Outcomes of synthesis - Runoff Prediction in Ungauged Basins: Synthesis across Processes, Places and Scales - page 366

in both cases R² of specific Q 95 low flows were used and
the selection of the catchments was similar. Also, for
floods, the results of L1 and L2 are consistent as in both
cases R² of specific 100-year floods were used. However,
in L2 there is a tendency to include studies with a larger
number of catchments and longer runoff records, which
explains why there are fewer studies with low perform-
ances. Finally, for the case of runoff hydrographs, there
was a tendency to select studies with a large number of
catchments with somewhat better NSE performance than in
the case of L1. It should be noted that the ranges in Figure
12.3 do not represent the variability of performance
between catchments but the variability of median perform-
ances of different studies. For hydrographs, the median
performances of the L2 assessment are remarkably similar
and vary little around 0.70.

1.0

NSE

0.8

0.6

R 2

0.4

Ch. 5
Annual
runoff

Ch. 6
Seasonal
runoff

Ch. 7
FDCs

Ch. 8
Low
flows

Ch. 9
Floods

Ch. 10
Hydro-
graphs

Figure 12.4. Same as Figure 12.3 but for Austria, to help interpret
Figure 12.3 , based on top-kriging of hydrographs as in Figure 12.1 .
Note: R² (open circles) for mean annual runoff, range of Pardé
coefficient, slope of the flow duration curve, Q 95 low flow, Q 5 high
flow as a proxy for floods, runoff integral scale (i.e., measure of the
average time lag over which runoff is correlated); median NSE
(open triangles) for monthly runoff and daily runoff. From Viglione
et al.( 2013b ).

How well can we predict signatures relative to each other?
Figure 12.3 shows a decreasing trend in the quality of
predictions as we move from annual runoff to seasonal to
daily (FDC) time scales; low flows and the predictions
of flood magnitudes were generally the poorest. The
hydrograph predictions are somewhat better than those of
the floods. The strength of the Level 1 and Level 2 assess-
ments is that they cover a wide spectrum of processes,
locations and methods, but this implies that the data set is
quite heterogeneous as the climates, methods, data quality
and observation periods differ between the signatures
( Figure 12.3 ). Also, not all the performance measures are
defined in exactly the same way. To help interpret Figure
12.3 and conduct a more consistent comparison across
signatures, one consistent data set and one consistent
method used in the chapters were selected to explore the
relative performance of the predictions for each of the
signatures. The method selected was top-kriging (a gener-
alised version of kriging that accounts for river network
structure) since it was assessed in most chapters. To
enhance consistency the hydrographs were regionalised
before the signatures were calculated, and the performance
was assessed by cross-validation, in the same way it had
been performed in all the chapters.
Figure 12.4 indicates that the performance is best for
season al and annual runoff and runoff hydrographs, and is
poorer for the prediction of low flows and floods. The
higher predictability of mean annual runoff and seasonal
runoff is due to the aggregation of runoff variation over a
relatively long time period. They therefore vary more
smoothly in space, which enhances their predictability. In
contrast, low flows and floods are extremes. Extremes are
generally harder to predict than averages, in part because
they represent considerable process change, compared to
the mean. Low flows are easier to predict than floods
because droughts tend to persist over larger areas and

longer time scales, making the estimation of low flows
from other stream gauges in the area fairly robust. The
Q 95 low flow signature used in Chapter 9 is a less extreme
runoff signature than the 100-year flood used in Chapter 9 ,
and can be estimated more reliably from a runoff record of
a given length, which may contribute to the better cross-
validation performance. Although floods are not predicted
well, runoff hydrographs can be predicted with some con-
fidence. This is because most parts of the hydrographs are
easy to predict. Although the extremes are harder to pre-
dict, the model efficiency metric treats all time steps with
the same weight, reducing the impact of poorer predictive
capacity for the flow extrema.
Due to their internal consistency, the model perform-
ances presented in Figure 12.4 can be used as a standard
against which the assessment results of the topic can be
evaluated. Generally the pattern across signatures of the L1
and L2 assessments remains consistent with those pre-
sented in Figure 12.3 . However, there are some subtle
but important differences that shed light on the relative
performance of the methods. The causes of the differences
depend on the signatures:

Annual runoff, seasonal runoff and flow duration curves:
For these three runoff signatures the median performances
of the L1 assessments (grey line in Figure 12.3 ) are some-
what lower than those of the Austrian data set ( Figure
12.4 ). This is because L1 involves a number of studies
with a lower stream gauge density than is available in