Incorporating Crown Shape Information
for Identifying Ash Tree Species
Haijian Liu and Changshan Wu
Abstract
Identifying ash trees from other common deciduous trees
is challenging due to subtle spectral differences of foliage
among species. Although many researchers have integrated
lidar-derived tree height and crown size metrics to improve
tree species classification accuracy, these simple biophysical
attributes provide inadequate explanatory power in distin-
guishing ash trees (Fraxinus, spp.) in urban ecosystems. To
address this issue, shape-related features, including crown
shape index (
SI
) and coefficient of variation (
CV
) of crown
height, were extracted from lidar data, and fused with treetop-
based spectra for ash tree species identification in Milwau-
kee City, Wisconsin, United States. Analysis results indicate
shape features including
SI
and
CV
play a big role in improv-
ing the accuracy for ash tree identification. Specifically,
Fusion of
CV
and treetop-based spectra improved the overall
accuracy from 81.9 percent to 89 percent, and McNemar tests
indicated the differences in accuracy between
CV
fusion and
tree height fusion was statistically significant (p = 0.016).
Introduction
Ash trees (
Fraxinu
s
spp
.) have become popular street trees for
decades in many urban areas of the United States, especially
in the eastern and middle parts of the country (Harlow, 1991;
MacFarlane and Meyer, 2005). They are widely planted due to
the tolerance of a wide range of soil and climate conditions,
large crowns for road greening, and commercial purposes.
However, ash trees are vulnerable to injury from the exotic
bark beetles like the emerald ash borer (
Agrilus Planipennis
Fairnaire
,
EAB
), which were introduced to North America
from Asia in the 1990s and first detected in the metropolitan
Detroit, Michigan in 2002 (Cappaert
et al
., 2005; Poland and
McCullough, 2006).
EAB
has killed tens of millions ash trees
in at least 18 states in the United States by 2011 (Flower
et
al
., 2013; Pugh
et al
., 2011) and was predicted to expand to 25
states by 2019 (Kovacs
et al
., 2010). Previous studies suggest
all ash species in eastern North American are susceptible to
EAB
due to their related genetic properties (Herms
et al
., 2004;
MacFarlane and Meyer, 2005), and ash trees infested by
EAB
experience mortality within three to five years, but the sign
of
EAB
infestation is not obvious before the primary visual
symptom occurs
(Pontius et al
., 2008). Therefore, finding ash
tree species among urban broadleaved trees is beneficial for
local residents to focus their efforts on checking and monitor-
ing the health status of ash trees; knowing the number and the
distribution of ash trees is helpful for multiple government
agencies to establish their feasible budget forecasts for ash
tree protection; and further controlling the spread of
EAB
is
essential for community to maintain species diversity (Pon-
tius
et al
., 2008; Sivyer, 2010).
In the field of remote sensing, tree species classification
mainly depends on spectral signatures of tree crowns (Ghosh
et al
., 2014), which correspond to their bio-chemical composi-
tions. Especially, the pure hyperspectral reflectance extracted
in laboratory conditions has the ability of capturing the
subtle difference of tree species due to hundreds of relatively
narrow but continuous spectral bands (Petropoulos
et al
.,
2012). Thus, a large amount of studies applied hyperspectral
information for tree species classification and disease iden-
tification. George
et al
. (2014) discriminated six broadleaved
evergreen and conifer forest tree species in western Himalaya
through employing the Earth Observation-1 (
EQ-1
) Hyperion
hyperspectral imagery with 242 bands. Both Xiao
et al
. (2004)
and Alonzo
et al
. (2013) mapped urban trees using airborne
visible/infrared imaging spectrometer (AVIRIS) imagery with
224 bands. Zhang and Qiu (2012) identified 40 urban spe-
cies in Dallas, Texas, United States through employing the
airborne imaging spectrometer for application (
AISA
) hyper-
spectral images with 492 spectral bands. In addition, Clark
et
al
. (2005) discriminated seven tropical rainforest tree species
through applying the HYperspectral digital imagery collec-
tion Experiment (
HYDICE
) data with 210 bands. Although
hyperspectral data has great contribution to tree species clas-
sifications, the similarity of spectral signatures of different
tree species were observed in deciduous species (Xiao
et al
.,
2004), and mixed pixel problem might exaggerate the spec-
tral confusion between species (Liu
et al
., 2011). Therefore,
deciduous tree species classification with only passive optical
imagery is still a challenging task (Fassnacht
et al
., 2016).
In addition to spectral information, crown structural fea-
tures extracted from lidar data have also been employed for
tree species classification due to their advantage of charac-
terizing biophysical properties. Ørka
et al
. (2009) extracted
height and intensity features from airborne laser scanner
(
ALS
) and discriminated Norway spruce and birch trees in
Norway, and resulted in an overall classification accuracy of
88 percent. Li
et al
. (2013) successfully classified four tree
species with an overall accuracy of 77.5 percent based on
several lidar derived structure features, such as 3D texture,
foliage clustering degree, foliage clustering scale, and gap
distribution of individual trees. Korpela
et al
. (2010) achieved
the best accuracy of 88 percent-90 percent in classification of
Scots pins, Norway spruce, and birch, using intensity vari-
ables, but spruce and birch were classified with the lowest
accuracy. In addition, Reitberger
et al
. (2006) extracted the
salient features including the outer tree geometry, internal
geometrical tree structure, and the intensity-related tree
structures from a full-waveform lidar data to separate the
Haijian Liu is with the Institute of Remote Sensing and Earth
Sciences, Hangzhou Normal.
Changshan Wu is with the University/Zhejiang Provincial
Key Laboratory of Urban Wetlands and Regional Change;
and also with the Department of Geography, University of
Wisconsin-Milwaukee (
.
Photogrammetric Engineering & Remote Sensing
Vol. 84, No. 8, August 2018, pp. 495–503.
0099-1112/18/495–503
© 2018 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.84.8.495
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
August 2018
495
