IEEE Geoscience and Remote Sensing Magazine - March 2021 - 237

the performance of these algorithms varies significantly
across different topographic and vegetation conditions
[90], [97]. Depending on the accuracy of tree identification, these methods can be prone to errors in both individual object detection and in tree crown delineation
[97], [98] and become less reliable in vegetation stands or
communities that are highly vertically complex and have
many suppressed and lower-level structures [90].
Recently, deep learning techniques have been successfully used to identify and extract features from 3D lidar point
clouds and have segmented individual plants with very high
accuracy [99], [100]. This includes separating individual organs (e.g., stem, branches, and leaves) of an individual plant
[101], [102] and, therefore, makes quantifying the structural
attributes of individual plant organs (e.g., branch length,
branching angle, and leaf length) possible [103], [104]. Although deep learning techniques in lidar data processing
are still under development, they are expected to play an increasingly important role in the future [105].
THE TRANSITION FROM 2D TO 3D ECOLOGICAL
OBSERVATIONS
The highly accurate and informative 3D information provided by lidar sensors is changing the way we study and understand terrestrial ecosystems and is leading the transition
from 2D to 3D ecological observations. In this section, we
illustrate recent progress in 3D ecological observations and
present examples from four terrestrial ecosystems: forest,
agricultural, grassland, and urban ecosystems.
FOREST ECOSYSTEMS
Lidar has been widely used to measure forests in 3D at
different scales [106]. Among the various lidar platforms,
terrestrial laser scanning (TLS) can capture the most accurate and detailed tree structure information and has
been largely used to improve the efficiency of forest inventory practices, especially with the help of terrestrial
mobile lidar systems (e.g., backpack lidar and mobile lidar) (Figure 4) [19], [43], [107], [108]. Forest structural attributes that can be measured from TLS data include, but
are not limited to, DBH, tree height, LAI, clumping index, LAD, leaf inclination angles, crown diameter, crown
base height, and crown volume [109]-[114]. Many of the
structural attributes derived from lidar data avoid the
saturation effect that is typical for passive optical sensors
and may even challenge the accuracy of traditional field
measurements [115]-[117]. For example, the accuracy of
TLS DBH measurements was reported to range from 0.74
to 3.51 cm and tree height measurements from 1.36 to
6.53 m [38], [118], [119]. The accuracy of TLS measurements is primarily influenced by the density and accuracy
of point clouds, tree density, and forest type [120], [121].
Moreover, structural features extracted from TLS data also
showed the potential to identify tree species by using deep
learning techniques [122]-[124]. The combination of TLS
volume estimations and tree species information can be
MARCH 2021

IEEE GEOSCIENCE AND REMOTE SENSING MAGAZINE

used to model aboveground biomass with an accuracy
equivalent to that obtained from allometric equations
[125, 126]. With the development of stem-leaf separation
algorithms, the extraction of tree structure measurements
has moved toward quantifying branching architecture,
including branching angle, branch diameter, and branch
length [25], [127]. Detailed tree structure models from TLS
data also provide the opportunity to parameterize 3D radiative transfer models to simulate light distribution and
gas exchanges of the canopy [128], [129].
Moving toward the landscape scale, airborne and
UAV-borne lidar data are the two main platforms used
for measuring wall-to-wall
forest attributes [45], [130],
[131]. Besides high-resolution
HOWEVER, FULL
topographic products (e.g.,
WAVEFORM LIDAR DATA
digital elevation model and
CAN BE RELATIVELY
slope) [71], [132], wall-toDIFFICULT TO INTERPRET
wall forest attributes-such
VISUALLY AND ARE
as t ree heig ht, L A I, L A D,
USUALLY DECOMPOSED
canopy cover, basal area, and
INTO DISCRETE LIDAR DATA
aboveground biomass-can
BEFORE USE.
also be estimated from lidar
data using either object- or
area-based approaches (Figure 4) [82], [130 ], [133]-[138]. Past research has demonstrated strong correlations between field and lidar measurements of height, volume, and biomass, among others
(R2 = 0.43-0.94) [139], with accuracies driven by topographic conditions, lidar point density, scan angle, and forest type [140]. For example, canopy cover calculated from
lidar data with large scan angles (>20°) tends to be overestimated [141]. By combining the structure information
provided by lidar with optical imagery acquired from satellite or aircraft, vegetation communities in homogeneous
forests can be identified more accurately [142], and shrubs
can be distinguished from trees more easily than when using optical imagery alone [143].
Airborne and UAV-borne lidar data can also be used to
estimate individual tree structural attributes through the
aid of individual tree segmentation algorithms [92], [93],
[144]-[146]. However, the accuracy of individual tree segmentation from airborne and UAV-borne lidar data is not
as high as using TLS data [94], [147]-[150]. Attributes that
can be extracted are limited to the upper canopy-e.g.,
tree height, crown base diameter, and crown base height-
due to the relatively low point density [140], [151], [152].
UAV-borne lidar data are more preferred for extracting individual tree-level parameters than are airborne lidar data
since they have higher point density (>100 points/m 2) [45],
[153]. Airborne and UAV-borne acquired lidar data provide
unique data sets that can be used to reconstruct and visualize forest landscapes and are increasingly used by forest
managers [154], [255].
At regional to global scales, none of the existing lidar
platforms can provide complete coverage alone, and the
237
IEEE Geoscience and Remote Sensing Magazine - March 2021

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