Outcomes of synthesis - Runoff Prediction in Ungauged Basins: Synthesis across Processes, Places and Scales - page 373

model structure everywhere, whereas regional regression
allows for different predictors and coefficients in different
parts of the region.
Overall, the reviews of Chapters 5

process-based methods that account for flow paths would
lead to better predictions of low flows. The message in
Chapter 9 is that process-based methods are useful but
must be balanced by regional mapping and other Darwin-
ian approaches.
In other words, there is now the recognition that a
synthesis of process-based and statistical approaches is
necessary to improve predictions. Chapter 9 highlights
the benefit of learning from a synthesis of all relevant
information in the landscape, a notion termed flood
frequency hydrology. In Chapter 10 , the new approaches
are geostatistical methods that make explicit use of the
stream network structure. Overall, the general trend is that
predictions of some signatures where the best prediction
methods are statistical could benefit from more process-
based methods (this is the case for Chapters 5 , 6 , 7 and 8 ),
while prediction of signatures where more process-based
methods are being used could benefit from the use of
statistical approaches. This suggestion about the value of
a synthesis of statistical and process-based approaches to
improve predictions is equivalent to a call for a synthesis of
Newtonian and Darwinian approaches.
There are two different types of process-based
approaches: (i) physics-based distributed models based on
laboratory-scale equations, and (ii) conceptual,

10 reveal a tendency
for the selection of predictors by optimising the correlation
coefficient between the runoff signature and the predictors.
We suggest that this may not always be a good choice, and
instead hydrological understanding of relevant controls
should be used to guide selection of predictors (along with
the statistical analysis) and, importantly, in interpretation
of the coefficients found to fit the regression model well.
This interpretation should involve consideration of co-
evolutionary processes such as landscape evolution and
the stream network characteristics.
The reviews also suggest that the index method has
received more recent attention than the regression method.
Each index method hinges on an underlying principle, e.g.,
the use of the aridity index in the Budyko method for
predicting annual runoff. The assessment suggests that
index methods work fairly well, a good example being
the Budyko method for predicting annual runoff. Also,
the index methods for flow duration curves and floods
appear among the top-ranked methods. This suggests that
the notion of hydrological similarity, based on universal
principles such as the Budyko curve, may have value for
predictions of runoff in ungauged basins. It seems that
much can be learned by representing regional patterns in
a way that leads to questions about the co-evolutionary
principle(s) underlying them. The key is to represent infor-
mation in a way that reveals possible universal relation-
ships. This can then lead to better understanding of the
signatures, including their best predictors and how they
have arisen through co-evolution.

-

-
output models, with index-type models falling somewhere
in between. In general, mechanistic models surveyed in
the topic are used mostly in groundwater, or mixed
groundwater

input

surface water applications. Formal compara-
tive assessments for mechanistic distributed models are
rare, perhaps because these models are most data hungry,
and setting them up for a particular catchment is labour
intensive, involving numerous subjective decisions
regarding model parameterisation; repeating this in many
catchments at the same time poses enormous difficulties.
On the other hand, most runoff models used for predictions
in ungauged basins use conceptual (top-down) lumped
models. Our comparative assessment has been for this latter
case only. While a few inter-comparisons of distributed
(mostly conceptual) models have been reported (notably
the Distributed Model Inter-comparison Project, DMIP;
Smith et al., 2004b ; 2012 ), it appears that the biggest miss-
ing element in Chapter 10 is a performance assessment of
distributed (mechanistic) process-based (Newtonian)
models for ungauged basins. The purpose of such an inter-
comparison would not be to ascertain which model or model
group is preferable, but to learn from the differences in
model behaviour in different catchments.

-

The need for greater synthesis
The assessment has not been very conclusive with regards
to the comparison of process-based versus statistical
methods. There is very little literature on consistent com-
parisons of this type. However, the literature reviews in
Chapters 5 10 do suggest that for those signatures where

traditionally the focus has been on the statistics much
interest now resides in developing process-based methods
and, conversely, for those signatures where process-based
methods have been the norm new statistical methods of
predictions in ungauged basins have been developed. In
Chapter 5 , there has been a major move towards index
methods such as the Budyko method. The general conclu-
sion in Chapter 6 is that we must move on from mapping
approaches to invoke more process-based methods. The
general conclusion in Chapter 7 and Chapter 8 is that
prediction methods in the future must use process under-
standing in the regionalisation of flow duration curves and
low flows. In particular,

-

Dependence on climate
Since there exist important dependencies between predict-
ive performance and climate, one would also expect that
some methods work better in a given climate, while other the conclusion is that more