stanbul 2004
veloped to
| spectral
r analysis
e spectral
pot, 2002)
nt for this
erwood et
Minimum
an usually
ts when
es with
research
before
en quickly
amples to
of target
| of high
r examine
in order to
As of the
ted, the
)urpose is
> analysis
n in this
eio texture
research is
located in
d in 1984
yurces and
an 33,000
and ocean
land- cover
er. in this
y of native
N.S-cc
N«22
NSS.LZ
ark and
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
2.3 Primary Materials for Analysis
The primary data used for analysis in this study are
three hyperspectral images acquired by NASA EO- I
Hyperion sensor and two sets of QuickBird high
resolution images. Hyperion sensor has 220 unique
channels covering from visible to short- wave
infrared (357 — 2576 nm) of spectrum with a
nominal ground resolution of 30m. Each Hyperion
image delivered by NASA (via USGS) is Level. |
radiomeric corrected consisting of 242 bands but
only 198 of them are calibrated. Because of an
overlap between VNIR and SWIR focal planes of the
sensor, there are only 196 unique channels in each
image. In addition, some of the calibrated bands
(mostly in water absorption regions) are excluded
for analysis in this study because of low signal- to-
noise ratio. Because the delivered images are not
geometrically corrected, they were registered using
ground control points before further analysis.
As mentioned earlier, the high resolution QuickBird
images were used to fortify the discrimination of
target plant in areas where spectral analysis of
hyperspectral images did not produce satisfactory
results. Two sets of QuickBird images have been
used in this study to improve the analysis accuracy
and more will be acquired if necessary as the project
continues. Each set of the images consists of a 60cm
resolution panchromatic image and a 2.4m spatial
resolution multispectral (4 channels) image. Detail
specifications about QuickBird products can be
found at Digital Globe's web site and are not
repeated here. The high resolution images were
orthorectified with DEM and accurate maps of the
study areca.
In addition, this study also uses in situ spectra
collected using a portable spectrometer (Ocean
Optics USB2000) to help select spectral features that
are helpful in separating — Leucaena Leucocephala
from other vegetation types. Also, several GIS data
layers obtained from the park administration and
other resources are used as reference data in this
study. They are used in pre- processing the images
(orthorectification and registration) as well as
selecting training samples. Some of them are also
used as ground truth to verify the analysis results.
3 RESULTS AND DISCUSSION
3.1 Analysis of Hyperspectral Images
After uncalibrated and low SNR bands were
removed, the hyperspectral images were registered
to the maps of the study arca using at least 10
ground control points for each scene. The RMS
errors of the registration were controlled to be less
than half of a pixel. Before any spectral analysis was
applied to the images, a set of training pixels were
selected from the images. The selection process was
helped by a vector layer generated from a previous
field investigation. The GIS layer consists of
polygons with 4 levels of population density of
Leucaena Leucocephala. Four hundred training
pixels were randomly selected from polygons that
73
were labelled as having more than 50% of target
plant. Part of the training samples are shown as
green dots in Fig. 2.
A
Figure 2. Training data (green dots).
A collection of "helpful" features that are capable of
effectively discriminating target plant from
surrounding vegetation clusters were extracted from
the training data and field- collected spectra for
subsequent analysis of the hyperspectral images.
As indicated earlier, the MNF operation is believed
to be a better base for species- level discriminaion.
Therefore, the images were transformed using built-
in MNF algorithm. Fig. 3 and Fig. 4 demonstrate two
of the MNF bands from one of the three Hyperion
images. Then, sub- images of selected MNF bands
were analyzed using parallel- pipe classifiers
according to selected spectral features.
Figure 3.Second MNF band of a hyperspectral image.
