VARIANCE APPROXIMATIONS FOR ASSESSMENTS OF CLASSIFICATION ACCURACY

Variance approximations are derived for the weighted and unweighted ka ppa statistics, the conditional kappa statistic, and conditional proba bilities. These statistics are useful to assess classification accurac y, such as accuracy of remotely sensed classifications in thematic map s when compared to a sample of reference classifications made in the f ield. Published variance approximations assume multinomial sampling er rors, which implies simple random sampling where each sample unit is c lassified into one and only one mutually exclusive category with each of two classification methods. The variance approximations in this pap er are useful for more general cases, such as reference data from mult iphase or cluster sampling. As an example, these approximations are us ed to develop variance estimators for accuracy assessments with a stra tified random sample of reference data.