The first step in the spectral rule generation algorithm
consists of preprocessing, which determines the target land
cover type, other land cover types in the same hierarchy (non target
land cover types), and all the training data in the same
hierarchy. Step 2 selects the candidate band(s) used in rule
generation. Theoretically, each rule can involve as many as 8
features (4 bands and 4 bands ratios). However, in our
applications we find that three features are sufficient for
generating spectral rules. If the classification accuracy over
the training data set X is higher than a threshold t2, then the
rule is accepted. Table 1 shows the rules generated for the
forest category. This rules were obtained from 26 training
vectors belonging to the non forest cover type. Figure 1 is
IRS 1C Image used for segmentation of the regions. Figure 2
is the image used for forest land cover delineation. In this
example, five rules were generated. Forest rules are used to
distinguish forest from other non forest vegetation, such as
agriculture and rangeland. Rule 1 can identify 20 out of the
26 true forest area and mislabeled 2 non forest areas. Rule 2
can identify 20 true forest areas and mislabeled none of the
non forest areas. Knowledge rules to distinguish shrub from
grass are not easy to obtain. It is known that shrub areas
should have higher visible band reflectance and higher near