Knowledge and Semantics in Landform Classification

The overall objective of this follow-up project is to develop a sound methodological framework for the integration of semantics in hierarchical landform modelling based on our previous research on detecting and analyzing characteristic land-surface patterns in the object based image analysis (OBIA) realm. Automated classification/extraction of features in OBIA requires the matching of computed image objects with concepts of real-world objects by making use of models that capture the meaning of the features of interest in a structured way. The main objectives of the methodology are: 1. To formalize existing knowledge about the morphology, morphometry and the context of glacial landforms by semantic modelling; 2. To match the semantic landform model with significant object patterns of land-surface models as well as with useful object and pattern properties in order to develop transferable rules for landform classification; 3. To investigate ways for assessing the stability and transferability of derived landform classes across spatial resolutions as well as across areas with comparable topography. The proposed framework will add to increased objectivity and transferability of OBIA-based landform classification. In the broader sense expected outcomes shall contribute to laying the scientific foundations for semantics-based extraction of any kind of features in OBIA from multidimensional and multiscalar field representations such as digital elevation models (DEMs).

In the context of the KnowLand project a semi-automated framework for the computer-aided mapping of landforms (e.g., glacial cirque, drumlin, etc.) in Digital Elevation Models (DEMs) was developed. Traditional mapping approaches were heavily dependent on the level of knowledge and preferences of individual users. As a consequence, different people produced different landform maps for the same DEM (in terms of quality and quantity). The ideal case would be the following: a mapping approach delivers the same, or at least very similar, results for the same area, no matter who the user is. The developed framework is close to this ideal situation: at the one hand, the framework involves automated techniques for extracting representative spatial scales in a DEM; at the other hand, it integrates methods that structure and store the common sense knowledge about specific landforms in an explicit way, i.e. in the form of semantic models. The common sense knowledge about landforms consists of qualitative statements such as high, depression, or below a ridge. For the digital mapping these statements have to be translated into a language that can be understood by computers. If possible, qualities have to be quantified. For instance, the landform quality high can be quantified by the expression > 2000 m. All of these translations make up the landform classification system. As tests showed, this system is objective and thus should also be independent of the DEM and the user. The increased objectivity (compared to previous approaches) results from the integration of explicit knowledge models that capture the standard knowledge about landforms on which the class system is built, and the integration of representative DGM scales to which the class system is finally applied.The developed framework has the potential to become a standard for the computer-aided mapping of landforms in DEMs. However, the framework may be applicable in any case where the interest is automated and objective mapping of spatial units in digital data such as satellite images, aerial photographs or medical images. Examples for such units may be brain areas, habitats or oil spills.