Geography Reference
In-Depth Information
two acquisitions. Pattern matching approaches will obvi-
ously be sensitive to such changes in the scenery. For
example, seasonal changes in vegetation and/or water
colour can, theoretically, have a significant impact on such
approaches. Additionally, it was found that rapid changes
in flow regime can significantly change river imagery and
negatively impact automated georeferencing attempts.
For example, in addition to the airborne data shown in
Figure 8.8, Carbonneau et al. (2010) also tested the use of
Quickbird imagery acquired three days after the sensed
imagery. In such a short time span, the surrounding veg-
etation and infrastructure remained virtually the same.
However, during this period, river discharge decreased
significantly and consequently, dry, exposed, gravel bars
in the Quickbird reference image occupy a significantly
larger area. In extreme cases, some very shallow riffles
became exposed thus causing an important change in
the rivers' appearance. This effect was at least in part
responsible for an increase of the RMS error to 5.5 m.
a new approach to normalise brightness values in flu-
vial hyperspatial image datasets. This method is based
on a similar principle to the automated georeferencing
described above. This georeferencing procedure relied
on the presence of a reference image with pre-existing
georeferencing data. In the radiometric correction pro-
cedure we propose here, the reference image is used to
provide baseline radiometric values which can be used to
normalise a set of sensed images.
The procedure is conceptually simple. First, both
sensed and reference images must be classified in order to
group pixels into vegetation, channel and dry classes. Sec-
ond, the sensed images must be georeferenced in order
to establish their position within the reference image.
Third, for each sensed image, pixel histograms are com-
pared separately for each class. Finally, by using histogram
matching, the histograms from the reference image are
used as target histograms and thus the histograms for the
sensed image classes are adjusted to those of the reference
image. To perform the matching, a process is realised not
on the RGB channel but LAB mode (L: luminance (%),
AB: colour range green/red and blue yellow). Given that
the reference image is a single image, all the sensed images
will be adjusted to the radiometry of the reference image.
Figures 8.9 and 8.10 give some sample results for the
same data shown in Figure 8.8. The figure shows that this
simple method has successfully normalised the brightness
values in the three images which now appear much more
continuous.
8.3.2 Radiometricnormalisation
The large hyperspatial image databases which can cause
georeferencing issues often suffer from non-uniform
illumination conditions. When hundreds or even thou-
sands of images are acquired, maintaining a uniform level
of illumination for the entire dataset is nearly impossi-
ble at the time of acquisition. First, over the timespan
required to collect the imagery, weather conditions can
change. Second, if camera settings are not adjusted prop-
erly, differences in exposure and/or aperture can lead to
very significant changes in brightness within the dataset,
even for images acquired within 1 second of each other.
This factor is further complicated by the use of small
commercial cameras designed for the mass market. These
camera designs often automate image acquisition param-
eters beyond user control. Unfortunately, since they were
not designed for aerial use, the automation parameters
are often sub-optimal in the case of airborne acquisition.
The resulting lack of uniformity in the radiometric levels
of the imagery can have a detrimental impact on the
data quality of the information derived from processes
such as depth and grain size mapping (see Chapter 9).
Whilst there are many standardised approaches to cor-
rect for illumination effects developed mainly for satellite
imagery (Dai et al., 2010), such approaches typically give
poor results for river imagery owing to the complexity
and spatial variability of riverine environments. With few
solutions present in the current literature, we propose
8.3.3 Shadowcorrection
A similar approach can also be applied to the correc-
tion of illumination problems within a single image. In
particular, shadows within the channel can pose a signif-
icant problem. One crucial application of river imagery
which is gaining in importance is the mapping of channel
bathymetry from image data. Depth can be extracted from
images of clear water rivers either through photogram-
metric analysis (Beal et al., 1997; Lane, 2000; Butler et al.,
2002; Ballester et al., 2003) or, more commonly, finding
a relationship between imaged river brightness and water
depth (Lyon et al., 1992; Garguet Duport et al., 1995;
Winterbottom and Gilvear, 1997; Gilvear, 2004; Legleiter
and Goodchild, 2005; Carbonneau et al., 2006; Lejot et al.,
2007). As such, shadows are one of the largest obstacles
to the applicability image based bathymetry as they rad-
ically change the relationship between image radiance
and channel depth. This shadow problem is most acute
during times of the year and/or day when the sun angle
Search WWH ::




Custom Search