Ground Truth in Classification Accuracy Assessment: Myth and Reality

Abstract

The ground reference dataset used in the assessment of classification accuracy is typically assumed implicitly to be perfect (i.e., 100% correct and representing ground truth). Rarely is this assumption valid, and errors in the ground dataset can cause the apparent accuracy of a classification to differ greatly from reality. The effect of variations in the quality in the ground dataset and of class abundance on accuracy assessment is explored. Using simulations of realistic scenarios encountered in remote sensing, it is shown that substantial bias can be introduced into a study through the use of an imperfect ground dataset. Specifically, estimates of accuracy on a per-class and overall basis, as well as of a derived variable, class areal extent, can be biased as a result of ground data error. The specific impacts of ground data error vary with the magnitude and nature of the errors, as well as the relative abundance of the classes. The community is urged to be wary of direct interpretation of accuracy assessments and to seek to address the problems that arise from the use of imperfect ground data.