Fig. 2: Point cloud reconstruction of a typical 0.4 m x 0.5 m plot
before destructive harvesting, including quadrat frame with coded
targets. Vegetation (primarily Taeniatherum caput-medusae) colored by
height.
Point Cloud Classification and Volumetric Processing
Following reconstruction, point clouds were imported into Matlab
(version R2021a) for classification into ground and vegetation classes
and volumetric measurements (Fig. 3). After clipping the point clouds
inside the frames, we applied the point cloud denoising algorithm
presented by Rusu et al. (2008) to remove the outliers from the point
cloud. To classify the point cloud into ground and vegetation classes we
applied the simple morphological filter (SMRF) algorithm in Matlab.
The simple morphological filter (SMRF) method (Pingel et al. 2013)
classifies the point clouds into vegetation and ground using the
elevation and slope threshold. The algorithm tiles the data by a grid
resolution, finds the minimum elevation points within each grid and fits
a surface to the minimum points. The points that are excluded from the
elevation difference between the minimum surface and morphological
opened minimum surface by linearly increased window sizes (from 1 to
user defined maximum window radius) in an iterative procedure are
considered as vegetation and the points that satisfy the elevation
threshold are classified as ground (more detail can be found in Pingel
et al. 2013).
We used an elevation threshold of 11 cm, slope threshold of 0.1 (as the
plots were mostly flat), elevation scale of 0.9 (to detect medium size
objects) and max window radius of 10. While this is effective, the
ground classified points still contain some vegetation biomass. To
separate the remaining vegetation points from the ground class, we
applied two gaussian mixture classification algorithms on the curvature
of the point clouds using 50 and 200 neighbors (200 neighbors is applied
on the ground class from 50 neighbors output). In this method we assume
that the ground surface is smoother than the vegetation surface and thus
the curvature on the surface is different for those classes. The
algorithm extracts curvatures and fits a gaussian distribution, and
assumes vegetation and ground curvatures follow two mixed gaussian
distributions. After retrieving the vegetation and ground class we
merged the vegetation class with the previous vegetation class from the
SMRF results.