Geography Reference
In-Depth Information
classification as detailed above is much easier. Notably
water areas and vegetation are well identified because of
the absorption in infrared band of the water contained
in the leaves (Jones et al., 2010). Better results have been
obtained from combined near and mean IR channel from
Spot 5 XS images and orthophotos (Tormos et al., 2009)
and from IR colour orthophotos (Wiederkehr et al.,
2010b). Tests performed on the Dr ome River were based
on an object-oriented method. Specific ratios were used
for detecting water and vegetation patches. For water we
used an average value of the infrared band below 80 and
a maximum of the red band between 160 and 200 over a
0-256 scale. For vegetation we used the ratio of infrared
band to the red band with a threshold between -0.7 and
0.1 to isolate the vegetation patches. Once these land-use
classes were identified, the remaining polygons were clas-
sified as gravel bars. Detection rates for each class reached
almost 100%. Note also that useful information can be
developed from a non-automated analysis, and a com-
parison of the two approaches (visual versus automatic)
must be done to evaluate which solution is the most
effective.
The second problem is associated with errors in the
imagery. Different ones have been identified all along
the process dealing with georectification, photo-inter-
pretation, and also water level measurement. Archived
hard copies (photographic film or paper reproductions)
need to be digitised and orthorectified. Old photos often
have strong distortions that lead to significant rectifi-
cation error. To prevent such problems, clear protocols
must be defined to minimise the georectification error
and the operator bias, notably because the errors accu-
mulate when overlying different layers. This problem is
critical when considering the channel shifting from chan-
nel overlays and associated sediment budgeting because
image errors can be multiplied by errors associated with
additional required field measurements (e.g., overbank
fine sediment depth estimate, sometimes bank height
estimate). Several authors have developed methods for
estimating errors in quantifying channel migration rates
that can be applied to GIS-based analysis of planform
change (Mount and Louis, 2005; Hughes et al., 2006)
and for estimating error in bankfull width comparisons
from temporally sequenced raw and corrected aerial pho-
tographs (Mount et al., 2003). The effect of scanning
resolution has been explored by Liebault (2003), and
Toone (2009) considered the observer bias in the draw-
ing of channel polygons over a range of photos whose
scale varies from 1:30,000 to 1:17,000. She showed stan-
dard error in width of 0.5 to 0.8 m (3.2 to 4.7% of
channel width) from repeat digitisations by four different
operators over a set of 90 channel cross-sections.
Following Liebault (2003) who performed a precision
test of the measured area, it is best to scan the archived
photographs at a resolution of 700 to 1000 DPI, depend-
ing on their scale, to provide a final resolution of ca
0.5 m, similar to that of modern orthophotographs. In his
case, the active channel surface area was slightly under-
estimated as resolution increased until a threshold was
reached above 700 DPI. The threshold did not depend
on the historical series used. After being georectified
primarily using first order polynomial transformations,
average RMS error was on the order of 2 to 3 m for each
photograph.
Interpretation of features and, in the case of manual
treatments, the associated drawing process, is also a major
source of error. Scale variation between different photo
series or light variation in a single photo series may affect
the capacity of even a human observer for detecting fea-
tures. For these reasons, it is sometimes worthwhile to
advocate identifying feature boundaries by hand rather
than using an image classification tool even if the manual
approach can be particularly time-consuming. The oper-
ator has to choose a single set of rules regarding level of
detail that is maintained throughout the drawing session.
Particular care should be taken to work at a consistent
photograph scale in order to minimise errors at this step
(although it is appropriate to zoom in and out from time
to time to help in interpretation). When considering his-
toric change, it is important to realise that riverscape fea-
tures are neither spatially nor temporally uniform and that
this can influence feature detection. For instance, when a
riparian forest is established along a channel, the canopy
partly covers the channel. If the phenomenon is system-
atic in space and time, the channel width/area detection
may be meaningful (at least in a relative sense), but if the
canopy closes over time, it can provide both alignment
error and classification error when overlaying different
states. Comparing channel width calculation from photos
and from field measures, Liebault (2005) found differ-
ences of 2.39 m (18% in average) between the two, the
measures on photo usually underestimating the real width
because of the canopy cover. This error also depends on
the scale, and increases when resolution becomes coarser.
Therefore, such analysis always needs to be checked for
precision. There is not 'right' or 'wrong ' way to digitise,
but the digitisation is not particularly helpful unless it
is checked by repeating the methods on other similar
photos or having another person independently repeat
the procedure on the single photograph in question.
Search WWH ::
Custom Search