Independent Component Analysis
Principal Component Analysis
(compute the PCT for x: W 1 )
lCA
(compute the unmixing transform for x 1: W 2 )
Fig.4.6. lCA Algorithm for processing hyperspectral image
127
relation of the components (Chang et al. 2002) or the equivalent sequential
projection pursuit method (Chiang et al. 2001).
The ICA algorithm was run on an AVIRIS image available at the website
http://dynamo.ecn.purdue.edu/~biehl/Multispec/. The image was acquired in
June 1992 over the Indian Pine Test Site in NW Indiana (see Fig. 4.7). This image
has been used as a test image for conducting several experiments reported in
different chapters in this book. It consists of 145 x 145 pixels in 220 contiguous
spectral bands, at 10 nm intervals in the spectral region from 0040 to 2045 }lm,
at a spatial resolution of20 m. Four of the 224 AVIRIS bands contained no data
or zero values. The advantage of using this dataset is the availability of the
reference data image (also called the ground truth) for accuracy assessment
purposes (see Fig. 4.8). This makes the dataset an excellent source for conducting experimental studies, and, therefore, this image has been used in many
other studies (e. g. Gualtieri et al. 1999; Gualtieri and Cromp 1998; Melgani and
Bruzzone 2002; Tadjudin and Landgrebe 1998a and 1998b) also.
Two-thirds of the scene is covered with agricultural land while one-third is
forest and others. Two major dual lane highways, a smaller road, a rail line, low
density housing, and other building structures can also be seen in the scene.
The scene represents 16 classes as defined in Fig. 4.8.
In this experiment, from the original bands, 34 bands were discarded (due
to sensor malfunctioning, water absorption, and artifacts not related to the
scene). The remaining 186 bands were first whitened using PCA and then
processed through ICA. A visual inspection of the results reveals that the
Principal Component Analysis
(compute the PCT for x: W 1 )
lCA
(compute the unmixing transform for x 1: W 2 )
Fig.4.6. lCA Algorithm for processing hyperspectral image
127
relation of the components (Chang et al. 2002) or the equivalent sequential
projection pursuit method (Chiang et al. 2001).
The ICA algorithm was run on an AVIRIS image available at the website
http://dynamo.ecn.purdue.edu/~biehl/Multispec/. The image was acquired in
June 1992 over the Indian Pine Test Site in NW Indiana (see Fig. 4.7). This image
has been used as a test image for conducting several experiments reported in
different chapters in this book. It consists of 145 x 145 pixels in 220 contiguous
spectral bands, at 10 nm intervals in the spectral region from 0040 to 2045 }lm,
at a spatial resolution of20 m. Four of the 224 AVIRIS bands contained no data
or zero values. The advantage of using this dataset is the availability of the
reference data image (also called the ground truth) for accuracy assessment
purposes (see Fig. 4.8). This makes the dataset an excellent source for conducting experimental studies, and, therefore, this image has been used in many
other studies (e. g. Gualtieri et al. 1999; Gualtieri and Cromp 1998; Melgani and
Bruzzone 2002; Tadjudin and Landgrebe 1998a and 1998b) also.
Two-thirds of the scene is covered with agricultural land while one-third is
forest and others. Two major dual lane highways, a smaller road, a rail line, low
density housing, and other building structures can also be seen in the scene.
The scene represents 16 classes as defined in Fig. 4.8.
In this experiment, from the original bands, 34 bands were discarded (due
to sensor malfunctioning, water absorption, and artifacts not related to the
scene). The remaining 186 bands were first whitened using PCA and then
processed through ICA. A visual inspection of the results reveals that the