Using Uncertainty Information to Combine Soft Classifications

The classification of remote sensing images performed with different classifiers usually produces different results. The aim of this paper is to investigate whether the outputs of different soft classifications may be combined to increase the classification accuracy, using the uncertainty information to choose the best class to assign to each pixel. If there is disagreement between the outputs obtained with the several classifiers, the proposed method selects the class to assign to the pixel choosing the one that presents less uncertainty. The proposed approach was applied to an IKONOS image, which was classified using two supervised soft classifiers, the Multi-layer Perceptron neural network classifier and a fuzzy classifier based on the underlying logic of the Minimum-Distance-to-Means. The overall accuracy of the classification obtained with the combination of both classifications with the proposed methodology was higher than the overall accuracy of the original classifications, which shows that the methodology is promising and may be used to increase classification accuracy.