Object·based Re-classification of High Resolution Digital Imagery for Urban Land-use Monitoring

This paper examines the application of object-orientated processing and artificial intelligence techniques to high spatial resolution satellite sensor images for urban land-use monitoring. Although these techniques have been applied to aerial photography for some time, their use in the analysis of digital images acquired by satellite sensors is much less well developed. Within this study, a two-stage approach is adopted to map urban land use from SPOT-HRV multispectral and panchromatic images. Firstly, a conventional per-pixel multispectral classification is performed to derive a map of the principal land cover types present within the scene. The second stage involves the spatial re-classification of these land cover types into land-use categories. Contiguous blocks of pixels with the same class label are grouped into 'objects' to form an object search map of cover types. This is used to derive an extended region-adjacency graph (XRAG). The XRAG contains information not only on the spatial relationships between individual objects within the scene, but also on the attributes of those objects (e.g. class label, size, perimeter), both of which are used in the re-classification. A priori knowledge of the previous extent of the urban area can also be used to guide the re-classification procedure. Preliminary results obtained using these techniques are shown to be considerably improved with respect to those obtained using a standard, per-pixel classification of the same scene.