CHAPTERS
Feature Extraction from Hyperspectral Data
Using ICA
Stefan A. Robila, Pramod K. Varshney
8.1
Introduction
Most of the image processing techniques for multispectral or hyperspectral
data have complexity that depends directly on the number of spectral bands
in the acquired data (Swain and Davis 1978). Due to the large number of
bands involved in the hyperspectral images, it is of interest to find methods
that transform the image cube into one with reduced dimensionality while, at
the same time, maintaining as much information content as possible. These
techniques are known under the general name of feature extraction (Richards
and Jia 1999). The term feature is used to refer to the spectral bands or other
transforms derived from combinations of bands.
In Chap. 4, we have seen that independent component analysis (lCA)
(Hyvarinen et al. 2001) has the promise of being an efficient feature extraction
tool. There were, however, several drawbacks. Application of ICA to the full
data set provides a way of separating the class information in different bands.
Unfortunately, the number of bands in the hyperspectral imagery makes the
ICA method extremely time consuming and decreases its attractiveness for use
with these datasets. In addition, automated identification of the bands associated with the relevant independent components (i.e. the ones containing useful
class information) is difficult. Previously known ranking methods (based on
signal to noise ratio, kurtosis, entropy, distances between the transform vectors, etc.) are not efficient for such identification (Robila et al. 2000). Moreover,
a very large number of observations may also decrease the efficiency of the
algorithm since ICA will try to produce more independent components than
the actual number of components in the dataset.
To overcome the above deficiencies, we suggest two ICA based approaches
for unsupervised feature extraction in this chapter. The first one is a hybrid
PCA/ICA algorithm, where PCA is used to decorrelate and reduce the data.
ICA is then applied to the reduced data cube. This hybrid ICA based feature
extraction algorithm (lCA-FE) is a direct improvement over the PCA based
feature extraction algorithm, the resulting bands being as independent as
possible.
In the second approach, we use PCA only for decorrelation of data without
performing any band reduction. To reduce the data, we design a new ICA
P. K. Varshney et al., Advanced Image Processing Techniques for Remotely Sensed Hyperspectral Data
© Springer-Verlag Berlin Heidelberg 2004
Feature Extraction from Hyperspectral Data
Using ICA
Stefan A. Robila, Pramod K. Varshney
8.1
Introduction
Most of the image processing techniques for multispectral or hyperspectral
data have complexity that depends directly on the number of spectral bands
in the acquired data (Swain and Davis 1978). Due to the large number of
bands involved in the hyperspectral images, it is of interest to find methods
that transform the image cube into one with reduced dimensionality while, at
the same time, maintaining as much information content as possible. These
techniques are known under the general name of feature extraction (Richards
and Jia 1999). The term feature is used to refer to the spectral bands or other
transforms derived from combinations of bands.
In Chap. 4, we have seen that independent component analysis (lCA)
(Hyvarinen et al. 2001) has the promise of being an efficient feature extraction
tool. There were, however, several drawbacks. Application of ICA to the full
data set provides a way of separating the class information in different bands.
Unfortunately, the number of bands in the hyperspectral imagery makes the
ICA method extremely time consuming and decreases its attractiveness for use
with these datasets. In addition, automated identification of the bands associated with the relevant independent components (i.e. the ones containing useful
class information) is difficult. Previously known ranking methods (based on
signal to noise ratio, kurtosis, entropy, distances between the transform vectors, etc.) are not efficient for such identification (Robila et al. 2000). Moreover,
a very large number of observations may also decrease the efficiency of the
algorithm since ICA will try to produce more independent components than
the actual number of components in the dataset.
To overcome the above deficiencies, we suggest two ICA based approaches
for unsupervised feature extraction in this chapter. The first one is a hybrid
PCA/ICA algorithm, where PCA is used to decorrelate and reduce the data.
ICA is then applied to the reduced data cube. This hybrid ICA based feature
extraction algorithm (lCA-FE) is a direct improvement over the PCA based
feature extraction algorithm, the resulting bands being as independent as
possible.
In the second approach, we use PCA only for decorrelation of data without
performing any band reduction. To reduce the data, we design a new ICA
P. K. Varshney et al., Advanced Image Processing Techniques for Remotely Sensed Hyperspectral Data
© Springer-Verlag Berlin Heidelberg 2004