Digital elevation models (DEMs) generated from raw TLS point clouds in vegetated dune environments represent a surface that includes both the ground and the vegetation (top right).   For the purposes of this research, I am only interested in those points that represent the ground surface, and so the two must be separated.  This is the challenge I am currently working through at this stage of my research.



What techniques are available for classifying / filtering unwanted points from TLS point clouds?



Unlike airborne LiDAR, most ground-based laser scanners only record first returns.  This means that last return is not an option for first-order vegetation filtering from TLS point clouds.  There are some alternatives:



-Lowest points analysis (ie. with Leica's Cyclone II Topo or Quick Terrain Modeller) - divide up the cloud into a grid and assume the point with the lowest z value in each grid cell of specified size represents the ground

-Using reflected laser intensity - use difference in laser intensity between ground surface and vegetation to filter cloud

-Using a multi-scale dimensionality criterion (after Brodu and Lague, 2012) - build a classifier for classifying ground and vegetation based on differences in dimensionality of the two classes at different scale intervals

-Using RGB values (after Lichti, 2005) - use difference in RGB values between ground surface and vegetation to filter cloud



How will I decide which technique generates a bare-Earth surface that most accurately represents the true ground surface?  



Recently I completed a ground-truthing survey whereby eighty known ground points were surveyed in patches of vegetation at each of my field sites using an electronic distance meter (EDM) - like the one you often see pairs of road surveyors using.  With this instrument, I can be sure the points surveyed represent the ground surface.  At the same time, a TLS survey was performed at each site.  Each method will be applied to the point cloud from the scan to filter the vegetation points and bare-Earth DEMs will be generated.  The corresponding known ground points will be compared against their location within the EDM to determine which method produces a DEM that most accurately represents the true ground surface.  Analysis of this data is currently ongoing.