Biology Reference
In-Depth Information
only pertains to showy, open-grown plant species. Cryptic, rare, or understory
species are more difficult to monitor with remote sensing. The usage of these
techniques is very species specific, and image acquisition must be timed with
phenological characteristics of the target species in relation to surrounding vege-
tation so that it can be distinguished from other species. For example, detection
of spotted knapweed (
Centaurea maculosa
) and babysbreath (
Gypsophila panic-
ulata
) in Idaho require the understory vegetation to be fully matured and
bleached, and spotted knapweed must retain some of a previous year's stems for
detection Lass et al. (2005).
Sensor technology and detection algorithms are improving, but this technique
is still limited to certain species. Pure pixels of the invasive species, meaning an
area on the ground corresponding to a grid cell in the image that is completely
covered in the species, are often required as training data to classify it in an image.
For many species it may be difficult to find pixels that are unmixed (e.g., some
spotted knapweed and some sagebrush rather than only spotted knapweed). It is
much more difficult to detect understory species in a dense forest or submerged
aquatic vegetation than a species on the prairie. For example, Underwood et al.
(2006) mapped two aquatic species, Brazilian waterweed (
Egeria densa
), which is
a submerged aquatic species, and water hyacinth (
Eichhornia crassipes
), an emer-
gent, floating species at two different spatial extents. At the local scale (average
size 51 ha), average accuracy was highest for Brazilian pepperweed (93% com-
pared with 73%), but dropped drastically when the extent was increased to
177,000 ha (29%) while the accuracy for water hyacinth decreased only slightly
(65%). The cost of developing these classifications for each individual species
renders developing a different model for every 51 ha unreasonable in most cases.
As with statistical techniques, there are several different algorithms that can be
used to classify a remotely-sensed image. Several different papers compare meth-
ods and promote some as better than others (Elith et al. 2006; Higgins and
Richardson 1996; Hirzel et al. 2001), but as with statistical techniques, different
methodologies probably work better for different species in different localities.
How small can an infestation be and still be detected? We do have time series
data, but are we able to tell when a species first appears in an area, or does it have
to reach a threshold level of abundance? If it does, is that level of infestation large
enough so that eradication at the new location would be impossible or incredibly
expensive? These questions still await answers. While it is true that covering large
area such as a state or province for field data collection over multiple years would
be difficult, remotely-sensed imagery cannot supply all the necessary information
to answer questions for management. For example, it would be difficult to use
products where cheatgrass, a small annual grass, was modeled at 30-60 m resolu-
tion (Bradley and Mustard 2006) for early detection of new infestations of cheat-
grass in central Nevada at a local level.
Detecting species at low levels of abundance even over small areas with high
resolution data is still difficult. Mundt et al. (2005) classified hoary cress (
Cardaria
draba
) in 3 m spatial resolution hyperspectral images for an area 1.75 by 22 km in
