Sung and Li, 2011), whereas we used
VHR
images. In particu-
lar, the Worldviw-2 images used have eight spectral bands,
which may provide more spectral and spatial information.
Second, the shaded area was further classified in this study,
whereas in the previous studies, shadow was considered as
an endmember and not classified further (Weng and Hu, 2008;
Sung and Li, 2011). In the Beijing area, the shaded area in the
winter image was approximately 5.72 percent greater than
that in the summer image. The impervious surface extracted
from the shaded area was 9.62 percent in the summer image
and 17.07 percent in the winter image. In the Tianjin area,
the shaded area in the winter image was approximately 14.95
percent greater than that in the summer image. The impervi-
ous surface extracted from the shaded area was 12.67 percent
in the summer image and 24.82 percent in the winter image.
By further classifying the shaded area, the underestimation
error was also further reduced. Third, the new class, the class
INS
(a mixture of impervious surface,
NPV
, and shadow) is
separable from the other classes. Although it may be confused
with those mixed areas of bare land, vegetation, ice, and
shadow, the mixed areas may be very small in well-planned
urban areas. Fourth, there are still evergreen trees and bushes
with healthy foliage in the study areas in winter which shows
a distinct spectral signature from the impervious surface. The
existence of evergreen trees and bushes with foliage in winter
in the study areas is largely attributed to some national or
industrial urban greening standards and city planning regula-
tions (e.g., China Academy of Urban Planning and Design,
1997; Shanghai Construction and Traffic Management Com-
mittee, 2007). Under these standards and regulations, vegeta-
tion in urban areas should be planted with a multilayer struc-
ture or deciduous trees and evergreen trees (also bush and
grass) should be inter-planted. Thus, there are some evergreen
species under or among the deciduous trees causing bare land
to remain unexposed even in winter. Moreover, the stan-
dards also require that trees planted along roads should be
deciduous in cold and snowy areas. The impervious surface
in the roads (and streets and sidewalks) obscured by decidu-
ous tree canopies in summer could therefore be exposed and
detectable in winter. Thus, the use of winter images could
also be effective for impervious surface extraction in similar
urban areas in temperate regions.
Given that the summer image also has some advantages for
the extraction of impervious surface, e.g., distinct spectral sig-
natures between impervious surface and vegetation, develop-
ing new and more sophisticated methods for a combined use
of summer and winter images in impervious surface extrac-
tion is a future study topic.
Conclusions
We extracted impervious surface from WorldView-2
VHR
im-
ages of summer and winter over Beijing and Tianjin urban
areas using a hierarchical object-based classification method.
The results were then quantitatively compared and analyzed.
The results showed that the proposed object-based extraction
method produced high accuracies for the two season images
from the two study areas and is suitable for impervious sur-
face extraction using
VHR
images. The results also demonstrat-
ed that the winter image produced accurate results of imper-
vious surface extraction, similar to that of the summer image.
However, the impervious surface area extracted from the
winter image is greater than that extracted from the summer
image in the Beijing and Tianjin areas, which are both approx-
imately 18 percent of the entire study area. Further analysis
also indicated that the difference in impervious surface area
between the two season images mainly occurred for streets,
pathways, and walkways, which were obscured by deciduous
tree canopies in summer and exposed as impervious surface
in winter. Thus, the difference mainly resulted from seasonal
variation of deciduous trees, which has a significant effect
on impervious surface extraction in temperate regions. Based
T
able
5. E
rror
M
atrix
and
A
ccuracy
E
stimates
(A
ll
in
P
ercent
)
of
I
mpervious
S
urface
E
xtraction
R
esults
from
S
ummer
and
W
inter
I
mages
in
B
eijing
A
rea
.
P
roducer
’
s
A
ccuracy
(PA), U
ser
’
s
A
ccuracy
(UA)
and
O
verall
A
ccuracy
(OA)
are
R
eported with
S
tandard
E
rrors
(SE)
in
P
arentheses
Summer image result
Winter image result
class
impervious
surface
pervious
surface
total
UA
class
impervious
surface
pervious
surface
total
UA
impervious
surface
50.92
3.52
54.44
93.53
(1.16)
impervious
surface
68.60
3.74
72.34
94.83
(0.88)
pervious
surface
2.63
42.93
45.56
94.23
(1.18)
pervious
surface
3.41
24.25
27.66
87.68
(1.89)
total
53.55
46.45
100
total
72.01
27.99
100
PA
95.09
(0.95)
92.42
(1.26)
PA
95.27
(0.69)
86.65
(1.99)
OA
93.85(0.83)
OA
92.85(0.82)
Validation sample size: 795 pixels.
T
able
6. E
rror
M
atrix
and
A
ccuracy
E
stimates
(A
ll
in
P
ercent
)
of
I
mpervious
S
urface
E
xtraction
R
esults
from
S
ummer
A
nd
W
inter
I
mages
in
T
ianjin
A
rea
.
P
roducer
’
s
A
ccuracy
(PA), U
ser
’
s
A
ccuracy
(UA),
and
O
verall
A
ccuracy
(OA)
are
R
eported with
S
tandard
E
rrors
(SE)
in
P
arentheses
Summer image result
Winter image result
class
impervious
surface
pervious
surface
total
UA
class
impervious
surface
pervious
surface
total
UA
impervious
surface
70.77
3.45
74.22
95.35
(0.76)
impervious
surface
86.36
1.65
88.01
98.13
(0.46)
pervious surface
1.87
23.91
25.78
92.76
(1.62)
pervious sur-
face
4.32
7.67
11.99
64.00
(4.00)
total
72.64
27.36
100
total
90.68
9.32
100
PA
97.43
(0.56)
87.39
(1.82)
PA
95.24
(0.50)
82.34
(3.67)
OA
94.68(0.71)
OA
94.03(0.63)
Validation sample size: 998 pixels.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
May 2016
347
