Environmental Engineering Reference
In-Depth Information
not be accurate everywhere, and that there will be
trade-offs in accuracy in different parts of the
domain with different sets of parameters. Pappen-
berger et al. (2007a), calibrating a one-dimensional
model of the River Alezette in Luxembourg in this
way within the GLUE methodology, showed that
none of the parameter sets tried could reproduce
observed inundation extent for all cross-sections.
They therefore suggested using parameter sets
that were shown to give good results for important
parts of the flow domain inmaking predictions for
those places in the future. Different places, how-
ever, might require different sets of models as
determined by matching the available observa-
tions from past events. New events could then be
used to refine the choice of models (Beven 2007).
field boundaries and other obstacles or pathways
for the flow over a floodplain can be represented.
The geometry of the channel itself is still a prob-
lem, since this requires relatively expensive field
surveys (although more widespread use of scan-
ning ultrasonic devices might also improve this
situation in the future).
It has often been assumed, however, that since
the physics of hydraulic models is well estab-
lished, and there is considerable experience
about what roughness parameters should be used,
the major source of uncertainty is probably in the
boundary conditions. This is the implicit assump-
tion made when, for example, roughness coeffi-
cients are defined a priori andwithout uncertainty
in flood risk mapping (in fact, often this is done
without taking any account of uncertainty in the
upstream input condition either). However, stud-
ies that have looked at the calibration of hydraulic
model parameters by comparing predictions with
observations of inundation extent have often
found that it is difficult to obtain good reproduc-
tion of the observations everywhere and that
rather wide ranges of roughness parameters
can give somewhat similar predictive accuracy
(see, e.g., Aronica et al. 1998; Bates et al. 2004;
Pappenberger et al. 2005a, 2007a, 2007b).
This is the result of similar calibration pro-
blems to those of distributed hydrological models.
In setting up distributed hydraulic models it is
necessary to provide depth/roughness functions
for every element (or for one-dimensional hydrau-
lic models, in each reach) of the discretized flow
domain. It is, in principle, possible to use different
functions or roughness coefficients for every ele-
ment or reach. This would, however, result in a
very high dimensional parameter space, so to sim-
plifymodel calibration it is often assumed that the
channel can be represented by one roughness co-
efficient or function and the floodplains (or each
type of surface on the floodplains) by another. This
is a gross simplification, remembering that it is
effective
roughness values that are required to
account for all forms of momentum loss within
each element, but it reduces the dimensions of the
calibration problem drastically. Consequently,
however, we should expect that the solutions will
Towards 'Models of Everywhere'
The application of distributed hydrological and
hydraulic models can be treated as a form of learn-
ing process about
places
. We should expect that
models that up to now have seemed to provide
acceptable predictions might not prove acceptable
into the future. Management based on the predic-
tions of suchmodels should consequently be adap-
tive. Beven (2007) has presented this learning
process in the context of 'models of everywhere'
- effectively the idea that distributed models
might be implemented for management purposes
for whole catchments or whole countries, and that
over time the accuracy in the local predictions
might be increased and the uncertainty of those
predictions might be reduced as either the sets of
appropriate parameters or the appropriate model
structures for different places are refined into the
future.
Model calibration has always been used to re-
flect the local idiosyncrasies of places (catchments
and river reaches) in the application of particular
model structures, at least where data have been
available to allowcalibration.What is envisaged in
the 'models of everywhere' concept is somewhat
different. Once everywhere is represented in the
system, and visualizations of model predictions
can be accessed by stakeholders interested in
