of obtaining the proportion of area of impervious surface
directly from the map classifications, we adopted an unbiased
“error-adjusted” estimator (Stehman, 2013), which includes
the area of map omission error of the impervious surface
class and leaves out the area of map commission error of the
impervious surface class. Proportion of area of the impervious
surface class was estimated following the formulas provided
in Stehman (2014) in accordance with the sampling design
implemented in which the strata are different from the map
classes. The standard error of the estimated area proportion of
the impervious surface class was also calculated along with a
95 percent confidence interval.
Furthermore, the difference in the spatial distribution of
the impervious surface between summer and winter images
was generated. The difference distribution map included four
types of area, which corresponded to the four strata used in
stratified random sampling (as previously defined in the
Ac-
curacy Assessment
Subsection). The areal proportions of the
four strata in each study area were calculated, and the factors
contributing to the difference were analyzed. Finally, the
influence of the image acquisition season on the impervious
surface extraction was discussed.
Results and Discussion
Impervious Surface Mapping Results for the Beijing Area
Impervious surfaces were extracted from both summer and
winter images using the object-based method described in the
previous section and quantitatively compared. For accuracy as-
sessment, the stratified random sampling method was adopted
to collect samples. For evaluation of general urban land cover
classification from the images of both seasons, a total of 1,000
validation samples (pixels) was separately selected for each sea-
son image. After removing a few points that were near training
pixels to avoid the spatial autocorrelation of the samples, 990
pixels for the summer and 991 pixels for the winter image were
finally used. For the evaluation of the shadow classification
from both season images, 215 validation samples were selected
separately in each season image. After removing a few points
that were near training pixels, 211 pixels for the summer and
209 pixels for the winter image were finally used, respectively.
For accuracy assessment of final impervious surface extrac-
tion, 800 pixels were initially collected using stratified ran-
dom sampling method, and 795 pixels were finally used after
removing a few points that were near training pixels. Specifi-
cally, 421 pixels from the impervious surface and 374 pixels
from the pervious surface were used in the evaluation of the
result from the summer image. In evaluation of the winter
image result, 574 pixels of impervious surface and 221 pixels
of pervious surface were selected. Most of the 795 pixels for
validation were the same in terms of both location and class
label for the two seasons. As mentioned previously, however,
there were some samples with the same locations but differ-
ent class labels, which is accounting for about 22.26 percent
(177 pixels in total) of all validation samples.
Tables 2 and 3 summarize the estimated accuracies of the
general urban land cover classification of summer and winter
images, respectively. The classification results from the images
of both seasons achieved similar overall accuracies: 71.74 per-
cent for summer and 73.91 percent for winter. For both results,
the shadow class was accurately identified, with
PA
and
UA
greater than 88 percent (Tables 2 and 3). The water class was
also correctly recognized from both season images, with
PA
and
UA
greater than 87 percent. The classes tree and grass also show
relatively high
PAs
and
UAs
, but with the exception of a relative-
ly low
UA
for the summer image (61.54 percent) due to spectral
confusion between tree and grass. Different impervious surface
types (road, rooftop, and other impervious surface) show rela-
tively low
PAs
and
UAs
because of their similar spectral features.
However, these misclassifications might not affect the final
impervious surface extraction result as these classes will be
merged into the class of impervious surface. The class bare land
T
able
2. E
rror
M
atrix
and
A
ccuracy
E
stimates
(A
ll
in
P
ercent
)
of
G
eneral
U
rban
L
and
C
over
C
lassification
of
the
S
ummer
I
mage
in
B
eijing
A
rea
class
tree
grass
bare
1
water
road
rooftop other
2
shadow total
UA
tree
23.60
0.85
0.00
0.00
0.11
0.11
0.11
0.21
24.98
94.47
grass
2.98
5.29
0.00
0.00
0.17
0.00
0.17
0.00
8.60
61.54
bare
1
0.31
0.24
0.63
0.00
0.08
0.00
0.16
0.00
1.41
44.44
water
0.10
0.00
0.00
1.93
0.00
0.00
0.00
0.19
2.22
86.96
road
0.42
0.00
0.00
0.10
10.16
7.44
0.73
0.63
19.48
52.15
rooftop
0.09
0.09
0.09
0.00
3.32
8.92
2.18
0.00
14.70
60.65
other
2
0.51
0.00
0.00
0.00
0.51
3.99
5.32
0.31
10.64
50.00
shadow 0.87
0.00
0.00
0.00
0.22
0.87
0.11
15.89
17.96
88.48
total
28.88
6.47
0.72
2.04
14.56
21.32
8.77
17.23
100
PA
81.71
81.75
86.88
94.85
69.78
41.82
60.65
92.22
Overall accuracy: 71.74%. PA, producer’s accuracy; UA, user’s accuracy. 1, bare land; 2, other impervious surface. Validation sample size: 990 pixels.
T
able
3. E
rror
M
atrix
and
A
ccuracy
E
stimates
(A
ll
in
P
ercent
)
of
G
eneral
U
rban
L
and
C
over
C
lassification
of
the
W
inter
I
mage
in
B
eijing
A
rea
class
INS
tree
grass
bare
1
water
road rooftop other
2
shadow total
UA
INS
11.42
0.00
1.09
0.11
0.11
0.11
0.22
0.11
0.11
13.27
86.07
tree
1.00
8.41
1.30
0.00
0.00
0.00
0.00
0.00
0.60
11.32
74.34
grass
0.00
0.60
4.83
0.00
0.00
0.00
0.00
0.20
0.00
5.64
85.71
bare
1
0.00
0.00
0.00
0.55
0.00
0.00
0.21
0.04
0.00
0.80
68.42
water
0.00
0.00
0.00
0.00
2.75
0.00
0.00
0.20
0.00
2.95
93.33
road
0.29
0.00
0.00
0.20
0.10
9.87
5.77
0.68
1.56
18.47
53.44
rooftop 0.10
0.00
0.00
0.73
0.00
4.49
9.93
0.94
0.10
16.30
60.90
other
2
0.00
0.00
0.43
0.11
0.00
0.54
2.81
3.67
0.00
7.57
48.57
shadow 0.30
0.00
0.00
0.00
0.00
0.50
0.40
0.00
22.48
23.68
94.92
total
13.12
9.02
7.66
1.69
2.96
15.52
19.33
5.85
24.86
100
PA 87.04
93.30
63.14
32.37
93.02
63.62
51.35
62.83
90.43
Overall accuracy: 73.91%. PA, producer’s accuracy; UA, user’s accuracy. 1, bare land; 2, other impervious surface. Validation sample size: 991 pixels.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
May 2016
341
