distance of zero (the closest distance possible) were selected
as the candidates for cottonwood crown objects since cot-
tonwood was distributed only in the immediate vicinities of
their own shadows, and these cottonwood trees were the only
features tall enough to cast shadows in the image.
While shadows can be cast only to one side of the cotton-
wood trees, many non-cottonwood objects, such as back-
ground soils, isolated saltcedar, or small patches of Sophora
in the vicinity of the sun-lit side of these trees, also were
included in some of the candidate objects due to their prox-
imity to shadow pixels. These non-cottonwood pixels among
the candidate objects can be properly classified with spectral-
texture data using the pixel-based classification method.
Therefore, in order to separate these non-cottonwood features,
a pixel-based classification result (using techniques previous-
ly mentioned in the Methods Sections) was first overlaid onto
the candidate objects. Then, in the overlaid image, a pixel
was labeled as cottonwood if it both was classified as cotton-
wood in the pixel-based result and fell within the boundary
of the candidate objects. The
SVM
classification result with the
highest user’s accuracy (lowest overestimation) of cottonwood
was selected for this overlaying process since only cotton-
wood needed to be extracted in this step and cottonwood
overestimation was the major error in pixel-based classifica-
tion results (as will be discussed in the Results Section).
In a final step, a pixel-based classification was performed on
the rest of the image after all cottonwood pixels were extracted.
Results
Pixel-based Classification Using Spectral Images
The overall accuracy of the pixel-based classification of the
multispectral image was 54.5 percent and 55.4 percent for
MLC
and
SVM
, respectively. Both classifiers had relatively low
overall accuracy.
Confusion matrices (Congalton, 1991; Story and Congalton,
1986) of the classification suggested that the biggest source of
error for both classifiers was the high commission of cotton-
wood (overestimation), caused mainly by misclassification of
saltcedar as cottonwood. In addition, overestimation occurred
in Sophora stands as well as a large number of pixels were
classified as cottonwood while they were identified as Sopho-
ra in reference data (Tables 2 and, 3). Another major source of
error was misclassification of saltcedar as Sophora. Both
MLC
and
SVM
results showed that most of these errors occurred
where small, isolated saltcedar individuals were distributed
in large Sophora stands.
T
able
2. C
onfusion
M
atrix
of
MLC C
lassification
R
esult
U
sing
M
ultispectral
I
mage
Reference Data
Classified Data
Cottonwood Saltcedar
Sophora
Soil
Road
Shadow
Total
Cottonwood 37
80
26
0
0
4
147
Saltcedar
2
102
6
0
0
3
113
Sophora
9
38
60
31
2
0
140
Soil
1
1
0
42
1
0
45
Road
0
0
0
0
4
0
4
Shadow 0
1
0
0
0
1
2
Total
49
222
92
73
7
8
451
Overall Accuracy = 246/451 = 54.5%
Producer’s Accuracy
User’s Accuracy
Cottonwood = 37/49 = 75.5%
Cottonwood = 37/147 = 25.2%
Saltcedar = 102/222 = 45.9%
Saltcedar = 102/113 = 90.3%
Sophora = 60/92 = 65.2%
Sophora = 60/140 = 42.9%
Soil = 42/73 = 61.6%
Soil = 42/45 = 93.3%
Road = 4/7 = 57.1%
Road = 4/4 = 100%
Shadow = 1/8 = 12.5%
Shadow = 1/2 = 50%
T
able
3. C
onfusion
M
atrix
of
SVM C
lassification
R
esult
U
sing
M
ultispectral
I
mage
Reference Data
Classified Data
Cottonwood Saltcedar
Sophora
Soil
Road
Shadow
Total
Cottonwood 28
88
46
0
0
1
163
Saltcedar
12
99
2
0
0
0
113
Sophora
5
23
43
4
1
0
76
Soil
2
10
1
69
2
0
84
Road
0
0
0
0
4
0
4
Shadow 2
2
0
0
0
7
11
Total
49
222
92
73
7
8
451
Overall Accuracy = 250/451 = 55.4%
Producer’s Accuracy
User’s Accuracy
Cottonwood = 28/49 = 57.1%
Cottonwood = 28/163 = 17.2%
Saltcedar = 99/222 = 44.6%
Saltcedar = 99/113 = 87.6%
Sophora = 43/92 = 46.7%
Sophora = 43/76 = 56.6%
Soil = 69/73 = 94.5%
Soil = 69/84 = 82.1%
Road = 4/7 = 57.1%
Road = 4/4 = 100%
Shadow = 7/8 = 87.5%
Shadow = 7/11 = 63.6%
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
October 2015
799
