On the Information Provided by Uncertainty Measures in the Classification of Remote Sensing Images

This paper investigates the potential information provided to the user by the uncertainty measures applied to the possibility distributions associated with the spatial units of an IKONOS satellite image, generated by two fuzzy classifiers, based, respectively, on the Nearest Neighbour Classifier and the Minimum Distance to Means Classifier. The deviation of the geographic unit characteristics from the prototype of the class to which the geographic unit is assigned is evaluated with the Un non-specificity uncertainty measures proposed by [1] and the exaggeration uncertainty measure proposed by [2]. The classifications were evaluated using accuracy and uncertainty indexes to determine their compatibility. Both classifications generated medium to high levels of uncertainty for almost all classes, and the global accuracy indexes computed were 70% for the Nearest Neighbour Classifier and 53% for the Minimum Distance to Means Classifier. The results show that similar conclusions can be obtained with accuracy and uncertainty indexes and the latter, along with the analysis of the possibility distributions, may be used as indicators of the classification performance and may therefore be very useful tools. Since the uncertainty indexes may be computed to all spatial units, the spatial distribution of the uncertainty was also analysed. It’s visualization shows that regions where less reliability is expected present a great amount of detail that may be potentially useful to the user. Keywords—Accuracy assessment, Minimum Distance to Mean Classifier, Nearest Neighbour Classifier, Non-specificity measures, Remote Sensing Images, Uncertainty.