Three algorithms for unsupervised per-field classifications of ERTS data are discussed: a hierarchical technique, an iterative technique, and a chain technique. The feature set used comprised the gray-scale histograms for all four spectral bands, reduced to 32 classes per band, and evaluated for fields containing 50 by 50, 100 ⨯ 100, and 200 ⨯ 200 pixels, yielding respectively 2816, 704 and 176 fields to an ERTS frame. Although it yielded excellent results, it was possible to use the hierarchical technique only on the smallest of these data sets. An iterative technique proved rapid for small data set, but used excessive amounts of computer time to achieve convergence for large data sets. However, an acceptable but non-convergent solution could be achieved within reasonable computation times. The chain algorithm proved to be the most efficient in operation, handling large data sets very rapidly, but yielding the least acceptable classifications. The results indicate that depending upon the scale and purposes of the investigation, a particular classification algorithm is appropriate.
Date of this Version