EM 1110-2-2907
1 October 2003
(5) Unsupervised Classification. Unsupervised classification does not re-
quire prior knowledge. This type of classification relies on a computed algorithm
that clusters pixels based on their inherent spectral similarities.
(a) Steps Required for Unsupervised Classification. The user designates
1) the number of classes, 2) the maximum number of iterations, 3) the maximum
number of times a pixel can be moved from one cluster to another with each it-
eration, 4) the minimum distance from the mean, and 5) the maximum standard
deviation allowable. The program will iterate and recalculate the cluster data until
it reaches the iteration threshold designated by the user. Each cluster is chosen by
the algorithm and will be evenly distributed across the spectral range maintained
by the pixels in the scene. The resulting classification image (Figure 5-20) will
approximate that which would be produced with the use of a minimum mean dis-
tance classifier (see above, "classification algorithm"). When the iteration thresh-
old has been reached the program may require you to rename and save the data
clusters as a new file. The display will automatically assign a color to each class;
it is possible to alter the color assignments to match an existing color scheme (i.e.,
blue = water, green = vegetation, red = urban) after the file has been saved. In the
unsupervised classification process, one class of pixels may be mixed and as-
signed the color black. These pixels represent values that did not meet the re-
quirements set by the user. This may be attributable to spectral "mixing" repre-
sented by the pixel.
(b) Advantages of Using Unsupervised Classification. Unsupervised
classification is useful for evaluating areas where you have little or no knowledge
of the site. It can be used as an initial tool to assess the scene prior to a supervised
classification. Unlike supervised classification, which requires the user to hand
select the training sites, the unsupervised classification is unbiased in its geo-
graphical assessment of pixels.
(c) Disadvantages of Using Unsupervised Classification. The lack of in-
formation about a scene can make the necessary algorithm decisions difficult. For
instance, without knowledge of a scene, a user may have to experiment with the
number of spectral clusters to assign. Each iteration is time consuming and the
final image may be difficult to interpret (particularly if there is a large number of
unidentified pixels such as those in Figure 5-19). The unsupervised classification
is not sensitive to covariation and variations in the spectral signature to objects.
The algorithm may mistakenly separate pixels with slightly different spectral val-
ues and assign them to a unique cluster when they, in fact, represent a spectral
continuum of a group of similar objects.
(6) Evaluating Pixel Classes. The advantages of both the supervised and
unsupervised classification lie in the ease with which programs can perform sta-
tistical analysis. Once pixel classes have been assigned, it is possible to list the
exact number of pixels in each representative class (Figure 5-17, classified col-
umn). As the size of each pixel is known from the metadata, the metric area of
each class can be quickly calculated. For example, you can very quickly deter-
5-35
Previous Page
