Lndmember analysls

Llnear and non Llnear unmlxlng

Pure pixel based approach
Minimum volume based approach
Statistical approach
The pixel purity index (PP) algorithm uses MNF as a preprocessing step to
reduce dimensionality and to improve the SNR. PP projects every spectral
vector onto skewers (large number of random vectors).
The points corresponding to extremes, for each skewer direction, are stored.
A cumulative account records the number of times each pixel (i.e., a given
spectral vector) is found to be an extreme. The pixels with
the highest scores are the purest ones.
N-FNDR is based on the fact that in spectral dimensions, the volume
defned by a simplex formed by the purest pixels is larger than any other
volume defned by any other combination of pixels. This algorithm fnds the
set of pixels defning the largest volume by infating a simplex inside the data.
ATGP starts by selecting the pixel vector with maximum length in the scene
as the first endmember. Then, it looks for the pixel vector with the maximum
absolute projection in the space orthogonal to the space linearly spanned by
the initial pixel, and labels that pixel as the second endmember. A third
endmember is found by applying an orthogonal subspace projector to the
original image where the signature that has the maximum orthogonal
projection in the space orthogonal to the space linearly spanned by the first
two endmembers.
The spatial-spectral endmember extraction tool (SSEE) uses spatial constraints to
improve the relative spectral contrast of endmember spectra that have minimal
unique spectral information, thus improving the potential for these subtle, yet potentially
important endmembers, to be selected
The SSEE algorithm searches animage with a local search window centered
around each pixel vector and comprises four steps.
First, the singular value decomposition (SVD) transform is applied to determine a set of
eigenvectors that describe most of the spectral variance in the window or partition.
Second, the entire image data are projected onto the previously extracted eigenvectors
to determine a set of candidate endmember pixels.
Then, spatial constraints are used to combine and average spectrally similar candidate
endmember pixels by testing, for each candidate pixel vector, which other pixel vectors
are sufficiently similar in spectral sense
CalculaLlon Cf lxel urlLy lndex
Skewer CeneraLlon
uaLa ro[ecLlon
lxel urlLy CounL
A
B
uaLa ro[ecLlon
3 3 3 - -
1 2 3
4 5 6
7 8 9
8 3 9 4 3 2 1 2
7 2 6 2 5 4 1 3
3 2 1 2 2 2 9 2
2 3 4 5 6 7 8 9
Pixels
1 2 5 2 3 2 9 2
1
1
2
9
6
Skewers
l CounL lmage
1 0
0 1
0 0
0
3
9
3 3 3 - -
Data
Projection
0
0
0 0
0
PP Count
mage
Selecting subset
of PP Count
mage
9
nllnu8
MNF or PCA Transformation Vector of size
) 1 ( - 5
) 1
`
...
` `
1 ... 1 1
det
)
`
, ,
`
(
2 1
1

=
5
'
5
5
0 0 0
0 0

0 0 =
`
size
1 ) 1 ( - 5
Select the combination that
makes' as large as possible
10
A1C algorlLhm
lnlLlal LargeL plxel
0
Lhe brlghLesL plxel
llnd LargeL plxels unLll Lhe Lh LargeL plxel by
wlLh
< < ) ( ) (

1
0 0
max arg 7 ! 7 ! 9
9 9 7
%
)
1 1 0 1
=

9 9 9 &
< < ) ( ) (

1 - 1 -
max arg 7 7 9 ! !
7

&
%
&
The basic idea of the ATGP is to search the most distinctive pixels by using the
following pixel spectral similarity metric the pixel whose projection in the
orthogonal subspace of other pixels is the most dissimilar pixel
SSLL AlgorlLhm
First step of the SSEE algorithm. A) Original data. B) Subset data
after spatial partitioning.
C) Set of representative SVD vectors used to describe spectral
variance
Second step of the SSEE algorithm. A) Original data. B)
Spectral distribution in 2-dimensional space. C) Projection of
data onto eigenvectors. D) Set of candidate pixels
Third step of the SSEE algorithm. A) Set of candidate pixels. B)
Updated candidate pixels
after including pixels which are spectrally similar to those in the
original set. C) Spatial averaging
process of candidate endmember pixels using a slidingwindow
centered on each candidate. D) First
iteration of spatial-spectral averaging. Averaged pixels shown as
thick lines, with original pixels
shown as thinner lines. E) Second iteration of spatial-spectral
averaging. F) Continued iterations
compress endmembers into clusters with negligible variance
12
2

0
1
0

0
2
Lndmember LxLracLlon AlgorlLhm
SCA
True endmember 2
1
, 2
2
, . 2
5
13
SCA
2

14
SCA
0
1
0

0
2
13
0
1
0

0
2
lCLSLLA ( ConL'd )
16
0
1
0

0
2
lCLSLLA ( ConL'd )
17
0
1
0

0
2
SCA
The MV approaches aim at fnding the mixing matrix M minimizing the volume of
the simplex defned by its columns and containing the observed spectral vectors
This is a nonconvex optimization problem much harder than those considered in
the previous subsection in which the endmembers belong to the data set
ALmospherlc correcLlon
Slgnal subspace ldenLlflcaLlon
Lnd member deLecLlon
unmlxlng(Abundance LsLlmaLlon)