After all cottonwood pixels were identified and extracted,
the highest classification accuracy of the rest of the image was
obtained in the texture-spectral space using a pixel-based
SVM
and a texture-window size of seven pixels.
Overall accuracy of this semi object-based method
achieved 76 percent, an increase of about 15 percent com-
pared to the best results among all pixel-based classifications
(67.6 percent). Especially for the cottonwood class, overes-
timation was reduced significantly, as there were far fewer
saltcedar and Sophora pixels misclassified as cottonwood in
this object-based result (Table 4).
Discussion
Maps for saltcedar and cottonwood that reveal their histori-
cal, current, and future distribution patterns are essential to
understanding the dynamic relationship between these two
species. Nagler
et al
. (2005) successfully identified cotton-
wood and willow trees within dense saltcedar stands along
the lower Colorado River using aerial photos. However, their
results were produced by time-consuming visual interpreta-
tion rather than automated machine-based classification.
Pixel-based Classification Using Spectral Images
Due to the spectral similarity of their foliage, it is difficult to
separate saltcedar from other riparian species in multispectral
remote sensing images. In a recent study along the Middle Rio
Grande River, confusion among several riparian plant spe-
cies, including saltcedar and cottonwood, were encountered
in classification using a summer-acquired aerial photograph
(Akasheh
et al
., 2008). Manual recoding was required to
correct the misclassifications. Similarly, Carter
et al
. (2009)
classified saltcedar along the Colorado River using a Quick-
Bird image obtained during the summer. Although overall
classification accuracy achieved over 90 percent, commission
error for saltcedar still was considerable. The large amount
of misclassification between saltcedar and cottonwood in our
pixel-based classification results also confirmed the previ-
ous assumption that spectral data alone is insufficient to
distinguish cottonwood and saltcedar in summertime images.
Much of the overestimation of cottonwood occurred where
saltcedar cover was relatively low (Plate 1). Most likely, this
was caused by the bare background soils in saltcedar stands.
The presence of wet soil would lower the reflectance value of
the corresponding pixels, making it match closer to the dark
cottonwood trees in the image. Similarly, many small frag-
mented saltcedar patches within large Sophora stands were
not detected. Therefore, spectral reflectance again, proves
T
able
4. C
onfusion
M
atrix
of
the
S
emi
-O
bject
-B
ased
C
lassification
R
esult
Reference Data
Classified Data
Cottonwood Saltcedar
Sophora
Soil
Road
Shadow
Total
Cottonwood 19
8
4
1
0
0
32
Saltcedar
24
175
9
2
0
0
210
Sophora
5
36
78
8
1
1
129
Soil
0
2
1
60
2
0
65
Road
0
0
0
2
4
0
6
Shadow
1
1
0
0
0
7
9
Total
49
222
92
73
7
8
451
Overall Accuracy = 343/451 = 76%
Producer’s Accuracy
User’s Accuracy
Cottonwood = 19/49 = 38.8%
Cottonwood = 19/32 = 59.3%
Saltcedar = 175/222 = 78.8%
Saltcedar = 175/210 = 83.3%
Sophora = 78/92 = 84.8%
Sophora = 78/129 = 60.5%
Soil = 60/73 = 82.2%
Soil = 60/65 = 92.3%
Road = 4/7 = 57.1%
Road = 4/6 = 66.7%
Shadow = 7/8 = 87.5%
Shadow = 7/9 = 77.8%
T
able
5. C
onfusion
M
atrix
of
SVM C
lassification
R
esult
U
sing
S
pectral
-T
exture
I
mage with
a
W
indow
S
ize
of
S
even
P
ixels
Reference Data
Classified Data
Cottonwood Saltcedar
Sophora
Soil
Road
Shadow
Total
Cottonwood 31
94
40
8
1
1
175
Saltcedar
16
114
8
10
1
1
150
Sophora
1
11
42
1
0
0
55
Soil
0
2
2
52
1
0
57
Road
0
0
0
2
4
0
6
Shadow
0
1
1
0
0
6
8
Total
49
222
92
73
7
8
451
Overall Accuracy = 249/451 = 55.2%
Producer’s Accuracy
User’s Accuracy
Cottonwood = 31/49 = 63.3%
Cottonwood = 31/175 = 17.7%
Saltcedar = 114/222 = 51.4%
Saltcedar = 114/150 = 76%
Sophora = 42/92 = 45.7%
Sophora = 42/55 = 76.4%
Soil = 52/73 = 71.2%
Soil = 52/57 = 91.2%
Road = 4/7 = 57.1%
Road = 4/6 = 66.7%
Shadow = 6/8 = 75%
Shadow = 6/8 = 75%
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
October 2015
801
