The more important point to this paper is that, not only there are
gain and bias variations within the band but also across the bands,
This means, the gain and bias corresponding to a detector in band 1
is not the same as the corresponding detector in band 2 and so on.
From this, one can conclude that there is not only a mismatch of
gains and biases between various detectors spatially, but also
spectrally. This indicates that the various gains and biases are
uncorrelated spectrally. The dynamic range of these biases and
gains are very small and hence they contribute to low variance noise
values. Hence this information is packed into the higher principal
components. In contrast the true image information has a wide
dynamic range which leads to high variance values and hence this
information is packed into the lower principal components.
To restate, applying the KLT in the spectral dimension would result
in all kinds of this uncorrelated or poorly correlated noise
information being packed into the higher principal components
whereas the lower ones contain the information of interest. This
has been verified by several researchers [2,5,11,15] on several
Landsat MSS and TM scenes. It has also been consistently found that
the higher order principal components have very low variance or
energy in them. For example, a typical KLT on a TM scene could
reveal about 1-2 $ of the total energy being contained in the
principal components 5 through 7. Note that this noise energy is
part of the original data energy and can be dropped in the KL domain
by setting the respective components to zero. Visual inspection of
the principal components of the TM and MSS scenes clearly illustrate
these facts. It is a primary reason why remote sensing researchers
pick the first three principal components from the TM or MSS scenes
for image classification and land use studies.
The KLT is an invertible transformation as was seen before.
Performing an inverse KL transformation, taking only the useful
components into account rejecting the noise or low variance
components (by setting them to zero values), then one could expect
the reconstructed image to be free from all unwanted noisy
information. Such a reconstruction also preserves vital image
information. The reconstructed image with the insignficant
components dropped, obviously does not have the same data variance
as the original noisy data. There is some loss of energy which is
contributed mostly by noise and very little (if any) by useful
information. A quantitative measure of such a loss can be obtained
through (4). This spectral filter operates on each pixel across the
bands and hence can remove noise spatially as well as spectrally.
Such a spectral filtering procedure in the KL domain also offers
some flexibility. Based on the variances of each component it
should be easy to find out a priori as to how much loss of energy
takes place by the filtering procedure.
Experimental Results
A 512 X 512 TM scene of the San Francisco bay area has been chosen
for the study. Statistics were then collected on the T-band data.
Table I provides the eigenvalues of the data. This indicates that
the KLT can pack 98.3 % of the total data energy into the first
three principal components and 99.5 % into the first four
266