Page 301 - Digital Analysis of Remotely Sensed Imagery

Spectral Image Analysis 263

During the actual implementation of K-means clustering, the
analyst needs to specify the convergence threshold in addition to
the number of clusters to be generated. Since it is easier to merge
several clusters into one than splitting one into a few, it is recom-
mended that more clusters than is necessary be specified initially.
As illustrated in Fig. 7.5, not every cluster corresponds to a unique

(a)

(b)

FIGURE 7.5 An example of unsupervised classifi cation in which the input
image is classifi ed into 8 (a) and 12 (b) clusters using a convergence
threshold of 0.950 and a maximum iteration of 10. See also color insert.