Outcomes of synthesis - Runoff Prediction in Ungauged Basins: Synthesis across Processes, Places and Scales - page 380

Figure 12.13. Synthesis of the Newtonian and Darwinian approaches
for the case of uncertainty estimation.

Figure 12.14. Comparative uncertainty estimation based on
Darwinian concepts consists of exploiting the differences between
regions to learn from them.

ensemble of predictions in different places. The difference
between the traditional approach and the one adopted in
this topic is multifaceted. First, there is no error propaga-
tion involved in the analyses of this topic; instead cross-
validation performance is used as an estimator of total
uncertainty. Second, uncertainty was analysed in a com-
parative way for a large number of catchments from around
the world, so as to exploit the differences between regions
to learn from them. An interesting question is why the
uncertainties in these catchments are different, and what
controls these differences. The focus, therefore, is not on
Monte Carlo techniques or other error propagation
methods, nor on reducing the uncertainty by optimisation
schemes or data assimilation techniques. The comparative
framework is about identifying the patterns so as to learn
from them. Sections 12.2 and 12.3 of this chapter, as
reviewed in detail, are about understanding the controls
on model performances. We have investigated why per-
formance decreases with increasing aridity index, i.e., why
runoff predictions are more uncertain in arid catchments
than they are in humid catchments, and interpreted these
differences in terms of the underlying controlling factors.

models. The comparative uncertainty model, on the other
hand, is more in tune with the notion of co-evolution.
There are patterns of predictability due the complex nature
of catchments that may not be easily foreseen, which has
led to the notion of ' '

(Blöschl and Zehe, 2005 ).
The presence of karst, for example, is hardly predictable in
standard uncertainty models. What is therefore needed is a
new uncertainty framework for PUB that accommodates
the synthesis of the Newtonian and Darwinian approaches
to predictions, to build on the strengths of both: a combin-
ation of error propagation methods with comparative per-
formance and uncertainty assessment of the kind
performed in this topic, including all error sources, as well
as model structure error ( Figure 12.13 ).
This unified uncertainty framework for predictions in
ungauged basins may start from comparative performance
analyses just like those conducted in the assessment chap-
ters in this topic. The comparative uncertainty assessment
focuses on why particular uncertainty patterns come about
and why the uncertainty differs between regions ( Figure
12.14 ). By contrasting different regions, the differences
in the uncertainty can be detected and Newtonian
approaches (such as process-based error propagation)
may help in understanding these differences in terms of
co-evolutionary processes. An interesting question is why
the uncertainty is different in different regions of the world
for a given model type, a given data availability and a
given runoff signature that is to be predicted. This will
contribute to enhanced understanding of the uncertainty
associated with hydrological predictions in ungauged
basins (more than can be extracted from individual case
studies). Rather than heralding (in a paper) the

outliers

Synthesis of the two uncertainty paradigms
Which of the two paradigms should be preferred for quan-
tifying predictive uncertainty in ungauged basins? The
truth is that both uncertainty paradigms have strengths
and weaknesses, just as the Newtonian and Darwinian
approaches do. Both are ways of understanding and quan-
tifying the uncertainty, and are in fact complementary. The
Newtonian uncertainty analysis approach is able to attri-
bute individual error sources, and optimisation schemes
and data assimilation techniques have an important role
in optimal prediction for operational use. On the other
hand, natural catchments around the world are complex
objects, so Newtonian type sensitivity analyses may not
explore the entire range of uncertainties, including any
feedbacks across processes and scales, and dependencies
in the error sources that may not be apparent in Newtonian

of
a modelling exercise because the uncertainty was small, or
was reduced by a particular method, the aim should be to
contribute to understanding. The combined comparative
and error propagation uncertainty framework will contrib-
ute to a unification of what has been learned about

' success '