References
215
of change needed to continue the clustering). In this context, and given the
available reference data, the PCA derived components provided a classification
accuracy of 43%, ICA-FE an accuracy of 53% and UICA-FE an accuracy of
58%, indicating superior performance of our ICA based algorithms. There are
several potential reasons for the lower values of classification accuracy. First,
the reference data contained large areas that were not classified, including
the road and railroad segments. Additionally, several of the crops were further
detailed in various sub-classes (such as soybeans - notill/min/clean). This finer
classification is useful in other experiments but these classes could not always
be clearly distinguished when only unsupervised classification is employed.
Another set of experiments using the same algorithms was performed on
HYDICE data (Robila and Varshney 2003). In that case, due to a lack of reference
data, the evaluation was based only on visual inspection or was coupled with
target detection (Robila and Varshney 2003; Robila 2002). The results were
consistent with those provided by the above experiment and, therefore, are not
described here.
8.6
Summary
In this chapter, we have presented two ICA-based algorithms for feature extraction. The first (ICA-FE) algorithm uses PCA to reduce the number of bands and
then proceeds to transform the data such that the features are as independent
as possible. The second (UICA-FE) algorithm produces the same number of
components, relying on PCA only for determination of the number and decorrelation of components. In the UICA-FE algorithm, the data are projected into
lower dimensionality directly through the ICA transform. Basically, both algorithms have the same goal: independence of the features to be extracted.
However, by using the information from all the components and not only the
ones selected by PCA reduction, UICA-FE is able to increase the separability
of the classes in the derived independent features. This was validated through
experiments on hyperspectral data. The UICA-FE derived features display an
increased class separation as compared to ICA-FE. Thus, both ICA-FE and
UICA-FE provide attractive approaches for unsupervised feature extraction
from hyperspectral data.
References
Achalakul T, Taylor S (2000) A concurrent spectral-screening PCT algorithm for remote
sensing applications. Information Fusion 1(2): 89-97
Hyviirinen A, Karhunen J, Oja E (2001) Independent component analysis. John Wiley and
Sons, New York
Lee TW (1998) Independent component analysis: theory and applications. Kluwer Academic
Publishers, Boston
Rencher AC (1995) Methods of multivariate analysis. John Wiley and Sons, New York
Richards JA, Jia X (1999) Remote sensing digital image analysis: an introduction. SpringerVerlag, Berlin
215
of change needed to continue the clustering). In this context, and given the
available reference data, the PCA derived components provided a classification
accuracy of 43%, ICA-FE an accuracy of 53% and UICA-FE an accuracy of
58%, indicating superior performance of our ICA based algorithms. There are
several potential reasons for the lower values of classification accuracy. First,
the reference data contained large areas that were not classified, including
the road and railroad segments. Additionally, several of the crops were further
detailed in various sub-classes (such as soybeans - notill/min/clean). This finer
classification is useful in other experiments but these classes could not always
be clearly distinguished when only unsupervised classification is employed.
Another set of experiments using the same algorithms was performed on
HYDICE data (Robila and Varshney 2003). In that case, due to a lack of reference
data, the evaluation was based only on visual inspection or was coupled with
target detection (Robila and Varshney 2003; Robila 2002). The results were
consistent with those provided by the above experiment and, therefore, are not
described here.
8.6
Summary
In this chapter, we have presented two ICA-based algorithms for feature extraction. The first (ICA-FE) algorithm uses PCA to reduce the number of bands and
then proceeds to transform the data such that the features are as independent
as possible. The second (UICA-FE) algorithm produces the same number of
components, relying on PCA only for determination of the number and decorrelation of components. In the UICA-FE algorithm, the data are projected into
lower dimensionality directly through the ICA transform. Basically, both algorithms have the same goal: independence of the features to be extracted.
However, by using the information from all the components and not only the
ones selected by PCA reduction, UICA-FE is able to increase the separability
of the classes in the derived independent features. This was validated through
experiments on hyperspectral data. The UICA-FE derived features display an
increased class separation as compared to ICA-FE. Thus, both ICA-FE and
UICA-FE provide attractive approaches for unsupervised feature extraction
from hyperspectral data.
References
Achalakul T, Taylor S (2000) A concurrent spectral-screening PCT algorithm for remote
sensing applications. Information Fusion 1(2): 89-97
Hyviirinen A, Karhunen J, Oja E (2001) Independent component analysis. John Wiley and
Sons, New York
Lee TW (1998) Independent component analysis: theory and applications. Kluwer Academic
Publishers, Boston
Rencher AC (1995) Methods of multivariate analysis. John Wiley and Sons, New York
Richards JA, Jia X (1999) Remote sensing digital image analysis: an introduction. SpringerVerlag, Berlin