mineral composition. Reflectance spectra of litter (non-
photosynthetic vegetation) and soils are often similar and
differ only in amplitude in the visible and near infrared
wavelengths (Biard and Baret, 1997; Daughtry
et al
., 2004).
In the shortwave infrared, the spectra of dry litter have ab-
sorption features associated with cellulose and lignin that
are absent in spectra of soils and green vegetation (Nagler
et al.,
2000). Soils and rocks have absorption features as-
sociated with minerals that are absent in green vegetation
and litter (Kokaly and Clark, 1999; Serbin
et al.
, 2009).
4. With changing sun or view directions, the proportions of
sunlit vegetation, shadowed vegetation, sunlit background,
and shadowed background viewed by the sensor will vary,
which could aid in distinguishing between forest and
other vegetative land covers. Canopy radiance as measured
by the sensor will change even when the spectral proper-
ties of canopy elements, canopy structure, and background
are constant (Ranson
et al
., 1985). Several satellite sensors,
e.g. WorldView-2 and -3, Hyperion, and
SPOT
, provide op-
tions for off-nadir view angles which must be considered
when analyzing their images.
5. Solar radiation is partially absorbed or scattered by
molecules and aerosols as it passes through the Earth’s
atmosphere affecting the quality of remotely sensed im-
ages of the surface. Atmospheric corrections are neces-
sary to quantitatively analyze multi-date images, which
could also aid in distinguishing between forest and other
vegetative covers. For example, the Fast Line-of-sight
Atmospheric Analysis of Hypercubes (
FLAASH
) algorithm
provides a physics-based approach to atmospheric correc-
tion (Adler-Golden
et al.,
1998).
Detecting Tree Cover in Multispectral Imagery
Conventional forest inventory methods often include field
measurements (e.g., diameter at breast height, crown diam-
eter, and crown height) of many trees within the regions of
interest. Manual interpretation of aerial imagery has also
been used extensively for forest inventory (Wang
et al.,
2004).
However, both techniques are labor intensive. Advances in
the spatial resolution of remotely sensed images and pattern
recognition algorithms have provided new opportunities for
automated detection of trees.
Hyperspectral imagery (
HSI
) and Light Detection and Rang-
ing (lidar) data have been used successfully for tree detection,
species identification, and tree crown delineation (Ke and
Quackenbush, 2011; Graves
et al.
, 2016; Chen
et al.
, 2006).
Unfortunately, both
HSI
and lidar data are not widely avail-
able and are often costly to acquire for large areas. Recently,
advanced multispectral imagery (
MSI
) with fine spatial resolu-
tion has become available commercially and may provide a
low-cost alternative to
HSI
and lidar.
Prior research has tested identification of trees in
MSI
in
mixed land cover areas and while results have been promis-
ing, confusion between forest cover, including individual
trees, and select land covers persists. Yuan
et al.
(2005), Ye
et
al.
(2014), Fu
et al.
(2014), and Gonzalez-Alonso and Cuevas
(1993) included forest in testing with Landsat to map differ-
ent land covers, while Wu
et al.
(2017) and Akar
et al.
(2017)
used WorldView-2, as explained in the following paragraphs.
Yuan
et al.
(2005) used a combined supervised-unsuper-
vised training approach to map land covers in Minneapolis
- St. Paul, MN using Landsat Thematic Mapper and Landsat
Enhanced Thematic Mapper imagery and reported forest
cover user’s and producer’s accuracies of 90 percent and
above during their time series. While accuracies were high,
they found that forest cover was most commonly confused
with agriculture followed by wetland and urban land covers
and less commonly confused with water and grass. They also
did not identify individual trees, as the resolution of Landsat
is inadequate for this purpose.
Ye
et al.
(2014) developed a forest index using Landsat
Thematic Mapper and Landsat Enhanced Thematic Mapper
comparing forested to non-forested lands and the user’s and
producer’s accuracies were over 95 percent accurate, but
they did not identify individual trees. Fu
et al.
(2014) used
Landsat 5 and 7 to map urban, agriculture, rangeland, barren,
ice/snow, and forest area using a supervised classification.
User’s and producer’s accuracies for forest cover across the
time series were between 77 percent and 95 percent. Most of
the confusion occurred between forest and rangeland with no
ability to identify individual trees.
Gonzalez-Alonso and Cuevas (1993) applied regression
methods to Landsat to examine conifers, green oak trees,
various crop covers, fallow, and river. Their tree cover results
were highly accurate at over 95 percent correctly classified,
but they did not identify individual trees.
Wu
et al.
(2017) and Akar
et al.
(2017) used WorldView-2
to classify land covers and included forest in their land cover
testing. Wu
et al.
(2017) applied a support vector machine to
WorldView-2 and WorldView-2 fused with lidar to classify an
image that contained buildings, trees, road/parking lots, grass-
lands, and bare soils. User’s and producer’s accuracies were
between 83 percent and 85 percent for trees for WorldView-2
alone, which increased when fused with lidar. Confusion with
WorldView-2 alone happened between trees and most land
covers. Individual trees were often confused with grassland.
Akar
et al.
(2017) merged WorldView-2
MSI
with panchro-
matic imagery and used a support vector machine to classify
an image that contained forest, rangeland, and other land
covers. They found that user’s and producer’s accuracies for
forest were 86 percent and 89 percent and with the most
confusion being with rangeland. Individual trees were among
the pixels misclassified as rangeland.
Overview of Automated Tree Detection Methods
Ke and Quackenbush (2011) defined tree detection algorithms
as procedures for finding treetops or locating trees with-
out necessarily delineating tree-crown outlines. They also
identified four broad categories of automatic tree detection
methods: (1) local maximum filtering, (2) image binarization,
(3) scale analysis, and (4) template matching. Many of these
methods rely on imagery from aerial sensors or cameras,
which is usually higher in spatial resolution but more dif-
ficult to acquire on a repeat basis than satellite imagery. The
methods varied in their ability to correctly identify trees and
reliably differentiate between trees and other vegetation.
For the local maximum filtering method, the maximum
pixel value within a moving window is assumed to represent
the sunlit treetop. First, vegetative land covers are identified
using a supervised or unsupervised classification and then
local maxima of various combinations of visible and near
infrared bands within the vegetative land cover are used to de-
lineate trees. Pouliot
et al
. (2005) used color infrared imagery
with pixel sizes of 6 cm. The percent correct in the color infra-
red imagery ranged from 40.6 to 90.4 percent for regenerating
cutover trees depending on the tree cover (Pouliot
et al
., 2005).
Image binarization is a straightforward thresholding ap-
proach where grayscale is converted into black-and-white
imagery and pixels on one side of the threshold represent pix-
els of interest, while pixels on the other side of the threshold
are background. When used to separate trees from other land
covers, the contrast between trees and background can vary
within an image, which can cause only partial success when
one threshold value is applied throughout the image (Ke and
Quackenbush, 2011). Pitkanen (2001) performed image smooth-
ing on grayscale digital imagery (0.5 m pixel size) and separated
trees from the rest of the image using a binarization to keep the
maximum values in the image. He achieved overall accuracies
of 50 to 96 percent depending on the threshold method.
506
August 2018
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
