A model that enables user to simulate classification error from ''true'' maps in order to elaborate error-contaminated maps was developed. Both types of maps, true and error contaminated, were used to evaluate accuracy assessment. A large number of random stratified samples were obtained and used to compute the estimates of accuracy indices along with their confidence intervals using the normal approximation to the binomial distribution and bootstraping. Both methods tend to overestimate accuracy, especially the producer accuracy, in case of a rare class confused with a large one. In such case, increasing the size of the sample is not effective and stratification using information about likely confused areas is recommended.