Geography Reference
In-Depth Information
these skills, but most are not familiar with the three-
dimensional river environment. This knowledge gulf has
been a problem for the development of the science of the
remote sensing of rivers. Finding technically competent
hyperspectral operators is thus an important consider-
ation when choosing to use hyperspectral imagery in a
river project.
Because imaging spectrometers break up incoming
light into a large number of channels, the number of
photons coming into each channel is much smaller than
would be entering into a multispectral instrument in the
same situation. While hyperspectral instruments often
have a high signal to noise ratio (to try to compen-
sate for this low number of photons), the main effect
has been to require the instrument designers to pro-
vide larger pixel sizes. The spatial (ground) resolution
of airborne hyperspectral imagers is therefore often sig-
nificantly larger than typical high-quality photographs.
Centimeter-to-decimeter resolution digital imaging is
now becoming quite straightforward for normal aerial
photography. Hyperspectral imagery is typically meter-
resolution or larger. The lower number of photons per
channel can also reduce themaximumdetectable depth in
shallow clearwater environments (Legleiter et al., 2009),
but this effect can be offset by using algorithms that use
several channels.
The primary uses of hyperspectral imagery in the river
environment to date have been (a) classification of river
habitats (Marcus et al., 2003), (b) production of water
depth maps (Legleiter and Roberts, 2005), (c) water qual-
ity/suspended sediment concentration mapping (Karaska
et al., 2004), and (d) riparian cover and submerged veg-
etation identification (Williams et al., 2003). In the past
decade, the increase in the spatial resolution of aerial
photography (to decimeter or centimeter scales) has
allowed this branch of river remote sensing to com-
pete with hyperspectral imaging in habitat mapping, by
substituting high-resolution textural information for the
increased number of channels in hyperspectral imagery.
As one example, larger grain size mapping is now typically
done using textural approaches on high-resolution pho-
tography rather than spectral analysis of hyperspectral
data. Also, LiDAR developments in the past decade have
included the development of water penetrating LiDAR
for depth mapping. Unfortunately, it is not totally clear
which instruments are currently best for mapping var-
ious fluvial environments in a given situation; logistics
may provide a better reason for choosing one technol-
ogy over another at the present time. In principle, the
fusion of hyperspectral and LiDAR approaches might be
very advantageous to fluvial research and management.
At present, however, the cost of such fusions has negated
their joint use in all but a tiny number of studies.
Many hyperspectral instruments operate in the whiskb-
room (cross-track) configuration; therefore an image
pixel's data is collected a tiny moment later than the pixel
next to it. As an airplane ismoving during thatmomentary
instant, the individual pixels are usually not in a standard
uniform rectangular array (with respect to the ground) as
are most air photos. In the case of most extensive aircraft
roll, pitch, yaw, and vibration, pixels may not line up well
(Figure 4.3). Such imagery also needs to be orthorectified
(just like aerial photography) to correct for distortions
such as topography and cross-track distance distortion.
The issue also exists with pushbroom scanning systems,
though to a slightly lesser degree. Rectification, therefore,
is a significant issue the project planners may need to
examine. More recent innovations have automated some
of these procedures, though these automations are not
yet standardised throughout the industry.
Because hyperspectral imagery is, by design, divided up
into a large number of narrow spectral channels, some of
these channels can be very sensitive to non-river effects,
such as atmospheric water vapour. Other channels may
be sensitive to imaging geometry, creating considerable
illumination differences from the middle to edge of a
flight line. Standard digital image processing techniques
(such as band ratioing) may alleviate some of these effects,
but more severe issues may require advanced techniques
such as physically-based optical approaches. While there
are somewhat standardised methods for atmospherically
normalising ('atmospheric correction') many multispec-
tral sensors (such as Landsat), these approaches were
often not designed for the spatial and spectral sensitivity
of the hyperspectral scanner. Atmospheric optical mod-
els such as MODTRAN and FLAASH allow automated
atmospheric correction, though they rely on (often rea-
sonable) assumptions about the atmospheric structure
and composition that may not be precisely known for the
time and location of study. The best approach, though
expensive, to such correction is tomake field spectroscopy
measurements at various sites on the ground at the time of
the flight. These can then be used to directly calibrate the
various channels of the hyperspectral imagery. Field spec-
trometers are now fairly common pieces of equipment in
remote sensing.
Another counterintuitive issue in using hyperspec-
tral imaging is the issue of ground validation. Typically,
remote sensing analyses, such as classification, is validated
by going to various sites on the ground and recording
Search WWH ::




Custom Search