RIVER NETWORK GENERALISATION BASED ON STRUCTURE AND PATTERN RECOGNITION
I.G.N., COGIT Laboratory, Paris, France
Spatial database generalization is a key issue in data presentation and integration. Selection is the generalisation step that consists in choosing the geographical objects (buildings, roads, rivers) that meet the specifications of the generalised database. This paper focuses on a specific point, river network selection. Compared to previous work, this method provides the management of imperfect and clipped networks, islands and island clusters and irrigation areas.
Network selection is a complex process that needs to enrich the initial dataset by automatically extracting implicit structures and patterns. The main enrichment used in this work is the creation of "stream strokes" introduced by Thomson and Brooks in 1999. Strokes are sets of stream segments that follow the "good continuation" principle of Gestalt theory. In strokes creation, the key problem is the definition of the continuity rules. In this method, at each confluence point the main upstream channel and downstream channel are chosen to continue or end the stroke using the following criteria : straightness, upstream stroke length or names.
Additional data enrichments are also used in this method. First, in order to deal with clipped drainage basins, the database is enriched with river sources and sinks that are labelled as "natural" or "due to clipping". The continuity rules for strokes creation are different if the upstream strokes' source is natural or not. River islands and complex islands (i.e. island clusters) are also extracted. Furthermore, the stroke creation process based on flow direction gives incorrect results in irrigation areas. An automatic detection of such areas allows to use a different process there. Finally, flow direction on stream segments is recomputed to correct errors present in initial data.
Using all enrichments, selection step of generalisation can be carried out. Selection focuses on stream strokes on which Horton ordering is applied. The small, low order strokes are eliminated. Strokes composing the outline of the islands and complex islands that are big enough, are kept. In irrigation areas, only natural streams are kept.
This process has been implemented in the GIS Clarity™ and tested on the river network of a French county with heterogeneous landscapes. BD TOPO®, the 1m resolution database of the French NMA, is used as initial data and the method is parameterised to derive a 10m resolution database. It provides results equivalent to BD CARTO® data, the 10m resolution database of the French NMA.