International Archives of the Photo
rellectance mode (each pixel of the image is presented by
a value of reflectance). At the end of this study, resulting
reflectance spectra should be identify to a various objects
and terrestrial materials, for that, ENVI 3.1 has a ^
spectral Libraries ". These libraries contain a whole of
various experimental calculated spectra reflectance,
representing a various materials terrestrial: minerals,
vegetation, rocks, water..Once the images has been
opened in ENVI 3.1, analysis techniques can then be
applied to the data. These techniques will be explained in
detail below. They consist of minimum noise fraction
transformation, pixel purity index, and n-dimensional
visualization. Overall, these analyses eliminate noise,
select pure pixels, create endmembers, visualize the
endmembers, and compare endmembers to various
spectral libraries.
3.1. Minimum Noise Fraction Transformation
The minimum noise fraction (MNF) transformation is
used to determine the inherent dimensionality of image
data, to segregate noise in the data, and to reduce the
computational requirements for subsequent processing
(See Boardman and Kruse, 1994). This step is often
completed as a precursor to other types of analysis.
Basically it is a way of simplifying the data. The MNF
transform is essentially two principal component
transformations. The first transformation, based on an
estimated noise covariance matrix, decorrelates and
rescales the noise in the data. This first step results in
transformed data in which the noise has unit variance and
no band-to-band correlations. The second step is a
standard principal components transformation which
creates several new bands containing the majority of the
information. For the purposes of further spectral
processing, the inherent dimensionality of the data is
determined by examination of the final eigenvalues and
the associated images. The data space can be divided into
two parts: one part associated with large eigenvalues and
coherent eigenimages, and a complementary part with
near-unity eigenvalues and noise-dominated images. By
using only the coherent portions, the noise is separated
from the data, thus improving spectral processing results.
Once applying MNF technique, on the 6 bands images
TM (which must be calibrated in reflectance mode), we
will have like result 6 new bands images MNF. The
image pixels are presented by eigenvalues. For the
purposes of further spectral processing, the
dimensionality of the data is determined by examination
of these values. In examining the eigenvalues it can be
seen that the first MNF bands ( 1 and 2) have the highest
values while the remaining bands have consistent low
values. It is the first two bands with the large values that
contain most of the information and it is these bands that
correspond to MNF images. The remaining low value
bands (3 and under for example) are seen as noise. The
images show the information compressed into only a few
bands. The redundancy of the data is eliminated and noise
is also removed. The result are more interpretable images.
You could say that the data has been simplified or the
dimensionality has been reduced. After the data has been
transformed, it is possible to examine scatterplots
between each fraction image in order to understand the
images better. Figure.2 shows scatterplots between the
first three bands :
grammetry, Remote Sensing and Spatial Information Sciences, \
1072
LEE ren orme
Ee e
pem
x: LIEU AUR x]
§ FV Dyers |
Band 2 .vs. Band 3
Figure.2 Scatterplots of MNF bands
—
The scatterplots plot the data of different MNF bands
against each other. This gives an ideas of the spectral data
distribution. Of special interest are the arms or separate
clusters outside of the main data cloud. These areas
represent unique spectra called "endmembers" which
represent unique ground components. From the
scatterplots the user can manually highlight endmembers
and identify them on the image.
3.2. PIXEL PURITY INDEX (PPI)
All remotely sensed images contain a phenomenon
known as mix ixels. These are pixels that contain a
mix of features (e.g. grass, forest water). In multispectral
analysis it is useful to separate purer [rom more mixed
pixels in order to reduce the number of pixels to be
analyzed for endmember separation and identification.
The pixel purity index is a way of finding the most
spectrally pure pixels in images (Boardman, 1993;
Boardman & al., 1995). The PPI stipulates how many
times the pixel is extreme in the simplex (Boardman &
Kruse, 1994; Boardman ef al., 1995). The most spectrally
pure pixels typically correspond to spectrally unique
materials. The PPI procedure generates an image where
pixel values correspond to the number of times that this
pixel was recorded as extreme. This way, threshold of the
PPI image can stipulate the most extreme pixel results in
further spatial reduction. In this paper the PPI was
performed using the six bands MNF. The PPI was
processed with 100 iterations. The PPI image was
generated using the threshold of 3. Brighter pixels
represent more spectrally extreme “hits” and indicate
pixels that are more spectrally pure. Darker pixels are less
pure. The value of the pixel corresponds to the number of
times it was recorded as extreme during the PPI process.
se
ed
3.3. N-DIMENSIONAL VISUALIZATION
The n -Dimensional Visualizer is an interactive tool that
allows the user to select endmembers in n-space. This
procedure generates clouds of points related to the pixels
in a n-dimensional space defined by the MNF
components. Once we have chosen the bands to display,
the visualizer can then go to work, it rotates interactively
the data cloud. The user is able to see the spectral data in
many dimensions from many angles. From here one can
interactively select endmembers. From different angles
we are able to visualize and highlight (in different
colours) the distinct spectra which appears as an
endmember (Figure.3)
^ol XXXV, Part B4. Istanbul 2004
Int
&
ol
In
i
