has low accuracies in both results, especially for the classifica-
tion result from the winter image, due to its similar spectral
signature with the classes: road, rooftop and other impervious
surface. The class bare land is rare representing less than 1 per-
cent of the study area, and therefore has little impact on overall
accuracy. The class
INS
which only appeared in the winter
image, had high
PA
and
UA
, which indicated that the class
INS
showed high spectral separability with the other classes.
A local area of general urban land cover classification from
the images of both seasons is shown in Plate 4. In general,
the classification maps are homogeneous. From the Plate, it
is also clear that some rooftops were misclassified as road,
other impervious surface and bare land, and some trees were
misclassified as grass in the classification result of the sum-
mer image (Plate 4C). In the classification result of the winter
image, the misclassification of rooftop as road and other im-
pervious surface was less significant (Plate 4D). By comparing
the classification results from both season images, most roads
and some rooftops obscured by tree canopies were classified
as vegetation in the classification result of the summer image
(Plate 4C), whereas these areas (road and rooftop) obscured by
deciduous trees were correctly classified as
INS
in the clas-
sification result of the winter image (Plate 4D, i.e., the areas
showed in the red rectangle and ellipse), since the impervious
surfaces were fully or partially exposed.
The statistics of the classes in both classification results in-
dicate that shadow accounts for approximately 17.96 percent
of the entire area in the summer image, and for approximately
23.68 percent of the entire area in the winter image. Table 4
shows the shadow classification accuracies for both season
images. The results from both images achieved almost the
same overall accuracy (higher than 91 percent). In terms of
the accuracy of individual classes, both impervious surface
and pervious surface in each shadow classification result
achieved high accuracies, e.g., the
PA
and
UA
of the two class-
es in both results were generally greater than 80 percent. After
examining the two shadow classification results, the misclas-
sifications mainly occurred in the boundary region between
the pervious surface (e.g., tree) and impervious surface. For
the winter image result, some
INS
areas (impervious surface)
were misclassified as tree (pervious surface).
The accuracies of the final impervious surface extraction re-
sults from the summer and winter images are shown in Table
5. The results from both season images achieved similar and
high overall accuracies: 93.85 percent for summer and 92.85
percent for winter image. In terms of the accuracy of individu-
al classes,
PAs
and
UAs
of both impervious surface and pervi-
ous surface from both images were greater than 87 percent.
Quantity and allocation disagreements (Pontius and Mil-
lones, 2011) of the result from the summer image were 0.89
percent and 5.26 percent, and those from the winter image
result were 0.33 percent and 6.82 percent, respectively. In
general, the overall disagreement was dominated by the al-
location disagreement between the reference map and the
results from the summer and winter images, whereas the
quantity disagreements were very low (less than 1 percent).
Plate 4. A portion of (A) the summer image and its (C) object-based classification result; and a portion of (B) the winter image, and its
(D) object-based classification result. 1, INS; 2, tree; 3, grass; 4, bare land; 5, water; 6, road; 7, rooftop; 8, other impervious surface; 9,
shadow. The red rectangle show some road areas and the red ellipse show some rooftop areas which were obscured by deciduous trees
in summer and classified as tree using summer image while classified as INS using winter image.
342
May 2016
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING