The application of uncertainty measures in the training and evaluation of supervised classifiers

Abstract

The production of thematic maps from remotely sensed images requires the application of classification methods. A great variety of classifiers are available, producing frequently considerably different results. Therefore, the automatic extraction of thematic information requires the choice of the most appropriate classifier for each application. One of the main objectives of the research described in this article is to evaluate the performance of supervised classifiers using the information provided by the application of uncertainty measures to the testing sets, instead of statistical accuracy indices. The second main objective is to show that the information provided by the uncertainty measures for the training set may be used to assess and redefine the sample sites included in this set, in order to improve the classification results. To achieve the proposed objectives, two supervised classifiers, one probabilistic and another fuzzy, were applied to a very high spatial resolution (VHSR) image. The results show that similar conclusions on the classifiers’ performance are obtained with the uncertainty measures and the traditional accuracy indices obtained from error matrices. It is also shown that the redefinition of the training set based on the information provided by the uncertainty measures may generate more accurate outputs.