Geography Reference
In-Depth Information
(Atkinson, 2000) and principal component analysis
(Matgen et al., 2006) may also be applied. Image
statistics-based active contour models have been used
by Bates et al. (1997), Horritt et al. (1999), de Roo et al.
(1999), Horritt et al. (2001) and Schumann (2005).
Classification accuracies of flooded areas (most of the
time defined as a ratio of the total area of interest where
classification errors are omitted) vary considerably and
only in rare cases do classification accuracies exceed
90%. Table 6.3 summarises strengths and limitations of
the most widely used image processing methods, most of
which are available in any standard commercial remote
sensing software package. From this list it is apparent
that no single method can be considered appropriate
for all images, nor do all methods perform equally well
for a particular type of image or flood event (Schumann
et al., 2009a).
Apart from flood area and extent, if an accurate and
high enough resolution digital elevation model (DEM,
generated from e.g. LiDAR) is available and provided a
flood boundary can be adequately extracted from SAR
imagery, it is possible to also map water levels and flood
depths for a given event. Various techniques for this
are listed in Table 6.4 and as early as the 1980s there
have been several attempts to successfully derive water
stages or heights from remote sensing datasets. In general,
accuracies of the resulting water stages or heights increase
with the complexity of the method and the resolution of
the datasets used.
binary wet/dry maps of flood area extracted from two
simultaneous SAR acquisitions of a flood on the River Dee
(NE Wales, UK) using five widely used image processing
algorithms leads to a multi-algorithm ensemble map that
contains more meaningful information than any of the
single binary wet/dry maps alone. The authors then went
on to show the value of this map for calibration of a flood
inundation model. More details on this aspect are given
later in Section 6.4.1.
In a similar context, Schumann et al. (2010a) used
uncertainty information in SAR derived flood edges as
well as in coarse resolution topographic heights from the
NASA SRTM (Shuttle Radar Topography Mission) DEM
to generate water surface gradients for large rivers reliable
enough to reject erroneous flood model simulations or
even to distinguish between flood models of different
complexities thereby helping select the most appropri-
ate model structure for a given flood event (Prestininzi
et al., 2011).
A general prerequisite to estimate and eventually reduce
uncertainties or use them in a meaningful way is to first
identify and understand the sources. Therefore, the next
section describes the most common flood detection error
sources in more detail.
6.3.2 Sourcesoffloodandwaterdetectionerrors
As stated previously, surface roughness is the main factor
affecting radar backscattering whereas the dielectric prop-
erties control the intensity of the signal. However, there
are a relatively large number of other factors affecting
image 'quality' as alluded to in the next paragraph.
Image classification or interpretation errors (i.e. dry
areas mapped as flooded and vice versa) may arise
from a variety of sources and adversely affect algorithm
performance as illustrated in Table 6.4. Error sources
may include inappropriate image processing algorithm,
altered backscatter characteristics, unsuitable wavelength
and/or polarisations, unsuccessful multiplicative noise
(i.e. speckle) filtering, remaining geometric distortions,
and inaccurate image geo-coding. Horritt et al. (2001)
state that wind roughening and the effects of protrud-
ing vegetation, both of which may produce significant
pulse returns, complicate the imaging of the water sur-
face. Moreover, due to the corner reflection principle (i.e.
where the structure of rectangular surfaces, e.g. buildings,
is such that the wave is returned to the SAR antenna
and thus causes complete sensor saturation resulting in
white image pixels (Rees, 2001)) in combination with the
relatively coarse resolution of many SAR systems means
6.3.1.4 Including error and uncertainty analysis
In recent years, there has been an increasing interest in
environmental science to assess uncertainty in observa-
tions and models (Beven, 2006). Error and uncertainty
analysis is a process which should be inherent in all stud-
ies involving remotely sensed data as these are prone to
much uncertainty, the sources and magnitudes of which
are relatively easy to identify and quantify. According to
Atkinson and Foody (2002), characterising the sources of
uncertainty and improving the uncertainty information
can lead to a reduction of uncertainty in remote sensing
products which in turn increases not only accuracy but
also credibility. Yet, in the case of SAR remote sensing
and flood mapping, the observed flood area or flood edge
data used are commonly treated as deterministic when in
reality these data are subject to considerable uncertainty.
In one of the first attempts to deal with uncertainty
in flood mapping from SAR, Schumann et al. (2009a)
demonstrate that fusing a large number of deterministic
Search WWH ::
Custom Search