other pixels yielding brighter values. Overall accuracies for
the FCI1 were over 95 percent in late May, including when ag-
riculture was pre-harvest and less dense than surrounding for-
est cover, and in December and January, when there was little
agriculture to confuse with forest cover. Results for the FCI2
were similar to FCI1 in August, December, and January im-
agery and were somewhat lower in late-May (Table 3). Trees
were the majority of the darkest pixels in the FCI outputs for
each scene, which allowed the user to reliably set a threshold
separating trees from other vegetative covers. Dark vegetation
features that were present in the imagery in small amounts,
such as dense agriculture, in some instances were confused
with forest cover in the FCI1 and FCI2 outputs. These pixel
values were low, similar to those of tree pixels.
Forest cover identification using the
NDVI
yielded low ac-
curacy in three of the four images. In the 27 May image,
NDVI
was roughly as accurate as the FCI1 output. Upon further
inspection of the output image, this was because most agricul-
tural fields were not yet planted or freshly planted with little
crop emergence, while leaves were already growing on trees.
These factors created a difference between the
NDVI
values of
trees and other vegetative land covers.
The Z test of statistically significantly different
KHAT
values
indicated the
NDVI
and FCI tests were statistically significantly
different from each other (Table 4) and in almost all compari-
sons, the FCI tests outperformed
NDVI
(Table 5).
Discussion
In this study, methods were created and tested to differentiate
forest from other vegetative land covers in a diverse land-use
scene. Forest delineation was possible by creating new vegeta-
tion indices that exploited the product of the red and red edge
bands (FCI1), and the red and first near infrared bands (FCI2) of
the WorldView-2 satellite. In this study, FCI1 and FCI2 delin-
eated forest cover throughout the year, including during times
when forest is spectrally similar to other vegetation covers.
NDVI
was less successful than the FCI1 for all imagery dates, and less
successful than the FCI2 in the August, October, and January
imagery. The results give an added value over previous studies
that mask out forest cover in imagery by offering a method that
does not require
HSI
or lidar data and can use archived
MSI
.
Potential Causes of Spectral Separability Between Trees and Other Leafy
Vegetation Cover
With FCI1 and FCI2, forest cover was consistently separable
from other vegetation cover as shown in Tables 3, 4, and 5.
Prior research comparing a canopy to an individual leaf has
shown that trees have lower overall reflectance than that of an
individual leaf because of radiation variations caused by illu-
mination angle, differing leaf orientations within the canopy,
shadows, and non-foliage backgrounds (Knipling, 1970).
Shadowing decreases the overall spectral curve (Zhang
et al
.,
2015) and foliage within trees and entire tree canopies creates
more shadows within trees than in agricultural crops, which
have less height and foliage. Zhang and Hu (2012) used high
spatial resolution imagery but found problems with identi-
fying trees because of shadows when using
NDVI
to identify
trees. For this research, shadowing within tree canopies was
used to distinguish between trees and other vegetation cover
types. Even in the winter when deciduous trees had lost their
leaves, shadowing among branches remained, allowing trees
to remain distinguishable from other land covers.
NDVI
accu-
racy was low for almost all imagery dates because
NDVI
values
Table 3. Producer's, user's, and overall accuracy for Tree versus Not Tree classifications using FCI1, FCI2, and NDVI on four
dates. The number of pixels in the ROIs differed by image due to differences in view angle, azimuth, and zenith.
Date
Class
Producer’s Accuracy (%)
User’s Accuracy (%)
Overall Accuracy (%)
FCI1
FCI2
NDVI
FCI1
FCI2
NDVI
FCI1
FCI2
NDVI
27 May 2012
Tree
95.1
99.4
88.1
99.1
74.1
94.7
97.8
86.4
95.2
Not Tree
99.5
78.2
99.7
97.0
99.5
93.0
05 Aug 2012
Tree
99.9
99.9
99.4
89.2
92.1
54.3
95.3
96.7
67.3
Not Tree
92.4
94.6
47.1
99.9
99.9
99.3
26 Oct 2014
Tree
99.8
99.9
99.9
72.0
62.8
61.2
84.9
77.1
75.9
Not Tree
75.6
62.7
60.1
99.8
99.9
99.9
18 Jan 2013
Tree
99.9
99.9
99.9
89.9
99.4
41.8
95.7
99.8
46.3
Not Tree
93.0
99.6
12.5
99.9
99.9
99.9
Table 5. Pairwise comparison of the error matrices. The results for the Z test of statistically significantly different
KHAT
values
indicated the
NDVI
and
FCI
tests were different in all cases with p < 0.05. Positive Z-scores indicated the first index of the pair
performed better in the comparison, negative Z-scores indicated the second of the pair performed better.
Date
FCI1 vs FCI2
FCI1 vs NDVI
FCI2 vs NDVI
Z
p Better index
Z
p
Better index
Z
p
Better index
27 May 2012
95.9
<0.001 FCI1
32.1
<0.001 FCI1
-65.5
<0.001 NDVI
05 Aug 2012
-14.4
<0.001 FCI2
139.7
<0.001 FCI1
150.6
<0.001 FCI2
26 Oct 2014
41.0
<0.001 FCI1
46.3
<0.001 FCI1
5.5
<0.001 FCI2
18 Jan 2013
-49.2
<0.001 FCI2
118.4
<0.001 FCI1
134.6
<0.001 FCI2
Table 4. Individual error matrix
KHAT
analysis results. A
Z-statistic of greater than 1.96 indicates the classification is
better than random.
Date
KHAT Variance Z-statistic p-value
27 May
2012
FCI1
0.95 0.97E-06
969.2 <0.001
FCI2
0.73 4.48E-06
344.6 <0.001
NDVI
0.90 2.08E-06
622.1 <0.001
05 Aug
2012
FCI1
0.90 2.08E-06
627.0 <0.001
FCI2
0.93 1.52E-06
753.7 <0.001
NDVI
0.40 10.7E-06
123.1 <0.001
26 Oct
2014
FCI1
0.71 4.95E-06
317.2 <0.001
FCI2
0.57 6.85E-06
215.9 <0.001
NDVI
0.54 7.16E-06
203.5 <0.001
18 Jan
2013
FCI1
0.91 2.81E-06
543.6 <0.001
FCI2
0.99 0.16E-06
2500.1 <0.001
NDVI
0.10 44.1E-06
15.0 <0.001
510
August 2018
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
