315.

[Cover] [Contents] [Index]

Page 82

classification being performed. The classification process is implemented by a set of rules that determine the path to be followed, starting from the root node and ending at one terminal node, which represents the label for the object being classified. At each non-terminal node, a decision has to be taken about the path to the next node.

Figure 2.11 illustrates a simple hierarchical decision-tree classifier using pixel reflectance as input. The aim of classification is to identify the pixel as either ‘vegetation’ or ‘water’. It is obvious that if the rule is not complete after tracing through the decision tree, some pixels will remain unclassified. Thus, we know that the efficiency and performance of this approach is strongly affected by tree structure and choice of feature subsets. Two approaches to the design of a decision tree can be considered. One is based on the user’s knowledge and relies solely on user interaction. This is called the manual design approach (Swain and Hauska, 1977). The second approach uses an automatic procedure.

In the manual design procedure, statistics for all classes are firstly computed, and a graph of the spectral range in each band is constructed. From the graph and the statistical parameters, estimates of the decision bound

Figure 2.11 An example of hierarchical decision tree classifier. See text for further discussion.

[Cover] [Contents] [Index]