Hierarchical mapping of landforms from Digital Elevation Models (DEMs)

Digital Elevation Models (DEMs) are used for the extraction of land-surface parameters and objects through geomorphometric analysis. Landforms are examples of objects that can be extracted or mapped through wall-to-wall classifications and further used in any application where discrete representations of land surface might serve as variable of interest. In a somehow counter-intuitive manner, most of landform classification systems work through the classification of cells, which could be further clustered to define the borders of objects. This approach is limited in several aspects, including the scattered aspect of classification in the so-called ‘salt-and-pepper effect’, tying the scale of analysis by the raster resolution, difficulties in including topological relationships in classification and also in developing hierarchies of landforms.
   This work aims at investigating methods of producing hierarchical mapping of landforms from DEMs. In our approach, homogeneous objects are produced first through image segmentation of DEMs and their derivatives, which are further used as building-blocks in classification/mapping of landforms. Image segmentation is coupled with multi-scale pattern analysis so that the objects are delineated at characteristic scales in a data-driven fashion. Thus, land-surface parameters as derived from DEMs are segmented into relatively homogeneous areas with eCognition Developer® at a range of scales. At each scale level, local variance (LV) is calculated as the mean value of standard deviation of segments. The values so obtained are plotted against scale levels. High values of LV and its rate of change (ROC-LV) indicate scale levels where objects are associated in patterns of land-surface parameters satisfying the condition of maximizing internal homogeneity while maximizing external heterogeneity. The whole procedure has been implemented as an algorithm called Estimation of Scale Parameters (ESP). This procedure produces homogeneous spatial entities with boundaries such that coarser scale entities have precise boundaries within which finer scales entities nest perfectly. This is a condition for developing hierarchical classifications of landform elements.
   We are currently investigating two methods of developing such hierarchies:
   1. Breaking down complexity through segmentation and successive partitions by nested means. The initially segmented DEM at the scale corresponding to the maximum value of LV is classified in two areas separated by the mean value of elevation. Each area is extracted as independent layer on which segmentations are performed again at the scale indicated by the maximum value of LV and then partitioned at the mean value of another land-surface parameter. This procedure is iterated to produce the third level of the hierarchy. This method is being applied at macro-scale to classify the physiographic units of Africa, as well as at micro-scale to classify landform elements in a flat Dutch landscape for archaeological purposes. Both applications have produced encouraging preliminary results.
   2. Semantic modeling. Real-world features and relationships between them (both horizontal and vertical) are conceptualized based on pre-existing knowledge about morphology, morphometry, and spatial context. Characteristic scales selected as above are integrated within a hierarchy where shape attributes and topologies are formalized so that targeted landforms are extracted or classified. This method is being applied to classify glacial landforms.