Table 1. WorldView-2 multispectral bands (DigitalGlobe, 2017).
Band Name
Band # *FWHM (nm)
Band Center (nm)
Coastal blue
1
400-450
425
Blue
2
450-510
480
Green
3
510-580
545
Yellow
4
585-625
605
Red
5
630-690
660
Red Edge
6
705-745
725
Near Infrared-1
7
770-895
835
Near Infrared-2
8
860-1,040
950
*FWHM = full width at half maximum.
The Worldview-2 images were orthorectified to the Uni-
versal Transverse Mercator coordinate system, calibrated to
at-sensor radiance values, and atmospherically corrected
to surface reflectance using the
FLAASH
module in the com-
mercial software, Environment for Visualizing Images (
ENVI
)
(Harris Corporation, Melbourne, Florida). All remote sensing
analyses were performed in
ENVI
.
Methodology
Forest Cover Indices and the Normalized Difference Vegetation Index
The Forest Cover Index 1 (FCI1) and Forest Cover Index 2
(FCI2) were developed to separate forest from other land cov-
ers using the following equations:
FCI1 = R
660
* R
725
(1)
FCI2 = R
660
* R
835
(2)
where R
660
, R
725
, and R
835
are reflectance in Red, Red Edge, and
Near Infrared-1 bands, respectively (Table 1). These equations
were developed after visually comparing spectra of tree cover
to those of other vegetation. While spectral profiles for trees
were similar to spectral profiles for other vegetative cover, the
red, red edge, and near infrared reflectance values for trees
were consistently lower. Traditional vegetation indices, sim-
ple ratios, and normalized difference indices do not exploit
the difference between trees and other land covers. We found
that multiplying reflectance in the pairs of bands emphasized
the difference between trees and other vegetative land covers.
The Normalized Difference Vegetation Index (
NDVI
) (Rouse
et al.
, 1974) provides a measure for vegetative vigor and is
frequently used to delineate green vegetation from non-vege-
tation and occasionally used to delineate trees from other land
covers (Peters
et al.
, 2002; White
et al.
, 2016; Bandyopadhyay
et al.
, 2013).
NDVI
was calculated using the following equation:
NDVI
= (R
835
– R
660
)/(R
835
+ R
660
)
(3)
where R
660
and R
835
are reflectance in Red and Near Infrared-1
bands, respectively (Table 1).
In the FCI1 and FCI2 images, darker areas corresponded to
trees and dense vegetation while brighter areas corresponded
to other vegetative features, including agriculture and grasses.
With this knowledge, a binary system was created to mask out
forest cover while retaining other land covers in the FCI1 and
FCI2 images using the steps below. The same steps were also
applied to the
NDVI
images.
1. Different land covers in the image were examined to de-
termine a user-defined threshold value that separated trees
from other vegetation covers.
2. A mask was built to separate out trees below the threshold
value.
3. A 200-pixel (800 m
2
) group minimum and 8-neighbor
sieve
were applied to the resulting image to remove small
clusters of other land cover pixels that should have been
identified as trees. The group minimum is the smallest size
grouping to keep. After reviewing the distribution of land
covers in the test site, a group minimum of 200 pixels was
chosen. An 8-neighbor sieve consists of all of the pixels
immediately adjacent to the original pixel and it removes
the entire group if the pixels are grouped in the same class.
4. A 3 × 3 pixel
clump
was executed to fill in small areas of
other land cover pixels that were incorrectly identified as
trees. This clump size was large enough to fill in errone-
ously identified pixels while retaining many individual
trees and most small tree groups.
5. The resulting output from Step 4 was used to apply a
mask to the original image, which resulted in an image
that masked all of the trees.
Accuracy Assessment
We selected 66 regions of interest (
ROI
) of various sizes within
the study area to assess the accuracy of these indices for
classifying trees from a broad range of other vegetative land
covers for each season of the year (Table 2). Approximately
100,000 pixels were included in these 66
ROIs
; however the
number varied slightly from date to date due to view angles
and clouds. The ‘Tree’ category included 10 conifer and 10
deciduous
ROIs
. The ‘Not-Tree’ category included 30
ROIs
of
annual crops (alfalfa, barley, corn, orchardgrass, rye, ryegrass,
soybean, turf grass, triticale, and wheat), 10 perennial grass
pastures, and 6 golf courses.
Table 2. Categories and number of
ROIs
selected for each tree
type and other vegetative land cover.
Pixels in ROI
ROI
27-May-12 5-Aug-12 26-Oct-14 18-Jan-13
Tree
Evergreen
1
4,513
3,992
4,301
2,790
Deciduous
2
34,492
31,106
32,900
21,350
Not Tree
Perennial
3
16,628
14,368
15,510
9,946
Annual
4
46,002
41,267
43,644
28,494
1
Evergreen trees are primarily coniferous trees including shortleaf
pines and Virginia pines.
2
Deciduous trees are primarily broadleaf trees including oaks,
maples, hickory, poplars, and sweetgums.
3
Perennial vegetation class included cool season grass pastures,
alfalfa, and turf grasses.
4
Annual crop class included corn, soybeans, and small grains (bar-
ley, oats, rye, triticale, wheat).
The accuracy assessment consisted of a comparison of tree
pixels to other vegetation pixels, which were recoded to “Not
Tree.” The
USDA
maintains records of land uses in the
BARC
,
which provided ground truth for the other vegetation pixels.
Error was calculated for FCI1, FCI2, and
NDVI
by using the
66
ROIs
for the four dates. User’s accuracy is a measure of the
probability that a pixel classified in an image represents that
category on the ground. Producer’s accuracy is the probability
a reference pixel is correctly classified. Overall accuracy looks
at the total number of correctly identified pixels compared to
the total number of reference pixels.
KHAT
is a comparison of
actual agreement between the computer classification of land
cover and reference data and chance agreement between the
computer classification and reference data (Congalton
et al
.,
1983). The
KHAT
statistic, its variance, and Z-statistic for a
single error matrix were computed to determine if agreement
between the remote sensing classification and the surface ref-
erence data was significantly better than random (Congalton
and Green, 2008). Non-vegetative
ROIs
were not selected.
508
August 2018
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
