Abstrakt
Sandstone rock outcrops form a unique part of the landscape of Central Europe. Due to climate conditions, they are mostly covered by dense vegetation.
With some limitations, laser scanning technology allows the ground to be captured even under the tree canopy, making it a powerful tool for mapping vegetated and inaccessible terrain. A challenge in deriving a digital terrain model from an acquired point cloud lies in filtering, i.e., discrimination between ground and non-ground points.
Conventional filtering methods applied on complex high-energy terrain with formations resembling man-made objects, e.g., rock walls, do not provide satisfactory results with respect to the accuracy of point assignment to ground and non-ground classes and consequently terrain modelling. However, the quality of digital terrain models is critical for geomorphometric applications and recognition of spatial patterns.
This study proposes three filtering methods adapted to various morphological conditions of sandstone landscape regions i.e., spatially conditioned filtering, object-oriented classification, and filtering with additional terrain data. Spatially conditioned filtering is based on a well-known method of triangulated irregular network densification, but it adjusts selected parameters in an iterative way.
This has been proven to be successful in filtering sandstone rocks by the application of distinct sets of parameters to areas with and without rock formations. Object-oriented classification distinguishes between rock pillars and trees in rock cities based on features describing the distribution of points inside these objects.
The method achieved an overall accuracy of 85 % with respect to manually filtered data and outperformed the conventional methods. Filtering with additional terrain data requires already filtered and co-registered references for spatial querying.
Evaluation by GNSS measurements showed that the digital terrain model derived using this method achieved a higher accuracy than that derived by conventional methods.