1 Spatial Modelling of the Terrestrial Environment - Georeferencial

98 Spatial Modelling of the Terrestrial Environment
dynamic flooding processes. However, the responsive mode operation of such systems
would also allow high temporal and spatial resolution sampling through individual flood
events to be conducted during a dedicated airborne campaign. This may be the only way
to generate multiple synoptic images of flood extent that show the dynamic extension and
retreat of the flow field. We currently lack such data, yet they are essential if we wish to
rigorously validate the dynamic behaviour of flood inundation models and develop them
beyond a relatively modest accuracy.
Finally, only a relatively limited number of model classes have been tested in this way and
the process needs to be extended to incorporate other standard model types such as ISIS and
HEC-RAS and the results inter-compared. This needs to be accomplished within a formal
uncertainty analysis framework and the result is likely to be an appraisal of limitations with
existing models. This will then form an objective basis for a model development programme
aimed at producing better inundation models
5.3.5 Uncertainty Estimation Using Spatial Data and Distributed
Risk Mapping
Prior to rigorous validation of the friction parameterization methodology of Mason et al.
(in press), poorly constrained calibration will remain a necessary step in any modelling
study. Even with this in place, we may still be left with a residual, if better constrained,
calibration problem, which will also lead to uncertainty in model predictions. Calibration
is, therefore, always likely to be present in any model application to real-world data, and as
observed data are typically sparse and contain errors, potentially many different calibrations
and model structures will fit available data equally well. Beven and Binley (1992) term this
the ‘equifinality problem’. These multiple acceptable model structures will, however, lead to
different predictions. Hence, a deterministic approach to conveyance estimation, whereby
single sets of calibrated coefficients are used to make single flood extent predictions, may
be problematic. This is particularly true if a set of calibrated coefficients from one event
is used to predict flood inundation for a further, larger event for which observed data
are unavailable. There will be multiple acceptable flood envelopes that fit the available
calibration data and hence, design flood extent is better conceived as a fuzzy map. This
approach has been advocated by a number of authors who have deployed Monte Carlo
analysis to derive inundation probability maps that take account of model uncertainties
(Romanowicz et al., 1996; Romanowicz and Beven, 1997; Aronica et al., 2002).
For example, Aronica et al. (2002) explored the parameter space of the two-dimensional
storage cell model, LISFLOOD-FP, developed by Bates and De Roo (2000). The model was
applied in two basins to separate floods, one of which (the Imera River, Sicily) represented
a basin-filling problem controlled by embankments, whilst the other (the River Thames,
UK) represented a more typical out-of-bank flow in a compound channel. The friction
parameter space was treated as two-dimensional and comprised single values of Manning’s
n for the channel and floodplain. A Monte Carlo ensemble of 500 realizations of the model
was conducted to explore this parameter space using a random sampling scheme. The
simulations were then validated against a single observed inundation extent available in
each basin using the F 〈2〉 performance measure given in equation (1).
Plots of this objective function were then mapped over the parameter space and are shown
in Figures 5.5 and 5.6. These show the calibration response of the LISFLOOD-FP model