Comparison of some classification methods on a test site (Kisköre, Hungary): Separability as a measure of accuracy

Abstract

The purpose of this study was to compare some classification methods on a typical Hungarian lowland agricultural area. For a multicrop test area of approximately 20000 ha several one-date and multi-date LANDSAT MSS data sets were classified using supervised and unsupervised methods. The best results were obtained by clustering two-band Kauth-Thomas transforms for each date, using Swain-Fu inter-cluster distance. In general, clustering Kauth-Thomas bands always showed slightly better accuracy than the ones for the MSS 5 and MSS 7 data sets. The accuracies achieved by polygon-vector classification were lower (3–6 percent) than those of multidate clustering. An effort was made to explain the results in terms of spectral separability of agricultural cover classes. The ratio of the average separability of an individual class and the total data set is proposed to measure classification accuracy on areas similar to the Kisköre test site.