Hierarchical object-based image analysis of high-resolution imagery for urban land use classification

With the concentration of artificial features and a combination of natural features, two types of aggregated objects can be defined in urban area: 1. Area objects consisting of a spatial contiguous set of detailed objects such as building roofs, grass land etc. 2. Linear objects such as roads, ditches and edges of lakes, which form the boundaries of area objects. RS image analysis for urban land use classification can be organized as a set of processes for image (sub)segmentation identifying urban (sub)regions according to their functions. Land use function can be inferred from land cover type, its distribution and proportion. In order to incorporate spectral and spatial information derived from imagery as well as data from other sources in land cover/use classification, a hierarchical approach of image analysis is proposed using meaningful feature-based segmentation to form a set of hierarchical image objects, characterized by land use, land cover, intermediate image objects etc. hierarchically. The hierarchy of image objects consist of four layers, from bottom upwards: 1. Pixels form the bottom layer. 2. Intermediate image objects are created based on hierarchical segmentation or homogeneity of neighbor pixels. 3. Intermediate image objects are merged to form land cover objects according to their spectral and spatial properties. 4. Land use classification is inferred according to properties of land cover objects. A hierarchical segmentation process is proposed in forming image objects hierarchically. A number of issues have been investigated such as hierarchical image segmentation, image object, semantic group, and classification problems. The experiment shows that the hierarchical image analysis approach is a powerful tool in incorporating spectral and spatial information from imagery as well as other ancillary data. Hierarchical image objects offer fundamental bases in collecting meaningful information derived from imagery and make it possible to apply various rules in image understanding and classification. The hierarchical approach has potential application in image object aggregation and land use generalization.