Geography Reference
In-Depth Information
bands (panchromatic or red/green/blue). However, the
infra-red colour is becoming more and more frequent
over large areas so that it should provide more capacity to
detect water and humid patches in the near future at this
regional scale. In any case, when covering a large area, sev-
eral sets of photos or sources of images have to be used, and
it is not possible to avoid radiometric variations between
tiles. Furthermore, there are artifacts such as shadows,
overlighting, variations in spatial resolution and radiom-
etry, particularly when using website images such as the
one in Google Earth. These may vary from one mosaic
tile to the next even within a larger mosaic. Furthermore,
one of the problems with using sources such as Google
Earth is associated with the difference in resolution from
one regional area to another. Earth is mostly covered with
a resolution of 15 m per pixel (Landsat images) with areas
with better resolution reaching 2.5 m (Spot images), 0.5 m
(Geo-Eye) or less when airborne information is provided
by national suppliers. This may have practical conse-
quences for automating classifications of polygons, and
preliminary treatments may be needed to homogenise the
light conditions (see discussions in Chapter 8). The prob-
lem is also present when comparing historic images with
one another. Variability in lighting conditions, shadows,
and discharge can lead to large changes in the patterns
visible in the photographs (Figure 11.12).
When the classification is done automatically, prelimi-
nary tests must be done to identify the best parameters for
the classification. An object-oriented procedure allowing
incorporation of radiometric values as well as other
parameters such as feature shapes, textures and context
is becoming more and more popular for maximising the
use of information available on low spectral resolution
images such as the orthophotographs. In preliminary
tests applied to the Rh one network, in particular on
the Dr ome River, a confusion matrix provided overall
accuracy of 91.8% for gravel bars, flow channels and
vegetation patches. The results varied according to the
types of features as well as the active channel width.
Gravel bars and vegetation patches were well identified,
with respectively more than 90% and 95% of polygons
detected whereas water areas were unevenly identified.
Moreover the quality of the classification increased with
the width of the active channel. Within the downstream
part of the network, the rate of detection was around 85%
of features. This rate reached only 50% of patches within
the upstream part of the network. Parameters used to
determine the water, gravel bars and vegetation polygons
included a textural index, spectral indexes (mean radiom-
etry of Red band and Green band and standard deviation
of Blue band) and a shape index (area) (Wiederkehr et al.,
2009). However, even though different classes could be
predicted reasonably accurately within a set of photos at
a regional level, the method cannot be applied directly
at the 45,000 km network scale of the Rh one because
of significant changes in light conditions. Consequently,
several separate classifications were required at this scale.
However, note that if an infra-red band is available, basic
(a)
(b)
(c)
2002/03/04
2005/06/02
2007/04/07
(d)
(e)
2007/07/31
2008/12/31
Figure 11.12
Images from Google Earth website of the Pierre-Benite reach (Rh one River) downstream from Lyon on different dates.
(a) corresponds to a flood period with turbid water, (b) and (e) have lightness diffusion with specular reflection, (e) also shows a tile
boundary, (d) water surface is rough, due to very small waves and associated light diffusion, (c) is the only one date on which surface
water is undisturbed, lightness conditions are good, and bathymetry can be predicted from image radiometry. Images of the
Pierre-Benite reach (Rh one River) sourced from Google Earth web site.
©
2012 Google http://www.google.com/earth/index.html
Search WWH ::
Custom Search