Page 270 - Digital Analysis of Remotely Sensed Imagery

232 Cha pte r S i x

into new bands to fulfill different purposes, such as decorrelation of
spectral bands, reduction in shared amount of information and the
number of bands used to represent it, and enhancement of certain
features. In this section, three image transformations will be covered:
principal component analysis (PCA), hue-intensity-saturation
(HIS) transformation, and Tasseled Cap transformation.

6.6.1 PCA
In remote sensing, spectral radiation from the ground is captured in mul-
tispectral bands so as to facilitate the identification and detection of
ground features. The recent trend of data acquisition is toward refine-
ment of spectral resolution so that subtle variations in radiance from the
objects of interest can be captured and detected from satellite data. For
instance, the spectral resolution of Airborne Visible/Infrared Imaging
Spectrometer (AVIRIS) imagery has improved to 10 nm with 224 bands.
Such a fine spectral resolution increases the spectral information content
about the target on the one hand. On the other, the huge number of spec-
tral bands undoubtedly leads to data redundancy as the spectral reflec-
tance of some ground objects scarcely varies over certain wavelength
ranges. In other words, not all multispectral bands contribute the same
amount of information toward the total. Chances are most of the infor-
mation contained in one band is also found in another because their
wavelength is so close to each other, hence their information content is
highly correlated. Such data redundancy increases the cost of data stor-
age and needlessly prolongs image classification. It is very desirable and
even necessary to reduce the number of multispectral bands without los-
ing substantial information through image transformation.
Data redundancy among spectral bands is best visualized by plotting
a scatter diagram of pixel values in these bands. Illustrated in Fig. 6.21 are
two-band scatterplots. If the pixels are distributed along a line, then the
information content of the two bands is highly correlated (Fig. 6.21a). If
the pixels are widely scatted without following any trend, then the infor-
mation content of the two bands is uncorrelated (Fig. 6.21b). There is little
data redundancy between them. The issue of data redundancy is addressed
by transforming the raw data into another domain using PCA. PCA is able
to fulfill three objectives:

• First, it can be used to ascertain the information content of each
multispectral band and to identify the most informative bands.
• Second, it is able to reduce the number of bands needed to represent
most of the information contained in the original spectral bands.
• Finally, transformation of the information content onto
orthogonal axes increases spectral separability of certain
spectrally adjacent classes that partially overlap each other in
the original spectral domain.

The undertaking of PCA is illustrated with eight pixels in two spec-
tral bands (Fig. 6.22). The values of these pixels are given in Table 6.2.