Urban scene classification with VHR images

Urban scenes refer to those spatial units that consist of diverse geographic objects but have special patterns, and thus can be abstracted as one scene category higher than object classes, such as residential areas, commercial areas, and industrial areas. Therefore, scene analysis consists of two steps: scene segmentation and scene classification. Since existing object- and pixel-based image analysis methods cannot work well for scene analysis, this paper aims at resolving the two issues. For scene segmentation, a multilevel aggregation mechanism is presented; while for scene classification, both the intra-scene feature similarity and interscene semantic dependency are considered. VHR images are used to test the presented methods and satisfactory results with large accuracies are obtained.