Geography Reference
In-Depth Information
Several researchers have developed alternative
approaches to supervised classification. Maruca and
Jacquez (2002) achieved good results using a spatially
agglomerative cluster technique, and Goovaerts (2002)
achieved high accuracies using a geostatistical co-kriging
approach (Table 2.3). Legleiter and Goodchild (2005)
examined the potential of a 'fuzzy' approach to stream
classification that provides a more realistic representation
of the gradual transitions between similar habitat
units. Their work revealed far more complexity in
the stream patterns than is captured in the simple
either/or dichotomy of biotype classification. None of
these approaches, however, are presently available in
existing software packages. Special programming skills
are therefore needed to implement them, making them
less accessible to the management community.
Regardless of the approach, no physical models yet exist
that allow remote mapping of biotypes in the absence of
field data. Because the spatial pattern of biotypes can
change with variations in discharge, personnel must be in
the field at or near the time of image acquisition to map
biotypes. The field maps are then used to select pixels
to train the classification algorithms. Given that field
mapping of biotypes is necessary regardless of mapping
approach, using remote sensing only makes sense if the
spatial extent of mapping far exceeds what a survey team
can readily achieve while in the field.
In contrast, mapping wood, at least against a background
of water or sediment, requires detecting the difference
between features with distinctly different compositions.
Automated mapping of wood in and along rivers can, in
fact, be relatively simple if one has hyperspectral imagery
or fine resolution, high quality colour imagery Marcus
et al. (2003) and Smikrud and Prakash (2006) both used
similar approaches to detect wood with relatively high
accuracies (Table 2.3). Both sets of researchers first cal-
culated principal components from their hyperspectral
imagery to isolate the spectral signatures of different fea-
tures. They then applied a matched filtering technique
to the principal component images to detect wood. The
matched filter operates by using some wood pixels within
the image to train an algorithm that then finds similar
pixels elsewhere in the image. A major advantage of the
matched filtering technique is that there is no need for
field teams if some wood can be seen clearly on the
imagery. Principal component transformations are avail-
able in all remote sensing software and matched filtering
is included in an increasing number of these packages.
If hyperspectral imagery is available, the matched filter
will almost certainly detect wood in areas where it cannot
be seen with the naked eye on the same image. In this case,
the algorithm is 'unmixing' the pixel to detect locations
where wood makes up only a portion of the pixel. This
can lead to confusion on the part of users, who may
believe the classification is showing wood where it does
not exist. In addition, finding pieces of wood smaller than
a pixel may cause problems if the user is seeking to map
wood of a given size or larger. For example, many fluvial
wood studies only seek to document wood that is 1 or 2 m
in length and at least 10 cm in diameter - large enough
to affect the flow. Hyperspectral imagery may detect
much smaller wood, making it difficult to determine
which pieces are of sufficient size to alter flow dynamics.
Pixel unmixing approaches or shape detection algorithms
might provide a solution to this issue, but no research has
yet been attempted along these lines.
2.9 Wood
Wood in rivers plays a major role in forming and
altering stream habitats in many streams. Wood can force
sediment deposition and transport, alter in-stream habi-
tats and stream morphology, provide organic debris for
macroinvertebrates, and create shelter for fish. Emplac-
ing wood is a major tool used in stream restoration.
Promoting wood accumulation is often a central goal
of riparian management strategies (e.g. Fox and Bolton,
2007; Jochem et al., 2007).
In theory, remote detection of wood in river channels
or on exposed bars should be relatively straightforward.
At its simplest level, if sufficiently fine resolution imagery
is available, one can see the wood on the images and map
it manually. Even from an automated perspective, detec-
tion of wood should be relatively simple. Mapping depth,
turbidity, channel change, and biotypes all require detect-
ing variations, often subtle, in one feature type (water).
2.10 Submerged aquatic vegetation
(SAV) and algae
SAV and algae are important to ecosystem health, provid-
ing food and cover for a wide range of species, removing
toxins from water and sediments, and stabilising stream
beds. On the other hand, an over abundance of SAV
Search WWH ::
Custom Search