Current methods of terrain classification by means of airborne multispectral observations are reviewed with emphasis on the selection of training sets for determination of the categorizer parameters. A method of selecting sample regions for assigning identities to the spectral signatures on the basis of statistically determined similarities, rather than on a priori considerations, is suggested. This method has been tested on data collected on two flights with the University of Michigan scanner over an agricultural region in California. We have found that a simple clustering algorithm, modified to take into account specific features of the crop-census problem, can be used to obtain the desired homogeneous regions with relatively little computation and that very sparse sampling of these regions is sufficient to assign the appropriate category to each cluster. Viewed as a two-stage sampling procedure, clustering improves the second stage classification on 15 crops from 20 to 50 percent over a random selection of the primary sampling units. The accuracy increases to 73 percent when only five classes are considered, with further improvement to 88 percent when a majority decision based on known field boundaries is used.
Note: The Institute of Electrical and Electronics Engineers, Incorporated is distributing this Article with permission of the International Business Machines Corporation (IBM) who is the exclusive owner. The recipient of this Article may not assign, sublicense, lease, rent or otherwise transfer, reproduce, prepare derivative works, publicly display or perform, or distribute the Article.