A Hybrid approach for land cover mapping based on the combination of soft classifiers outputs and uncertainty information

In this article, the authors present a hybrid approach to produce more accurate land cover maps of diverse landscapes representing Mediterranean environments. The innovation of the proposed methodology is to use, in the combination of soft classifiers outputs, information about the classification uncertainty associated with each pixel and its neighbourhood. The hybrid combined classification method developed includes the following steps: 1) definition of the training areas; 2) pixel-based classification using several soft classifiers; 3) computation of the classification uncertainty; 4) application of rules to combine the outputs of the pixel-based soft classifications and the uncertainty information obtained with uncertainty measures, on a pixel and neighbourhood based approach; 5) image segmentation; and 6) object classification based on decision rules that include the results of the combined soft pixel-based classification and its uncertainty. The proposed methodology was applied to an IKONOS and a SPOT-4 multispectral image with, respectively, 4m and 20m spatial resolution. The overall accuracy of the hybrid classification obtained with the proposed methodology was higher than the one obtained for the individual pixel-based classifications, which shows that this approach may increase classification accuracy. The approach showed to be particularly powerful to obtain Land Cover Map for landscapes representing Mediterranean environments.