Conditional tau coefficient for assessment of producer's accuracy of classified remotely sensed data

Abstract The error matrix is frequently used for accuracy assessment of the classification of remotely sensed data. The classification accuracy for each individual category is often expressed by the percentage correct classified and/or the conditional kappa coefficient. In the classification of remotely sensed data, the distribution of the reference data over various categories is often unknown beforehand. In such cases, conditional kappa may give erroneous estimates of classification accuracy. An alternative measure for classification accuracy of individual categories is proposed. It is denoted as conditional tau, and it expresses the agreement obtained after removal of the random agreement expected by chance. Formulae for computation of conditional tau appropriate for classifications based on equal and unequal probabilities of category membership are presented. Three numerical examples are used to demonstrate the calculation and interpretation of the measure. In the examples, conditional tau correctly estimates the classification accuracy, whilst the results show that conditional kappa may overestimate as well as underestimate the classification accuracy.