Effect of errors in ground truth on classification accuracy

The effect of errors in ground truth on the estimated thematic accuracy of a classifier is considered. A relationship is derived between the true accuracy of a classifier relative to ground truth without errors, the actual accuracy of the ground truth used, and the measured accuracy of the classifier as a function of the number of classes. We show that if the accuracy of the ground truth is known or can be estimated, the true accuracy of a classifier can be estimated from the measured accuracy. In a series of simulations our method is shown to produce unbiased estimates of the true accuracy of the classifier with an uncertainty that depends on the number of samples and the accuracy of the ground truth. A method for determining the relative performance of two or more classifiers over the same area is then discussed. The results indicate that, as the number of samples increases, the performance of the classifiers can be effectively differentiated using inaccurate ground truth. It is argued that relative accuracies computed using a large number of inaccurate ground truth points are more representative of the true relative performance of the classifiers as they are being evaluated over a larger portion of the scene. An example is presented that uses this method to evaluate the relative performance of two Landsat classifiers.