IEEE Geoscience and Remote Sensing Magazine - March 2021 - 242

Unlike in other ecosystems, lidar applications in urban
ecosystems have the unique prerequisite of classifying vegetation (e.g., trees and grass cover) and artificial objects (e.g.,
buildings, power lines, bridges, vehicles, and railways)
from lidar point clouds (Figure 7). Various classification
algorithms have been introduced that can be generally divided into four categories: pixel-wise classification based
on lidar-derived surfaces, object-based segmentation
based on lidar-derived surfaces, hierarchical semantic
segmentation based on lidar
ONE OF THE PARAMETERS
point clouds, and deep learnTHAT HAS DRAWN THE
ing-based methods.
ATTENTION OF
Pixel-wise classification
RESEARCHERS IS THE
methods apply traditional
machine learning methods
URBAN GREEN SPACE-THE
(e.g., artificial neural netPUBLIC AND PRIVATE
work, maximum likelihood,
URBAN SPACES PRIMARILY
support vector machine, ranCOVERED BY VEGETATION.
dom forest, Gaussian mixture
modeling, rule-based classification, conditional random
fields, and Markov random fields) on lidar-derived surfaces
(e.g., digital surface models and intensity maps) to classify urban objects [217]-[223]. Object-based segmentation
methods, which employ a user-defined procedure to classify
segmented objects from lidar-derived surfaces [224]-[226],
are designed to avoid the salt-and-pepper effect resulting
from pixel-wise classification [227], [228]. Numerous studies have reported that object-based methods can achieve an
overall accuracy of >80%, which is higher than what can be
derived with pixel-wise methods [229]-[234].
To date, much of the focus in an urban context has been
on the identification of large urban objects, while small
objects (e.g., trees, street lamps, and power lines) are often neglected. Instead of classifying urban objects from
lidar-derived surfaces, hierarchical semantic segmentation methods rely on the entity and contextual information (e.g., points, voxels, planar segments, and objects)
within lidar point clouds, which can be used to classify
specific urban objects (e.g., roofs, walls, pavements, grass
fields, trees, and street lamps) with an accuracy of greater
than 90% [235]-[237]. Deep learning-based methods apply convolution neural networks directly to lidar point
clouds or lidar-derived surfaces to classify urban objects,
and studies have shown that they are highly robust and
can achieve an overall accuracy of greater than 96% [238].
Moreover, recent studies have suggested that combining
lidar data with color information, such as from aerial or
street view photos, can further improve the classification
accuracy [218], [224].
Vegetation structure and functional attributes-such as
tree height, DBH, crown diameter, crown base height, LAI,
tree species, vegetation volume, vegetation coverage, and
aboveground biomass-can be extracted from the classified lidar point clouds in urban environments (Figure 7)
242

[162], [214 ], [239]-[242]. The accuracy of these extracted
attributes is usually higher than in forested areas due to
the relatively simple tree species composition and regular
tree arrangement in urban ecosystems [239], [243]. One
parameter that has drawn the attention of researchers
is the urban green space-the public and private urban
spaces primarily covered by vegetation [244], [245]. Lidar
offers a new way to quantify urban green space from 2D
horizontal arrangements to 3D horizontal and vertical
arrangements, greatly benefiting studies of urban function connectivity, urban planning and management, and
biodiversity [242], [246], [247]. Besides vegetation indices, attributes related to buildings can be also developed
from the lidar points, including building height, building mass, building volume, building 3D model type, urban envelope index, and sky openness (Figure 7) [238],
[242], [248]-[252]. These 3D building indices can be
used together with urban vegetation attributes to study
microclimates and urban flora and fauna, as well as urban planning and management.
Overall, lidar applications in urban ecosystems share
common characteristics with other ecosystems in terms
of extracting vegetation information, but they involve
more efforts in recognizing and processing artificial urban components. The current major obstacles lie in how
to handle the huge amount of lidar data at a city scale, especially for data collected by mobile lidar systems. Moreover, how to incorporate lidar-derived 3D parameters
into urban ecological studies and modeling still needs to
be further explored.
LIDAR OBSERVATIONS TO INFORM 3D
ECOLOGICAL MODELING
The observation and quantification of ecological attributes from lidar data are providing key insights into
ecological modeling studies. In forest ecosystems, accurate forest structure observations from lidar allow the
integration of vegetation structure information into solar
radiation modeling, vegetation dynamic modeling, and
biodiversity modeling across scales, which ultimately can
lead to a better understanding of forest ecosystems. Approaches like voxel-based transfer models and discrete
anisotropic radiative transfer models can be used to simulate accurate 3D radiation distribution [128], [253], [254]
and thereby investigate the influence of canopy structure
on the canopy light regime and photosynthetic partitioning [129]. Accurate 3D tree structural attributes also provide the opportunity to validate hypotheses in current
vegetation dynamic models. For example, Su et al. [256]
used TLS data to prove that canopy architecture displays
a strong spatial variability along climate gradients and to
reject the current hypothesis in vegetation dynamic models of a constant ratio between canopy height and canopy
size. Tao et al. [257] found that canopy height exhibits a
hump-shaped correlation with water availability, rejecting the prevailing assumption that canopy height follows
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
