Abstract |
Properties of statistical analyses of error matrices generated for accuracy assessment of remote sensing classifications were evaluated for three sampling designs: systematic, stratified systematic unaligned, and simple random sampling (SRS). The population parameters investigated were the proportion of misclassified pixels, P, and the Kappa coefficient of agreement, K. Systematic designs were generally more precise than SRS for the populations studied, except when sampling in phase with periodicity in a population. Bias of the estimated proportion of misclassified pixels, P, was negligible for the systematic designs. The common practice of estimating the variance of P for systematic designs by using an SRS variance estimator resulted in over- or under-estimation of variance, depending on whether the systematic design was more or less precise than SRS. A small simulation study showed that the usual standard error formula for the estimated Kappa coefficient of agreement can perform poorly for systematic designs. (Copyright (c) 1992 American Society for Photogrammetry and Remote Sensing.) |