Number of signatures necessary for accurate classification
This paper presents a procedure for determining the number of signatures to use in classifying multispectral scanner data. A large initial set of signatures is obtained by clustering the training points within each category (such as 'wheat' or 'other') to be recognized. These clusters are then combined into broader signatures by a program that considers each pair of signatures within a category, combines the best pair in the light of certain criteria, saves the combined signature and repeats the procedure until there is one signature for each category. The result is a collection of sets of signatures, one set for each number between the number of initial clusters and the number of categories. With the aid of statistics such as an estimate of the probability of misclassification between categories, the user can choose the smallest set satisfying his requirements for classification accuracy.