An assessment of some factors influencing multispectral land-cover classification

Abstract

Experiments to evaluate the accuracies of different stages of land-cover classification are described. Four feature groups, two training strategies, three classifiers, and three accuracy assessment methods have been analyzed. The features used are three original SPOT HRV multispectral images, two principal component images and one edge-density image generated from the original multispectral Band 1 image. Single-pixel training and block training are evaluated. Classifiers used are the minimum Euclidian distance, the minimum Mahalanobis distance, and the maximum likelihood. Pure-pixel sampling, stratified random sampling, and stratified systematic unaligned sampling are used to generate Kappa coefficients for accuracy assessment. Results show that single-pixel training makes the largest contribution to improving classification accuracies. The second largest improvement results from use of the maximum-likelihood classifier rather than the minimum-Euclidian-distance classifier. The third largest contribution is from the inclusion of the edge-density image. Different sampling strategies used for accuracy assessment result in significantly different accuracy values measured by the Kappa coefficient.