Supervised and Forest Type-specific Multi-scale Segmentation for a One-level-representation of Single Trees

In this paper a supervised approach for the object building / segmentation process utilizing a priori knowledge about the specific scale domain of the target features is proposed. One premise was that the result should finally represent the entire scene content in a spatially contiguous one-level-representation (OLR, Lang & Langanke, 2006). High-level segmentation and pre-classification of multi-spectral line scanner data were used to generate an initial set of regions, characterised by their spectral behaviour and height information and accordingly assigned to image object domains (i.e. forest types). Five different forest types were distinguished: coniferous spacious vs. coniferous non-spacious forest, deciduous-spacious vs. deciduous non-spacious forest, and mixed forest. All types were treated independently: optimized multi-scale segmentation was used to build up objects in a region-specific two-level hierarchy. The sub level objects correspond to single trees or tree crowns. The region-specific segmentation is controlled by a combination of rule-sets as being developed by Tiede & Hoffmann (2006) for single tree crown delineation on laser scanning data, taking into account the forest characteristics (i.e. deciduous vs. coniferous or spacious vs. non-spacious) of the respective types. Since working below the initial (super-) level, OLR is finally accomplished. The approach was realised using Cognition Network Language (CNL) for the Definiens Developer Software. CNL offers possibilities to address single objects and to manipulate and supervise the process of building new region-specific scaled objects. The results concerning single tree crown delineation are promising due to the fact that instead of applying one single algorithm for the entire scene a sequence of adapted algorithms for each initial region is performed.