Assessing the Accuracy of Satellite Image Classifications for Pollutant Loadings Estimation

We examined the accuracy measures to evaluate maps created from knowledge-based classifications of remotely sensed data. The automated classifications involved categories that showed different levels of annual loadings of six pollutants. From the classification error matrices that used spectral information and ancillary data, we computed the overall accuracy and the kappa coefficients. These common measures, however, assume that misclassification errors are equally serious. We propose a procedure, directly related to the pollutant loadings, to calculate weights for the cells in the error matrix to reflect the severity of the misclassification errors. With the weights we were able to calculate the weighted overall accuracy and the weighted kappa coefficient. By using the weighted equivalents of the usual measures of accuracy, we find that there is more specificity in the measures of quality of the classifications for the individual pollutants. * Corresponding author