Since the late 1960's automated methods for map generalisation have been studied, but thus far no comprehensive system has been achieved. This is duet to the general complexity of the matter, part of which is caused by the inability to separatethe :conceptual
and the graphic issues. These aspects of map generalisation are considered separate issues ever since the advent of GIS but in practice it has been difficult to disconnect the conceptual issues from the impediments of graphic representation, either in the form of a paper map or on a computer screen. Current research into automated map generalisation generally appears to be in a cul-de-sac for this reason.
This study therefore aims to concentrate on strictly non-graphic operations and large generalisation steps, i.e. big scale changes. Whereas most existing methods work towards a clear end result, this approach does notInstead, it is entirely based on the input data. Minimizing generalisation errors is a priority and assessment of the generalisation results is also an issue to consider. The goal is to develop a system for the generalisation of object- and vector-based categorical maps, such as large-scale topographic data, that is to a large extent automated and can be operated by non-expert users. In the past, several generalisation procedures have been developed for individual objects and dichotomous maps, but the number of procedures for categorical maps is still limited and the methods that do exist rely on similarity and importance factors that are hard to determine.
Large-scale categorical data mostly form an area partition, i.e. the whole spatial extent of the dataset is covered by objects and the objects do not overlap. This implies that objects cannot simply be removed - since this would cause 'holes' - but have to be combined or aggregated.
Objects can be aggregated based on taxonomy orpartonomyrelationships. Taxonomy relationships are based on similarity between the objects or classes. Aggregation based on taxonomy relationships has been described extensively in map generalisation literature, but only works within a limited spatial range. Since this study is aimed at large-scale changes it is based on the much less describedpartonomyrelationships. Inter-object and inter-class relationships are used to determine functionally related classes in order to aggregate the object instances of the class. It is assumed that spatial correlations indicate functional relationships. The class adjacency index is used as a measure of spatial correlation between classes. Combinations of classes with a high class adjacency index are likely candidates for the creation of composite classes. Adjacent objects of these classes can subsequently be aggregated and reclassified to create composite objects.
The class adjacency index is determined based on adjacency measures of the member objects. The input dataset must therefore form a topologically correct, object-based area partition. The implementation is based on a stored adjacency graph and uses regular relational database management software. The data model is object-based and supports the concept of composite objects. In the process a multiple representations dataset is produced by connecting the composite objects created in every aggregation cycle to the constituent parts in the previous level.
The process can be fully automated but it is also possible to allow user interaction at several points in the process without compromising the approach. Since it is entirely based on characteristics of the input dataset, the method is also suited for exploratory purposes. To a certain degree, the meaning of the classes is not even relevant, although in interactive mode the user naturally has to be aware of the classes.
The method was applied to two Dutch topographic datasets:TOPlOvectorand GBKN.Theresults show that this is a very promising method for conceptual generalisation. The concept of composite classes makes that generalisation errors are not an issue. Therefore, it cannot be evaluated using conventional generalisation effect measures. The output of the aggregation process is not readily suitable for mapping purposes, and additional cartographic generalisation is in that case required. The current implementation is notintendedas a complete solution for conceptual generalisation. But since it is set in an environment of other conceptual generalisation operations, such as structural generalisation and extended adjacency graphs, it can be extended to create such a comprehensive system.