LARS Tech Report Number



Presently, automatic classification of multispectral data images is most commonly effected on a point-by-point basis, as if the data vectors from one resolution element to the next were uncorrelated. In other words, no use is made of the spatial information contained in the scene. This is a useful but suboptimal approach, since in a practical situation strong correlations are certain to exist. "Sample classification" is a name used to refer to those classification schemes in which points a reclassified in sets (samples) rather than individually. It is assumed that all the points in a set belong to the same class. This assumption is met in applications such as crop species identification, where each field contains just one crop. Before sample classification can be applied to the fields, however, the fields themselves must be located in the data image. This is the

role of boundary finding.

Date of this Version

January 1973