Evaluation of soft possibilistic classifications with non-specificity uncertainty measures

The aim of this paper was to investigate the usefulness of non-specificity uncertainty measures to evaluate soft classifications of remote sensing images. In particular, we analysed whether these measures could be used to identify the difficulties found by the classifier and to estimate the classification accuracy. Two nonspecificity uncertainty measures were considered, the non-specificity measure (NSp) and the U-uncertainty measure, and their behaviour was analysed to evaluate which is the most appropriate for this application. To overcome the fact that these two measures have different ranges, a normalized version (U_n) of the U-uncertainty measure was used. Both measures were applied to evaluate the uncertainty of a soft classification of a very high spatial resolution multispectral satellite image, performed with an object-oriented image analysis based on a fuzzy classification. The classification accuracy was evaluated using an error matrix and the user's and producer's accuracies were computed. Two uncertainty indexes are proposed for each measure, and the correlation between the information given by them and the user's and producer's accuracies was determined to assess the relationship and compatibility of both sources of information. The results show that there is a positive correlation between the information given by the uncertainty and accuracy indexes, but mainly between the uncertainty indexes and the user's accuracy, where the correlation achieved 77%. This study shows that uncertainty indexes may be used, along with the possibility distributions, as indicators of the classification performance, and may therefore be very useful tools.