HOW WELL DO TERRAIN OBJECTS DERIVED FROM PRE-EVENT DIGITAL ELEVATION MODELS SPATIALLY CORRESPOND TO POST-EVENT LANDSLIDES?
The aim of this study is to analyze how well terrain objects that are derived from a pre-event (outdated) DEM spatially correspond to post-event landslides. The multiresolution segmentation as implemented in eCognition was employed to partition terrain variables into homogeneous terrain objects. Next to curvatures and slope, more complex variables such as topographic openness and sky-view factor were segmented, individually and in combination. By changing the scale parameter of the algorithm different scales of terrain objects – ranging from hillslope to sub-landslide – were generated for the same variable(s). Multiresolution segmentation was statistically optimized through multi-scale analysis of the local variance of terrain objects. The optimized terrain object scales were finally intersected with manually derived landslide reference polygons to determine the overlapping areas. Hence, the total spatial coincidence between terrain objects and reference polygons was analyzed in relation to the input variable(s) and scale factors.
Results show that the overlaps are generally small indicating a relatively low landslide predictive capacity of terrain objects that are based on pre-event DEMs. However, some terrain variables generated significantly higher overlaps with the post-event landslides than others. These variables should be preferred for automated EO-based landslide mapping in cases where no post-event DEM is at hand.