way only. Of course, it would be very useful to
extend the assessment to include how the landscape simi-
larity indices perform also, but this is left for future research.

'

The distinction between the different methods of predicting
flood and low flow behaviour highlights an important point:
namely that improved hydrograph fitting should not neces-
sarily be the ultimate or only goal of predictions in ungauged
basins. Instead, methods must be optimised to predict spe-
cific signatures and their characteristics. In the Austrian
example, a targeted method for floods gave significantly
better performance (e.g. R² 0.76, see Section 11.10) than

those from the regionalised hydrographs (R²

¼

0.58) even
though the hydrographs used to estimate these floods had a
median regionalisation performance of NSE

¼

0.85, much
better than most runoff models in L1 and L2. A detailed
comparative approach focused on understanding individual
signatures and how they are connected may provide more
insights and eventually lead to better predictions than solely
focusing on reproducing the full hydrograph.

¼

12.2.2 Synthesis across places
Generic findings from chapters (non-assessment)
Because of the wide variety of runoff processes around the
world, it is particularly challenging to overcome the frag-
mentation across places. To achieve synthesis across

Assessment of performance as a function of climate
A key element of this topic has been a comparative hydrol-
ogy approach to address the fragmentation of hydrological
findings across individual case studies. Among other