CLUSTERING APPROACHES FOR HYDROGRAPHIC GENERALIZATION

Abstract

In many GIS applications users need to visualize and inspect data in different scales. This case requires that the different representations are stored at different levels of detail. One possibility is to store maps from the representation levels in a multiple representation database. Generalization methods are used to maintain and update the spatial databases. Hydrographic data at different resolutions or scales are needed in various spatial studies. This paper presents a generalization operation, selection/elimination, for river networks based on clustering methods, namely k-means and Self Organizing Maps (SOM). This is a model generalization operation, that considers geometric, topological and semantic river attributes as input. Clustering methods group all rivers into different categories according to similarities of the attributes, and then select the rivers for the reduced map scale based on these categories. The clustering methods are compared using the data of United States Geological Survey (USGS) National Hydrography Dataset (NHD) at scales of 1:24000 and 1:100000. Clusters are eliminated based on the coefficient of line correspondence (CLC) is used to evaluate how well the features of clusters match the target 1:100000-scale. If the features of a cluster do not match any features of the target 1:100000-scale (CLC value is zero) or CLC value is different from zero, yet the selection of the cluster decreases the total CLC value of the derived network, then the cluster is eliminated. The derived networks are visually compared to USGS 1:24000-scale and 1:100000-scale NHD. Tpfer's Radical Law is also used for a quantitative comparison. The case study applied to the network illustrates that both clustering approaches for hydrographic generalization have pleasing visual impact. However, SOM-based approach can be used as an effective method for the selection of rivers via data visualization and exploration for multi dimensional geospatial data.