Evaluating Landsat classification accuracy from forest cover-type maps

Abstract

The use of complete enumeration in the form of photointerpreted forest cover-type maps to evaluate the accuracy of Landsat classifications was compared with assessments made directly from the aerial photography. A computerized, geographic information system was utilized to compare the Landsat classifications with the cover-type maps on a pixel-by-pixel basis. Error maps of pixels which were similarly misclassified by a variety of algorithms contained a larger number of errors than were verified from the aerial photography. For the two test sites studied, only 67 and 52 percent of the pixels which were originally considered to be in error were substantiated as being in error. Discrepancies between the two results were primarily caused by definitional differences between the cover-type maps and the Landsat classifications, especially with regard to minimum-type size and crown closure estimates of forest land.