Hyperspectral Classification Using ICA Based Mixture Model
227
Table 9.1. Classes and the number of pixels in each class
Class
Number of pixels
Background
594
Corn-notill
640
Grass/Trees
383
Soybeans-min
1583
Total
3200
crop canopies had approximately 5% cover, the rest being soil, covered with the
residue of previous year's crop. Given this low canopy cover, the variation in
spectral response due to - i) the soil type variations and ii) the varying amount
of residue from last season's crop, may have a much greater influence upon the
net pixel spectral response than the variation in the type of class. The other
two classes, which are background and grass/trees, are distinguishable from
each other as well as from the other two classes in most wavelength regions.
Figure 9.3a-d show the results of the first six features produced by PCA,
SPCA, OSP and PP respectively. The first 10 principal components obtained
from the PCA have a total data variance of 99.67%, and thus have been used
for classification by the ICAMM algorithm, with N = M = 10. Similarly,
the features extracted by the remaining three techniques have been ranked
based on their corresponding ranking criterion. We select the first 10 features
extracted by each of these feature extraction techniques.
The first six features obtained by SPCA, correspond to the first two principal
components obtained by performing PCA transformation on each block of
correlated bands. It can be seen that the first feature obtained by both PCA and
SPCA extract almost the same information. We investigate the cause of this
similarity. Since PCA has a significant property that it concentrates the data
variability to the maximum extent possible into the first few components and
the variability of the data is scale-dependant, PCA is sensitive to the scaling of
the data to which it is applied.
For example, if the intensity values of one or some of the hyperspectral
bands are arbitrarily doubled, their contribution to the variance of the dataset
will be increased fourfold, and they will therefore be found to contribute more
to the earlier eigenvalues and eigenvectors. With this remark, it is of interest
to carefully observe the first feature in Fig. 9.3a,b as well as the image of
a spectral band centered at 0.48 JIm (Fig. 9.2a). These three images appear
almost identical and hence we conclude that it is one or some of the bands
in the first block of highly correlated bands employed for SPCA, which have
contributed the most to the first feature obtained by PCA performed on the
entire data.
The improvement provided by SPCA over PCA, in enhancing interclass separability as determined from visual inspection, can be comprehended through
the fact that the third, fourth, fifth and the sixth features generated by PCA do
not exhibit high data variability, while all the six features (shown in Fig. 9.3b)
227
Table 9.1. Classes and the number of pixels in each class
Class
Number of pixels
Background
594
Corn-notill
640
Grass/Trees
383
Soybeans-min
1583
Total
3200
crop canopies had approximately 5% cover, the rest being soil, covered with the
residue of previous year's crop. Given this low canopy cover, the variation in
spectral response due to - i) the soil type variations and ii) the varying amount
of residue from last season's crop, may have a much greater influence upon the
net pixel spectral response than the variation in the type of class. The other
two classes, which are background and grass/trees, are distinguishable from
each other as well as from the other two classes in most wavelength regions.
Figure 9.3a-d show the results of the first six features produced by PCA,
SPCA, OSP and PP respectively. The first 10 principal components obtained
from the PCA have a total data variance of 99.67%, and thus have been used
for classification by the ICAMM algorithm, with N = M = 10. Similarly,
the features extracted by the remaining three techniques have been ranked
based on their corresponding ranking criterion. We select the first 10 features
extracted by each of these feature extraction techniques.
The first six features obtained by SPCA, correspond to the first two principal
components obtained by performing PCA transformation on each block of
correlated bands. It can be seen that the first feature obtained by both PCA and
SPCA extract almost the same information. We investigate the cause of this
similarity. Since PCA has a significant property that it concentrates the data
variability to the maximum extent possible into the first few components and
the variability of the data is scale-dependant, PCA is sensitive to the scaling of
the data to which it is applied.
For example, if the intensity values of one or some of the hyperspectral
bands are arbitrarily doubled, their contribution to the variance of the dataset
will be increased fourfold, and they will therefore be found to contribute more
to the earlier eigenvalues and eigenvectors. With this remark, it is of interest
to carefully observe the first feature in Fig. 9.3a,b as well as the image of
a spectral band centered at 0.48 JIm (Fig. 9.2a). These three images appear
almost identical and hence we conclude that it is one or some of the bands
in the first block of highly correlated bands employed for SPCA, which have
contributed the most to the first feature obtained by PCA performed on the
entire data.
The improvement provided by SPCA over PCA, in enhancing interclass separability as determined from visual inspection, can be comprehended through
the fact that the third, fourth, fifth and the sixth features generated by PCA do
not exhibit high data variability, while all the six features (shown in Fig. 9.3b)