Contextual Description of Superpixels for Aerial Urban Scenes Classification

Remote Sensing Images are one of the main sources of information about the earth surface. They are widely used to generate thematic maps that show the land cover. This process is traditionally done by using supervised classifiers which learn patterns extracted from few image pixels annotated by the user and then assign a label to the remaining pixels. However, due to the increasing spatial resolution of the images, pixelwise classification is not suitable anymore, even when combined with context. Moreover, traditional techniques used to aggregate context are unsuitable in the scenario of thematic maps generation since they depend on a previous labeling of image pixels/segments and, thus, are computationally inefficient and require a large amount of training data. Therefore, the objective of this work is to develop a description for superpixels which is able to encode their visual cues and local context without labeling them in order to generate more accurate land cover thematic maps. Keywords-contextual descriptor; land cover; thematic maps; remote sensing.