Validating sub-pixel classifications: A comparison of approaches using an artificial dataset

In this paper, different methods for validating sub-pixel classifications are compared with respect to their ability to correctly represent artificially induced confusions. A distinction is made between map-level accuracy measures, which allow a fast comparison and ranking of classifications and confusion matrices that provide a more detailed overview of the confusions between the individual classes. A newly developed method for calculating the soft confusion matrix was also added to the comparison. Our results show that this new approach is able to compete with the existing soft confusion matrices and in some cases even outperforms them.