Geography Reference
In-Depth Information
example, Legleiter et al., 2002). This classification advan-
tage is partly based on the number of channels for exact
types, but also because of the ability to better categorise
mixtures; a more difficult problem for multispectral
(underspecified) imagery. Most published work on this
aspect of hyperspectral data in fluvial systems has used
standard supervised classification techniques, the same
or similar to those used in undergraduate remote sensing
classes. The high number of channels, however, has raised
the possibility that classes derived from the imagery itself
(unsupervised classification) may be better at mapping
rivers than standard supervised methods for classifying
river habitats on the ground. This unsupervised approach
to mapping rivers is a form of data mining, and many
techniques developed for standard remote sensing
topics (such as self-organising maps, a type of artificial
neural network algorithm) could be imported to river
remote sensing using hyperspectral data that has been
done previously.
One of the chief advantages of hyperspectral data, as
noted already, is its ability to quantify mixtures of types.
This type of process, often called 'soft classification', can
yield a variety of new ways of measuring rivers. One
typical example of soft classification, linear unmixing,
uses pixels covered by a single pure feature (also known
as an endmember for example a 'pure' water habitat such
as a pool) as a baseline, and then analyses all other imaged
pixels in order to extract what proportions of each pixel
is made of the various habitats - in other words, it is a
subpixel mapping approach. A separate soft classification
approach treats mixtures not as combinations of pure
habitats, but rather shows how 'impure' habitats are;
this approach is known as fuzzy classification (Legleiter
and Goodchild, 2005). Soft classification approaches tend
to yield more realistic data products compared with
standard 'hard' classification approaches. They are also
more comparable to continuous-data ground validation
methods in the field.
when compared to traditional river mapping. For very
large rivers (
>
30m width minimum), it may be possi-
ble to use spaceborne hyperspectral instruments such as
NASA's Hyperion (on the experimental EO-1 satellite)
at relatively low cost. Future spaceborne hyperspectral
instruments, such as EnMap and HyspIRI, will have simi-
lar 30-50meter ground resolutions, though their lowered
repeat times and increased global coverage will add to the
number of rivers that can be examined from space using
hyperspectral imaging. The problem of pixel mixing usu-
ally means that a river must be at least a few pixels wide
in order for high quality information to be extracted;
therefore, spaceborne hyperspectral platforms do not, as
yet, allow most of the world's rivers to be imaged. For
most rivers, existing hyperspectral datasets with the nec-
essary resolution to image river environments are few,
and the cost can be high. Most projects desiring hyper-
spectral imagery generally hire outside consulting firms
to produce new hyperspectral imagery. This sometimes
can cost tens of thousands of dollars (US). Some gov-
ernmental programs can provide limited hyperspectral
imagery at lowered cost. NASAAVIRIS (Airborne Visible/
Infrared Imaging Spectrometer), for example, flies hyper-
spectral imagery in support of US government programs
at reduced prices. Additionally, the AVIRIS program has
a competitive program to provide one free flight line to a
graduate student per year, subject to the flight schedule of
that year. But these opportunities are often competitive
in nature and projects can therefore not control the tim-
ing and logistics of their flight. Buying a well-calibrated,
good quality hyperspectral scanner that is useful for river
remote sensing is in the low hundreds of thousands of
dollars (US), similar to the cost of buying a LiDAR air-
borne system. The paucity of agencies and companies that
fly contract imagery also makes the imaging difficult to
control in terms of exactly when things are flown; this in
turn can affect the quality of the imaging. A few days can
turn a clear water stream to a muddy morass. Time will
tell if the number of hyperspectral instruments flying will
increase and the costs will decrease.
Most students of river remote sensing do not have
extensive hyperspectral imaging skills. While many tech-
niques of hyperspectral digital image processing are
similar to those in the multispectral arena, some of the
best uses of hyperspectral data require advanced skills.
These include physically-based approaches for separating
various water and channel characteristics and advanced
statistical algorithms for mapping mixtures of classes
such as Spectral Angle Mapper (SAM) and linear unmix-
ing. Specially-trained remote sensing students often have
4.4 Logistical and optical limitations of
hyperspectral imagery
The single biggest downside to utilising hyperspectral
imaging in fluvial environments is its cost. Hyperspectral
work on rivers has been almost entirely based on airborne
imaging, and hired hyperspectral flights can be expen-
sive. An economy of scale exists if larger areas are to be
mapped; only part of the total costs are associated with
flying time, and larger river areas might be cost-effective
Search WWH ::
Custom Search