Comparison of accuracy and efficiency of five digital image classification algorithms

Abstract

Accuracies and efficiencies of five algorithms for computer classification of multispectral digital imagery were assessed by application to imagery of three test sites (Roanoke, VA., Glade Spring, VA., and Topeka, KA.) A variety of land cover features and two types of image data (Landsat MSS and Thematic Mapper) were represented. Classification algorithms were selected from the General Image Processing System (GIPSY) at the Spatial Data Analysis Laboratory at Virginia Polytechnic Institute and State University, Blacksburg, Virginia and represent a range of available techniques including:

a) AMOEBA (an unsupervised clustering technique with a spatial constraint) b) ISODATA (a hybrid minimum distance classifier) c) BOXDEC (a discrete parallelepiped classifier) d) BCLAS (a Bayesian classifier) e) HYPBOX (a combined parallelepiped-Bayesian classifier)

Two sets of training data, developed for each study site were combined with each technique and applied to each study site.

Parallelepiped classifiers provided the highest classification accuracies but failed to categorize all pixels. The number of classified pixels could be altered by the method of selecting training data and/or adjusting the threshold variable. The minimum distance classifiers were most accurate when the spectral sub-class training data were used. Use of the land cover class training data provided the most accurate results for the Bayesian techniques and decreased the CPU requirements for all of the techniques.

The most important consideration for accurate and efficient classification is to select the classification algorithm that matches the data structure of the training data.