Comparing classification metrics for labeling segmented remote sensing images

Image segmentation and labelling are the two conceptual operations in image classification. As the remote sensing community uses more powerful segmentation procedures with spatial constraint, new possibilities can be explored for labelling. Instead of assigning a label to a single observation (pixel), whole segments of image are labelled at once implying the use of multivariate samples rather than pixel vectors. This approach to image classification also offers new possibilities for using a priori information about the classes such as existing maps or object signature libraries. The present paper addresses the two issues. First a labelling scheme is presented that gathers evidence about the classes from incomplete a priori information using a "cognitive reasoning" approach. Then, five different metrics are compared for the label assignment and are combined through a voting scheme. The results show that very different results can be obtained depending on the metric chosen. The metric combination through voting, being a suboptimal approach does not necessarily provide the best results but could be a safe alternative to choosing only one metric.