278
dicted Amur honeysuckle cover than a single image. Evangelista et al. (2009) used
a species distribution model to predict tamarisk (Tamarisk ramosissima) distribution over time based on vegetation indices derived from Landsat ETM+ data, with a
90% classification accuracy. Kimothi et al. (2010) used Indian RS satellite data to
map another understory IAS, the West Indian lantana (Lantana camara), using texture analysis of images from September, February, and April. The dense leaf canopy
meant separation of IAS was not possible in September, but classification accuracies were >90% in the other images after leaf fall. Barbosa et al. (2016) mapped
subcanopy strawberry guava (Psidium cattleianum) outbreak with imaging spectroscopy and lidar and tested the accuracy of a machine learner, biased-SVM
(BSVM), and mixture-tuned matched filtering (MTMF; a partial unmixing classification algorithm similar in principle to MNF) across canopy layers. While both
methods allowed the estimation of the fraction of canopy layers that were invaded,
the BSVM used information across the entire spectrum, while the MTMF did not,
which may limit the applicability of MTMF when spectra of IAS are similar to
“background” native species.
Indirect methods are another alternative to study understory IAS. Joshi et al.
(2006) mapped Siam weed (Chromolaena odorata) in the understory using Landsat
ETM+ and an artificial neural network to predict forest density and canopy light
penetration and then subsequently predict Siam weed seed production. They found
that 93% of the IAS seed production was predicted by the light intensity reaching
the understory and concluded that this method worked relatively well to detect the
IAS, despite the spatial resolution limiting detection to well-established IAS
patches.
In summary, the most common method to detect tree IAS and map their distribution are to use their characteristic spectral signatures and dissimilarity with that of
the native vegetation (Lass and Prather 2004). Tree IAS likely affect both the forest’s spatial structure as reflected in texture metrics (Pearlstine et al. 2005) and its
3-D structure, as shown with lidar (Asner et al. 2008a, b). To maximize the ability
to detect invasive tree species, the use of the full visible (VIS) to shortwave infrared
(SWIR) spectrum with imaging spectroscopy has shown clear advantages (Martin,
Chap. 5), for example, in detecting the fire tree (Asner et al. 2008a, b) and for detecting bamboo (Dendrocalamus sp.) and slash pine (Pinus elliottii; Amaral et al. 2015).
Alternatively, other studies selected specific bands that maximized discrimination
and eliminated potential noise from nondiscriminating parts of the spectrum
(Boschetti et al. 2007). While the advantages of imaging spectroscopy are obvious,
data are not yet readily available to detect and map many tree IAS, especially in
early stage invasion stages, although the upcoming launch of several hyperspectral
satellite sensors will soon change this. Many tree IAS have different phenology than
the native forest, either staying green longer, greening earlier, or flowering or budding later (Landmann et al. 2015); or they may be evergreen in a deciduous forest
(Diao and Wang 2016). Timing imagery acquisition to maximize phenological differences has resulted in good classification accuracy (Ramsey III et al. 2002).
Finally, using pixel sizes that match a tree canopy allows the detection of single
E. A. Bolch et al.
dicted Amur honeysuckle cover than a single image. Evangelista et al. (2009) used
a species distribution model to predict tamarisk (Tamarisk ramosissima) distribution over time based on vegetation indices derived from Landsat ETM+ data, with a
90% classification accuracy. Kimothi et al. (2010) used Indian RS satellite data to
map another understory IAS, the West Indian lantana (Lantana camara), using texture analysis of images from September, February, and April. The dense leaf canopy
meant separation of IAS was not possible in September, but classification accuracies were >90% in the other images after leaf fall. Barbosa et al. (2016) mapped
subcanopy strawberry guava (Psidium cattleianum) outbreak with imaging spectroscopy and lidar and tested the accuracy of a machine learner, biased-SVM
(BSVM), and mixture-tuned matched filtering (MTMF; a partial unmixing classification algorithm similar in principle to MNF) across canopy layers. While both
methods allowed the estimation of the fraction of canopy layers that were invaded,
the BSVM used information across the entire spectrum, while the MTMF did not,
which may limit the applicability of MTMF when spectra of IAS are similar to
“background” native species.
Indirect methods are another alternative to study understory IAS. Joshi et al.
(2006) mapped Siam weed (Chromolaena odorata) in the understory using Landsat
ETM+ and an artificial neural network to predict forest density and canopy light
penetration and then subsequently predict Siam weed seed production. They found
that 93% of the IAS seed production was predicted by the light intensity reaching
the understory and concluded that this method worked relatively well to detect the
IAS, despite the spatial resolution limiting detection to well-established IAS
patches.
In summary, the most common method to detect tree IAS and map their distribution are to use their characteristic spectral signatures and dissimilarity with that of
the native vegetation (Lass and Prather 2004). Tree IAS likely affect both the forest’s spatial structure as reflected in texture metrics (Pearlstine et al. 2005) and its
3-D structure, as shown with lidar (Asner et al. 2008a, b). To maximize the ability
to detect invasive tree species, the use of the full visible (VIS) to shortwave infrared
(SWIR) spectrum with imaging spectroscopy has shown clear advantages (Martin,
Chap. 5), for example, in detecting the fire tree (Asner et al. 2008a, b) and for detecting bamboo (Dendrocalamus sp.) and slash pine (Pinus elliottii; Amaral et al. 2015).
Alternatively, other studies selected specific bands that maximized discrimination
and eliminated potential noise from nondiscriminating parts of the spectrum
(Boschetti et al. 2007). While the advantages of imaging spectroscopy are obvious,
data are not yet readily available to detect and map many tree IAS, especially in
early stage invasion stages, although the upcoming launch of several hyperspectral
satellite sensors will soon change this. Many tree IAS have different phenology than
the native forest, either staying green longer, greening earlier, or flowering or budding later (Landmann et al. 2015); or they may be evergreen in a deciduous forest
(Diao and Wang 2016). Timing imagery acquisition to maximize phenological differences has resulted in good classification accuracy (Ramsey III et al. 2002).
Finally, using pixel sizes that match a tree canopy allows the detection of single
E. A. Bolch et al.