International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
able to produce results with high quality and
confidence, especially when dealing with detection
and mapping tasks down to the species level (Chen
et al., 2003).
the availability of hyperspectral and
high resolution satellite data provides researchers
an opportunity to pursue more complex analysis.
Hyperspectral imagery consists of tens to hundreds
of contiguous spectral bands therefore can provide
Fortunately,
more complete coverage of spectral information
about targets. Previous studies have demonstrated
the possibility to perform species- level vegetation
classifications using hyperspectral data (Cochrance,
2000; Laba et al, 2003; Schmidt & Skidmore, 2003).
However, currently available hyperspectral satellite
data have a limited ground resolution. This may
create difficulties in producing precise mappings of
vegetation types. On the other hand, although
limited to single or few spectral bands, high
resolution images provide detail spatial information
about the target areas. Therefore, a combination of
both types of data is likely to be an optimal
approach to the mapping of nonnative plants.
In consequence, this research undertakes an effort
to develop a systematic method to identify an
invasive plant (Leucaena Leucocephala) in the
Kenting National Park and vicinity of southern
Taiwan using hyperspectral images and high
resolution satellite images. Verification with ground
truth samples indicates that the developed method
of combining high resolution and hyperspectral
images is an effective and efficient approach to
detect invasive plants in a large area.
2 METHODS AND MATERIALS
2.1 Analysis Procedure
The method developed in this study consists of
several analysis phases. First, candidate areas of
interest where target plant (Leucaena Leucocephala)
is most likely to be spotted are selected from the
study area. The selection is done using NDVI and
other vegetation indexing schemes from
multispectral images or sub- images generated from
hyperspectral images. In other words, regions with
low possibility of vegetation distribution are filtered
out. This is a practical necessity to reduce the data
volume for analysis and to avoid negative impacts
stemming from uninterested land- cover types.
The second phase of analysis is to perform spectral
analysis on hyperspectral images to identify
Leucaena Leucocephala from vegetation covered
areas. The process begins with extracting helpful
spectral features from training data and field-
collected spectra of the target plant. The purpose of
this process is to collect spectral features that are
most helpful in discriminating Leucaena
Leucocephala from other vegetation types. Then, a
parallelpiped classification is applied to the
hyperspectral images according to the selected
features. One thing to note is that some of the
features may be difficult, if not impossible, to obtain
72
from original data. Therefore, methods developed to
identify subtle features in transformed spectral
domains should be used to achieve better analysis
performance. For example, the derivative spectral
analysis (Tsai & Philpot, 1998; Tsai & Philpot, 2002)
has been proved to be effective and efficient for this
purpose. Also, in a previous study (Underwood et
al.; 2003) researchers concluded that Minimum
Noise Fraction (MNF) analysis method can usually
produce better classification results when
identifying specific vegetation types with
hyperspectral imagery. Accordingly, this research
also applies MNF transformation before
classification process.
The preliminary classification result is then quickly
evaluated with known ground truth samples to
identify regions with poor discrimination of target
plant. In these spots, texture analysis of high
resolution images will be used to further examine
the canopy structures and other features in order to
increase the accuracy of target mapping. As of the
time this paper is being formatted, the
computerized texture analysis for this purpose is
still under development. Therefore, the analysis
results of high resolution images shown in this
paper were produced from interactive texture
analysis by experienced human interpretors.
2.2 Study Area
As shown in Fig. 1, the study area of this research is
the Kenting National Park and vicinity located in
southern Taiwan. The park was established in 1984
for the purpose of preserving natural resources and
ecosystems. It covers an area of more than 33,000
hectors of land (about 18,000 hectors) and ocean
with diversified coastal zones, terrains, land- cover
types and wildlife. The vegetation cover in this
region consists of a rich and fertile variety of native
and exotic species.
120°45'E
120*50'E
z S
N of
N Z
N =
= n
8 =
ui Foi
= e
e z
Figure 1. Study area (Kenting National Park and
vicinity).
Internati
EME LICE
2.3 |
The pri:
three h
Hyperio
resoluti
channel
infrared
nomina
image
radiom«
only !'
overlap
sensor,
image.
(mostly
for ana
noise 1
geomet
ground
As men
images
target
hypersj
results.
used ir
and mc
continu
resolut
resolut
specific
found
repeate
orthore
study :
In add
collecte
Optics
are he
from €
layers
other
study.
(orthor
selecti
used a
3.1
After
remove
to the
ground
errors
than h
appliec
selecte
helped
field
polygo
Leucae
pixels
