IEEE Geoscience and Remote Sensing Magazine - March 2021 - 236

FIGURE 3. The general processing procedure of discrete or decomposed waveform lidar data for ecological studies and example algorithms for each processing step. DEM: digital elevation model.

Point Cloud Based
Met ods
M Based
Met ods
Slope Based Filters
Morp olog Filters
Interpolation Based
Filters
Bilateral Filtering
Mean S ift Filtering
mean Cluster
Filtering

round Points
egetation Points
Terrain Surface
Noise
Lidar Points
Terrain Surface

utlier Removal

round Point
Filtering

Normalization

Individual Plant
Segmentation

CHM Based
Met ods
Point Cloud Based
Met ods
eep Learning Based
Met ods

Cro n
iameter
eometric Based
Met ods
Regression Based
Met ods

egetation ttribute
traction

Heig t

236

filters, interpolation-based filters, morphology filters, and
segmentation-based filters [73], [74]; however, the performance varies significantly under different terrain and
vegetation conditions [75], [76]. Special attention needs
to be paid to the selection of filtering algorithms based
on the characteristics of the study area. Normalization is
designed to remove the influence of terrain elevation on
lidar height measurements [77]. A normalized point cloud
can be calculated by subtracting the terrain elevation from
lidar points.
With normalized lidar point clouds, vegetation attributes can be extracted using a variety of approaches.
Two well-established types of approach for deriving vegetation attributes are regression-based approaches (also
known as area-based approaches) and individual plant
segmentation-based approaches [78]. Regression-based
approaches estimate vegetation attributes for preestablished grid cells (typically 20-30 m across) by building
regression models between field measurements and lidar
metrics that summarize the distribution of lidar points
within each cell [79], [80]. Among the various regression
methods, linear regression has been recognized as one of
the most effective because of its ease of interpretation and
relatively high accuracy [81]. In addition to linear regression, the random forest method also shows a remarkable
performance in deriving vegetation attributes because it
can overcome the overfitting problem seen with decision
tree algorithms and does not require an assumption of
normality of the data [81].
To date, vegetation attributes that have been successfully modeled using regression-based approaches include
height, canopy cover, basal area, biomass, and volume at
the community level [37], [82]-[84]. Since regression-based
approaches can use a sampling strategy to derive wall-towall maps of vegetation attributes [85], they can potentially
reduce the need for highly intensive field-based programs
and, as a result, offer more spatial details [86].
As the quality of lidar data improves (e.g., point density and accuracy), the demand for deriving vegetation
structural attributes at the individual plant level also
increases [87]. Individual plant segmentation, i.e., delineating individual plants from the lidar point cloud,
has therefore become the prerequisite first step, and the
procedures can be generally divided into two categories:
canopy height model (CHM)-based methods and point
cloud-based methods. CHM-based methods rely on gaps
among plants in lidar-derived CHMs and use image segmentation techniques (e.g., watershed segmentation)
to delineate the boundary of individual trees [88]-[91].
Point cloud-based methods segment points of each individual plant directly from lidar point clouds [92]-[95].
Vegetation structural attributes at the individual plant
level [e.g., tree height, diameter at breast height (DBH),
leaf area index (LAI), crown base height, and crown diameter] can then be extracted from individual plant
points using geometry-based methods [96]. However,
IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

MARCH 2021
IEEE Geoscience and Remote Sensing Magazine - March 2021

Table of Contents for the Digital Edition of IEEE Geoscience and Remote Sensing Magazine - March 2021
