Domain-specific class modelling for one-level representation of single trees

As a synthesis of a series of studies carried out by the authors this chapter discusses domain-specific class modelling which utilizes a priori knowledge on the specific scale domains of the target features addressed. Two near-natural forest settings served as testing environment for a combined use of airborne laser scanning (ALS) and optical image data to perform automated tree-crown delineation. The primary methodological aim was to represent the entire image data product in a single, spatially contiguous, set of scale-specific objects (one-level-representation, OLR). First, by high-level (broad-scale) segmentation an initial set of image regions was created. The regions, characterised by homogenous spectral behaviour and uniform ALS-based height information, represented different image object domains (in this case: areas of specific forest characteristics). The regions were then treated independently to perform domain-specific class modelling (i.e. the characteristics of each region controlled the generation of lower level objects). The class modelling was undertaken using Cognition Network Language (CNL), which allows for addressing single objects and enables supervising the object generation process through the provision of programming functions like branching and looping. Altogether, the single processes of segmentation and classification were coupled in a cyclic approach. Finally, representing the entire scene content in a scale finer than the initial regional level, has accomplished OLR. Building upon the preceding papers, we endeavoured to improve the algorithms for tree crown delineation and also extended the underlying workflow. The transferability of our approach was evaluated by (1) shifting the geographical setting from a hilly study area (National Park Bavarian Forest, South-Eastern Germany) to a mountainous site (Montafon area, Western Austria); and (2) by applying it to different data sets, wherein the latter differ from the initial ones in terms of spectral resolution (line scanner RGBI data vs. false colour infrared orthophotos) and spatial resolution (0.5 m vs. 0.25 m), as well as ALS point density, which was ten times higher in the original setting. Only minor adaptations had to be done. Additional steps, however, were necessary targeting the data sets of different resolution. In terms of accuracy, in both study areas 90% of the evaluated trees were correctly detected (concerning the location of trees). The following classification of tree types reached an accuracy of 75% in the first study area. It was not evaluated for the second study area which was nearly exclusively covered by coniferous trees.