Environmental Engineering Reference
In-Depth Information
parameter values needed to run the model (see
'Calibration issues in using distributed models'
below) but onlyadistributedmodel canprovide this
type of local information. Under Section 105 of the
Water Resources Act (1991) in theUK, the Environ-
mentAgencyhas beenchargedwithprovidingflood
riskmaps for all areas at risk of flooding in England
andWales. It has done soby commissioning distrib-
uted inundation model predictions for all major
floodplains, generally for the100-year returnperiod
event (with and without an allowance for future
climate change). This requirement is being extend-
ed by the EUFloods Directive, which states that all
countries intheEUshouldproducesuchdistributed
floodmaps by 2013, and catchment flood riskman-
agement plans by 2015.
It is therefore worth considering why the initial
optimism about the future use of distributed
models has not been borne out by more recent
developments and applications. We have, after all,
access to far more computer power than 40 years
ago; we have access to far better topographic data
than 40 years ago; we have access to geographic
databases on soil and vegetation; we have access to
the distributed information in remote-sensing
images; we have strong drivers to use those dis-
tributed data sources to make local predictions of,
for example, the impact of land use and manage-
ment on flood runoff production and water qual-
ity. Legislation, such as the EUWater Framework
and Flood Directives, effectively requires such
predictions and both data and available computing
power should allow us to be much better at
distributed modelling.
A primary reason why the distributed
modelling effort has not been more successful in
hydrology and hydraulics is the result of
uncertainty: uncertainties in the representation
of hydrological processes (model structure);
uncertainties and incommensurability in input
data; uncertainties in estimating model para-
meters; and uncertainty and incommensurability
in the observations with which model predictions
are compared. We will return to these limitations,
and their implications for distributed flood inun-
dation prediction, after considering the range of
distributed catchmentmodels currently available.
computers some 40 years ago has meant that the
computational restriction has become less of an
issue over time. More runs of distributed models
with finer discretizations can now been made. It
has not disappeared as an issue, however, since the
parameter definition problem has not gone away
(which might require many runs to be made in
model calibration or uncertainty estimation), and
computational times for fine-resolution models
over large catchment domains may still be long
compared with the lead times required in flood
forecasting. Thus when using a distributed model
for large-scale systems, resolution will generally
still need to be compromised, such as the 5-km
grid used in the European Union (EU) European
Flood Alert System, which makes distributed
predictions of runoff generation for the whole of
Europe driven by rainfall forecasts from the Euro-
pean Centre for Medium-term Weather Forecasts
up to 10 days ahead (see De Roo et al. 2003).
There is a further problem in the use of distrib-
utedmodels in forecasting, which is the number of
state variables that could be updated in real-time
data assimilation. Weather forecasting models
also have very large numbers of state variables of
course, and now use data assimilation as a matter
of course. This is one reason for the improvements
in forecast accuracy over the last two decades. In
that case, however, there are also large numbers of
observations to be assimilated. In the flood case,
we may have observations of water levels at only
a small number of sites. The information content
of a forecast innovation (difference between
observed and predicted values) may then not be
sufficient to support assimilation of the large
number of distributed model variables without
making rather strong assumptions (see 'Data
assimilation issues in using distributed models'
below). Thus distributed models might still not
be the best strategy for real-time forecasting
problems (see Chapter 9 [Young], this volume).
In flood risk assessment, however, distributed
models might be much more useful. The require-
ment is then to prioritize the local areas at greatest
risk by coupled hydrological and hydraulic
routing. The results will be dependent on the
specification of the large number of inputs and
