| Summary: | The generalization process works on independent image-objects formed by a closed-boundary and a corresponding enclosed-region on the classified image. Initial image-objects are obtained by integrating the classified image with a geographically corresponding satellite-derived edge-segmented image. The generalization process is organised in two main levels of abstraction. Firstly Geometric generalization, responsible for the spatial and thematic simplification of the classified image, elaborating the initial image-objects in to higher-level polygons (the spatial basis of the final product). Secondly Semantic generalization, responsible for the thematic conversion of the simplified product, associating each higher-level polygon to the most appropriate land use class. The CORINE land cover classification scheme was taken as the target product during this thesis. The classification scheme is however overly detailed for direct comparison with satellite-derived products. To overcome this an intermediate classification was defined in this thesis: Pseudo CORINE (Pcor), which is a 1-level scheme containing: bottom-level CORINE classes which are automatically recognisable by image processing techniques, and 2nd-level CORINE classes as substitutes for those CORINE classes not automatically recognisable. The definition of the Pcor scheme allowed an automatic nomenclature conversion, organised in two steps: 1)re-labelling, based on syntax matching, of low-level classes presenting a one-to-one relationship with a single Pcor class. 2)contextual reasoning, based on mutually exclusive hierarchical rules, for the conversion of low-level classes which present a one-to-many relationship with Pcor classes. A fully automatic generalization process has been developed and verified during this work. The automatic generalization process has produced generalized products which are in excellent agreement with the target CORINE map. The simplification of geometry and content of the input information based on image-statistics and contextual rules is fully automatic, unsupervised, consistent, objective, repeatable and generally applicable.
|
|---|
