Methods: Computing the DEM/DSM
H. Mitasova, M. Overton, E. Hardin, M.O. Kurum
Lidar point clouds can be transformed into raster-based Digital Elevation Models (DEM) representation that is well suited for surface analysis.
- Point cloud analysis using binning
LIDAR point data needs to be interpolated and converted to raster data. But first a resolution must be chosen. If the resolution is too low, important features can be lost. If the resolution is too high, computing time and memory issues can arise. The lowest resolution without losing topographic detail needed for the given application is the resolution that should be chosen. This is the resolution at which at which the range in z values of all LIDAR points in a given raster cell is less than or equal to the accuracy in the LIDAR. This can be determined by computing preliminary bin statistics. The optimal resolution can be found by computing bin statistics at different resolutions, for example, 0.5m, 1m, 2m, 5m, and 10m.
- Systematic error assessment
Due to inaccuracies in georeferencing during the LIDAR survey, a Digital Elevation Model (DEM), especially if it is from older survey, can include systematic error that needs to be corrected. In order to correct this error, pick a feature in the region that had a stable elevation throughout the study period and use that feature's elevation to calibrate each DEM. This feature will often be a man-made structure like the centerline of a road. As an example, systematic error in the Outer Banks, NC can be corrected using accurate geodetic benchmark points taken along the centerline of NC12 by the NC DOT. A complete list of these points can be downloaded from http://www.obtf.org/NC12Alignment/NC12.htm. If a number of geodetic benchmarks lie inside of the study region, a correction can then be applied to the entire DEM equal but opposite to the median error. To this end, the median error becomes zero. The median may be more appropriate because it is less susceptible to outliers.
- Interpolation and smoothing
LIDAR data may include gaps, especially if the DEM that we want to compute has resolution close or higher than the distances between the data points. Lidar data can be interpolated using any of the standard interpolation functions. We use simultaneous interpolation and smoothing by regularized spline with tension in this tutorial. The topographic surface is modeled from the LIDAR data as a thin plate suspended from the LIDAR points by springs. There are two parameters that must be tuned specifically for each LIDAR point cloud based on the data density and terrain geometry:
- Tension adjusts the stiffness in the plate. The greater the tension, the tighter the surface will bend to the LIDAR points (the surface behaves like rubber sheet), low tension leads to smoother transition between points (the surface behaves like thin steel plate - it is stiffer).
- Smoothing adjusts the stiffness of the springs. The greater the smoothing, the more the surface will be allowed to deviate from LIDAR points and the smoother the resulting surface will be.
Additional parameters tune the segmentation and control the selection of points. In areas that are not mapped with sufficient data, rectangular segments are visible on aspect or curvature maps. The segments can be removed by: (a) increase npmin parameter to make the overlap between segments larger (this slows down the computation); (b) add few points on the area with visible segments, e.g. by randomly sampling this area iwith sparse points in the interpolated DEM, adding these points to the original data set and re-interpolating the DEM; (c) reduce the points in the original data set while preserving the geometry information using generalization. Any interpolation and data addition that is not based on the direct measurement should be done with caution and awarness that we don't know how topography looks like in these areas - visible segments can be interpreted as uncertainty indicator of insuficient data in particular location.