as the principal components transformation. The KLT is applied in
remote sensing [11], generally in the spectral dimension. It has
been widely used by remote sensing people [3-4] mainly to reduce the
feature space for better olassification in land use studies. The
name Principal Components is derived from the fact that most of the
energy in the original data is compacted into fewer components which
lends itself to better classfication.
The KLT is an optimal transform in a statistical sense under a
variety of criteria. The basis vectors of the KLT are the
eigenvectors of the input covariance matrix. The corresponding
eigenvalues is an indication of the energy packed into each one of
the components. Also most of the energy or variance is packed into
the fewest number of coefficients possible by any transform. Hence
it is the optimal transform. The KLT completely decorrelates the
input data in the transform domain. The KLT simply diagonalizes the
covariance matrix. The diagonal non-zero values of this diagonal
matrix are the eigenvalues. Each eigenvalue is also the variance of
that particular coefficient. The KLT is derived from the statistics
of the image data. Hence the KLT varies if the image statistics
vary. This can be considered as its primary disadvantage. One
needs to compute the transformation every time the statistics
change. At the same time, if the statistics do not change, the
transformation matrix (or the eigenvector matrix) does not change.
As an example, the same transformation can be used for various
images if all the images to be transformed are assumed to obey the
Markov-I process. Another important disadvantage about the KLT is
that it has no fast algorithms in general. However fast algorithms
have been proposed [12-13] for the KLT for a class of process such
as a stationary Guass-Markov process. Additionally signficant
computational effort is involved in computing the basis vectors of
KLT.
In general the eigenvectors are ordered as per the decreasing order
in corresponding eigenvalues. Now most of the energy is packed into
the first few transform coefficients. Also the transform components
are uncorrelated. Since each eigenvalue is the variance of the
corresponding principal component, it is an indication of the energy
packed into the particular principal component. The variance
distribution curve in the KLT domain drops down steeply from the
first component onwards. Since the KLT is an orthogonal transform
[14], the total data variance is preserved in the transform domain.
The very same property guarantees that the inverse transform matrix
is the same as the transpose of the forward transform matrix. In
other words if [T] denotes the forward transform matrix, then
«1
[T)[IT] s [T] [T3'.s [1] — n
where [T]' denotes the transpose of [T] and [I] denotes the identity
matrix. The KL transformation on a data vector [D] is generally
obtained as [11]
[Pi =: (TI [:.[D1:- [M1 ] a vn 2)
264