Digital image classification assigns pixels to classes. Each pixel is
treated as a set of values in several spectral bands, derived from the
brightness of the same area on the earth's surface. By comparing
pixels to one another, and to pixels of known identity, it is possible
to assemble groups of similar pixels into classes that are associated
with the informational categories of interest to clients. Because
these classes form regions on an image, after classification, the
digital image is presented as a mosaic of uniform parcels, each
identified by a color or symbol. These classes are, in theory,
homogeneous-- pixels within classes are spectrally more similar to one
another than they are to pixels in other classes. In practice, of
course, each class will display some variation, because category
exhibits natural variation. Image classification forms one of the
most important tools for examination of digital images--sometimes to
produce a final product, other times as one of several analytical
procedures applied to derive information from an image.
Image classification requires the analyst to match
informational classes to spectral
classes. Informational classes are the categories
of interest to the ultimate users of the data. Informational classes
are (for example) the different kinds of forest, or the different kinds
of land use that convey information to planners, managers, and
scientists who will use information derived from remotely sensed data.
These classes convey the information that we wish to derive from the
data--they are the object of our analysis. Unfortunately, remotely
sensed images do not directly convey informational classes-- we can
derive them only indirectly, using the brightnesses that compose each
image. For example, the image cannot directly show geological units,
but rather only the differences in topography, vegetation, soil color,
shadow, and other factors that lead the analyst to conclude that
certain geological conditions exist in specific areas.
In contrast, Spectral classes are groups of
pixels that are uniform with respect to brightnesses in their several
spectral channels. The analyst defines spectral classes within
remotely sensed data; then must define links between spectral classes
on the image and informational classes that are of interest to the
client. In this manner, image classification proceeds by matching
spectral categories to informational categories. If the match can be
made with confidence, then the information is likely to be reliable.
If spectral and informational categories do not correspond, then the
image is unlikely to be a useful source for that particular application.
Informational classes are typically composed of numerous spectral
subclasses--spectrally distinct groups of pixels that together may be
assembled to form an informational class. In digital classification,
we must often treat spectral subclasses as distinct units during
classification, but then display several spectral classes under a
single symbol for the final image or map to be used by planners or
administrators (who are, after all, interested only in the
informational categories, not the intermediate steps required to
generate them).
Overview of Process of Supervised Image Classification
Unsupervised classification can be defined as the identification of natural groups, or structures, within multispectral data. The notion of the existence of natural, inherent groupings of spectral values within a scene may not be intuitively obvious, but it can be demonstrated that remotely sensed images are usually composed of spectral classes that internally are reasonably uniform in respect to brightnesses in several spectral channels. Unsupervised classification is the definition, identification, labeling, and mapping of these natural classes.