International Archives of the Photogrammetry, Remote Sensing a
nd Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
In our previous study using the AISA sensor, the classification
process was performed using 25 bands. The Airborne
Visible/Infrared Imaging Spectrometer (AVIRIS) sensor allows
more specific spectral analysis with wavelengths from 380 nm
to 2500 nm. AVIRIS data with spectral radiance in 224
contiguous spectral channels, spectral resolution of about 10
nm, and a ground pixel size of 18m x 18m were available to
this study. The 2001 AVIRIS scene of the Santa Barbara
coastal zone was obtained from the NASA Jet Propulsion
Laboratory (JPL) in California. The whole track scene consists
of 13 segments; each segment consists of 512 x 614 pixels.
Four scenes were selected, each one representing an oil spill at
a different stage from north to south. AVIRIS data allow the
study of different types of oil spills in the ocean, oil spill
classification, and quantitative measurements for on apparent
oil spill.
2. METHODOLOGY AND RESULTS
In this research study, hyperspectral processing techniques are
applied to an AVIRIS image using a total of 224 bands from
0.374 um to 2.50 pm. This process involved several steps
beginning with looking at the change in spectral signatures in
both space and time, as the oil spill characteristics changed
with time. Three segments were selected to detect the different
oil slick shapes. Data browsing was used to look at spatial and
spectral changes in reflectance.
2.1 Model Description
In this paper an attempt was made to use newer and simpler
techniques for target identification building on the PU
technique. This PU method is used for partially unmixed
AVIRIS data. High quality data of spectrally complex areas are
highly dimensional and are consequently difficult to fully
unravel (Boardman, et al., 1995). PU technique provides a
method of solving only that fraction of the data inversion
problem that directly relates to the specific goals of the
investigation. In our previous study, the spectral linear
unmixing technique showed many limitations for operational
applications, because signatures for all the target materials in
the scene must be spectrally identified. The PU method does
not requires a prior knowledge of the background material
spectral signatures. Our investigation included only oil spill
targets that were not mixed with other materials except water.
The data are subjected to dimensional reduction processes using
Minimum Noise Fraction (MNF) technique for reducing high
dimensionality of the data.
2.2 Data Reduction
The data reduction phase of the analysis strives to identify
Regions of interest (ROIs) in the data set by separating noise
from information and reducing the data set to its true
dimensionality by applying the minimum noise transform, then
determining spectrally pure (extreme) pixels using the Pixel
Purity Index (PPI) function on the minimum noise fraction
results (Wagtendonk, et. al, 2000). N-dimensional
visualization can then be performed using the grand tour of the
high PPI value pixels to cluster the purest pixels into image-
derived ROIs.
AVIRIS data were put through a dimensionality analysis using
the MNF transform. The data are translated to have a zero
Inti
mean, and then the data are rotated and scaled so that the noise pU
in every band is uncorrelated and has unit variance. The fes
inherent dimensionality of the data is determined by examining ru
the final eigenvalues and the associated images. The data space Ee
is divided into two parts: one part associated with near-unity t
eigenvalues and coherent eigen-images, and a complementary .
part with near-unity eigenvalues and noise-dominated images. Dix
By using only the coherent portions, the noise is separated from col
the data, which improves the spectral processing results. Image RES
linking and overlaying are used to identify pixel locations for
thick oil slicks to selected oil spill spectra. A target weighting 23
variance for high eigenvalues for each class provides clues to 5
spectral features, contributing to the classification.
2.3 Target Identification Using Partial Unmixing
In our study another experiment was performed using the MNF
(principal component) technique for target identification using
data reduction and PU. The results indicated good separation
of oil slicks, oiled water, and polluted water. Results are here
used as a preliminary identification of oil targets, since no
ground truth data were available. PU allows mapping of the
apparent target abundances in the presence of an arbitrary and
unknown spectrally mixed background.
We have developed a partial unmixing technique as a method
of mapping target ROIs using two MNF bands in a 2-D scatter
plot to separate the oil spill target from the background pixels
and to produce ROIs. The purest pixels in the scene are then
compared against the target spectra. If any are close matches
for the target materials, they are identified and separated from
the other purest pixels. High-purity pixels that do not closely
match a target spectrum are used to determine the subspace
background. This analysis does not require a prior knowledge
of the background material signatures. The steps are indicated: In F
(visi
The:
2.3.1 Eigenvector Weighting max
sepa
brov
MNF Band2 Eigenvector Weightings clus:
| El {Thick dick yellc
10 3 |
gu 2
3s 9 Wr pir
Sen | ;
Do p
Gi 7
-10 E
-15 7
bo
Cn
=
i
1 i A 1 1 1 1 1 1 1 L A 1 1
500 $000 1500. 2000
Wavelength (400-2500 nm)
Figure 7. The MNF Spectrum For Thick Oil Slick Bands
Shows High Eigenvalues.
Band 12 (fODDBOB5tü1 POI C2 se03+1.img}aanta_s
i
Figure 7 shows a plot of the MNF eigenvalues, showing the
amount of covariance in each output MNF band. Most of the
information is derived from differences in 640 to 890 nm.
The output results of data reduction processes are 25 MNFs;
some of these bands have significant pixels for oil spills and the In Fi
betwe
1330
