Environmental Engineering Reference
In-Depth Information
may be used to attribute differences in model
behaviour to differences in channel roughness
(Schumann et al. 2007b). This allows the defini-
tion of a model structure that uses additional
roughness parameters in order to strike the bal-
ance between model complexity and performance
at the local level where accurate field observations
are available.
aerial photography; and Hostache et al. (2009) for
an application to SAR]. Used in conjunction with
topographic maps or LIDAR this might lead to
vertical RMS accuracies of around 20-30 cm.
Retrieval techniques that combine imagery
with LIDAR topography and statistical data anal-
ysis have also been suggested. River cross-sections
are drawn perpendicular to the main channel,
and elevation data are extracted at the SAR flood
boundaries, assuming a horizontal water level at
each section. A smoothed linear trend of water
levels is estimated by using either a moving
average filter or spline interpolation (Matgen
et al. 2007a). A least squares estimation in flow
direction is another approach for water surface
approximation with respect to localized flow
behaviour (Schumann et al. 2007a). Additionally,
regressionmodelling allows reliable simulation of
stages at any location along the stream centreline
(with an RMS accuracy of
<
20 cm). These data can
be used in a GIS to generate a triangular irregular
network (TIN) mesh of coherent flood area and
stage across the inundated floodplain. As regres-
sionmodelling, particularly linearmodelling, may
be undesirablewhen integratedwithmore dynam-
ic hydraulic models, multiple water stage data
points may be extracted on river cross-sections
(Schumann et al. 2008a). This allows descriptive
statistics (e.g. mean, median or quartiles) to be
applied instead of a least squares estimation. The
advantage is that levels are now considered vary-
ing perpendicular to as well as in the direction of
stream flow, with a median accuracy generally
better than 50 cm.
Quantifying model performance based on
flood area observations
The most common procedure to assess model
performance is through an overlay operation of
single or multiple simulations of flood inundation
models with binary maps from remote sensing
based on wet/dry cells in a GIS. For this operation,
map outputs from 2-D models can be readily used
with a contingency table (also called confusion
matrix) that counts the number of correctly and
incorrectly flooded/non-flooded cells. One-
dimensional models require adequate post-proces-
sing of their output to render a binary map com-
parison possible. There exist a vast number of
performance measures based on these categorical
data, of which Table 11.3 provides a short sum-
mary and recommendation for flood studies
(Hunter 2005). However, the trouble with area-
based performance measures is that (i) there is not
a single best one, and (ii) each one is difficult to
interpret, mostly because easily predictable
dry cells within the contingency table are
misleading.
As an alternative to single area-based
performance measures, the modeller may use
a comparison of multiple measures or use a
fuzzy-rules-based measure, where a simple yes/no
(i.e. wet/dry) answer is augmented by a 'maybe'
relative to the certainty of a given cell being
flooded (Pappenberger et al. 2007). Such fuzzy
membership functions have been used success-
fully to evaluate multiple model simulations
within an uncertainty framework, such as the
Monte Carlo-based Generalised Likelihood Un-
certainty Estimation (GLUE; Beven and Bin-
ley 1992). A membership function reflects the
lack of knowledge about the real flood extent as
Integration with inundation models
Building and understanding model structures
As reported in the first part of this chapter,
remotely sensed data are crucial to build models
and define boundary conditions. Integration of
remote sensing has also been used to understand
the difference in behaviour of different model
structures. Localized error information, resulting
from a comparison of model simulations with
spatially distributed SAR-derived water stages,
