ABSTRACT
It has been suggested that different categories of urban land use may be identified in high spatial resolution remotely sensed images by analysing the structural characteristics of their constituent land cover parcels (e.g., buildings, roads and open spaces), in terms of their size, shape and spatial distribution within the scene (Wharton 1982; Meller-Jensen 1990; Gong and Howarth 1992; Eyton 1993; Barnsley and Barr 1996; Barnsley and Barr 2000). The implicit assumption is that each land use exhibits distinct, broadly consistent patterns of buildings and other land cover objects by which it may be recognized. This hypothesis has yet to be fully and rigorously tested, although encouraging results have been obtained in studies making use of multispectral scanner images (Barnsley and Barr 1997; Barr and Barnsley 2000; Bauer and Steinnocher 2001) and, separ ately, rasterized digital map data (Barr and Barnsley 1998a). These studies have employed a graph-based data-processing system, known as SAMS (Structural Analysis and Mapping System), that takes as input rasterized, thematic (land cover) maps (Barr and Barnsley 1995; Barr and Barnsley 1997; Barr and Barnsley 1998b). The system aggregates contiguous pixels sharing the same thematic label into discrete regions and derives a number of morphological properties (e.g., area, perimeter, compactness) and structural relations (e.g., adjacency, containment, distance and direction) from them. This information is stored in a graph-theoretic data model, known as XRAG (eXtended Relational Attribute Graph). In this model, each building (or other land cover parcel) is represented by a node in one or more graphs, and for which morphological properties, such as the area and compactness of the corresponding region, may be stored. Structural rela tions, such as adjacency, are represented by edges connecting these nodes in
the appropriate graph, while structural properties, such as the distance between buildings, are represented as properties of these edges.
