The quantity and allocation disagreements indicate that the
classifications correctly assign the total number of pixels to
each cover class (lower quantity disagreement), but classes
may be shown in the wrong location on the map (higher al-
location disagreement). Despite the slightly erroneous class al-
location, however, the proportion of impervious surface in the
classification result was close to that from the reference map.
Final impervious surface extraction results from the two
season images are shown in Figure 3. The area of impervious
surface extracted from the winter image (Figure 3B) is greater
than that from the summer image (Figure 3A). It is also clear
that most streets and sidewalks obscured by deciduous tree can-
opies in the summer image were correctly identified as imper-
vious surface in the winter image, e.g., the street shown in the
grey rectangle (Figure 3). In some gardens or parks, the propor-
tion of pervious surface in the winter image was less than that
in the summer image, as rooftops and pathways (impervious
surface) obscured by tree canopies were exposed and identified
as impervious surface, e.g., the grey ellipse in Figure 3.
Area proportions of the impervious surface class from
the summer and winter images were estimated by using the
method proposed by Stehman (2014), taking into account
the classification error. The results indicated that the area of
the impervious surface class accounted for 53.55 percent of
the entire image (with a standard error of 0.83 percent, and
51.93 ~ 55.18 percent at the 95 percent confidence interval),
whereas the estimated area proportion of the impervious
surface from the winter image was 72.01 percent (with a stan-
dard error of 0.82 percent, and 70.39 ~ 73.62 percent at the 95
percent confidence interval). The area of impervious surface
extracted from the winter image was greater than that from
the summer image, by 18.46 percent.
The analysis of the proportions of impervious surface in
the shaded area indicated that the impervious surface ex-
tracted from the shaded area in the summer image accounted
for 9.62 percent of the entire image, while impervious surface
extracted from shaded area in winter image accounted for
17.07 percent of the entire image. Although these statistics
were not adjusted for classification error, the difference in
area of impervious surface between the summer and winter
T
able
4. E
rror
M
atrix
and
A
ccuracy
E
stimates
(A
ll
in
P
ercent
)
of
S
hadow
C
lassification
R
esults
from
S
ummer
A
nd
W
inter
I
mages
in
B
eijing
A
rea
Summer Image Result
Winter Image Result
class
impervious
surface
pervious
surface
total
UA
class
impervious
surface
pervious
surface
total
UA
impervious
surface
51.13
2.41
53.54
95.50
impervious
surface
69.24
2.86
72.10
96.03
pervious
surface
6.50
39.96
46.46
86.00
pervious
surface
5.77
22.13
27.90
79.31
total
57.63
42.37
100
total
75.01
24.99
100
PA
88.72
94.31
PA
92.30
88.54
OA
91.09
OA
91.37
Validation sample size: 211 pixels for the summer image and 209 pixels for the winter image.
Figure 3. The impervious surface extraction results from (A) the summer image, and from (B) the winter image: 1, impervious surface;
2, pervious surface. The grey rectangle shows some road areas and the grey ellipse shows some rooftop and pathway areas which were
obscured by deciduous trees in summer and classified as pervious surface using summer image while extracted as impervious surface
using winter image.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
May 2016
343
