Semantic labeling of urban areas in remote sensing imagery using multiple exemplars-based matching

In this study, we are going to focus on the exploration of color based features on labeling remote sensing images. The common widely used color descriptors are based on color histogram or Gaussian Mixture Models. However, the problem of these methods is to lack of the spatial layout information. We propose a new color description and matching approach, which allows to relax the assumption of independence of the observed pixels and incorporates the spatial information naturally by an iterative estimation and a regeneration process. We compare our results to the traditional descriptor based on the labeling of urban scenes using IKONOS imagery and show that our color descriptor outperforms the color histogram and Gaussian Mixture Models, furthering with combination of linearly interpolation and over-segmented map give much better performance.