two season images (e.g., the blue area in grey ellipse shown
in Plate 6), and land cover changes due to construction. The
remaining 79.85 percent of the full image is the same in both
classification results from summer and winter images: 71.04
percent of impervious surface (shown as white in Plate 6) and
8.81 percent of pervious surface (shown as black in Plate 6).
Discussion
The object-based impervious surface extractions from both
summer and winter images achieved high accuracies, where
the overall accuracies from both summer and winter images
were greater than 92 percent (Tables 5 and 6) and quantity
and allocation disagreements were relatively small in the
two study areas. The results therefore validated the effective-
ness of the proposed method. When comparing the results
from the two season images in the two study areas, although
the results from the summer image produced slightly higher
accuracies (Tables 5 and 6), the estimated area proportion of
the impervious surface extracted from the winter image was
much greater than that from the summer image, by approxi-
mately 18 percent of the entire study area for the Beijing
area (53.55 percent from summer image versus 72.01 percent
from winter image) and for the Tianjin area (72.64 percent
from summer image versus 90.68 percent from winter image).
This is because 22.26 percent of the validation samples in
the Beijing area and 18.94 percent in the Tianjin area were
assigned different class labels for the summer and winter im-
ages due to the transition between leaf-on and leaf-off states.
Thus, with comparable accuracy, the use of the winter image
in impervious surface extraction significantly reduced the un-
derestimation error caused by the obscuring effect of decidu-
ous tree canopies in summer.
There are several contributing factors to the difference in
the impervious surface extraction between these two season
images. The main factor is the seasonality variation of decidu-
ous trees, mostly on the sides of roads, sidewalks, and around
buildings (e.g., Plate 5A, and 5B, and Plate 6). In summer,
deciduous tree canopies hide the impervious surface under-
neath. In winter deciduous trees drop their leaves, and the
impervious surface beneath the canopies is exposed or mixed
with
NPV
and shadow. Plant phenology therefore has a signifi-
cant effect on urban impervious surface extraction in temper-
ate regions. Other factors include illumination and view angle
variations and change in land cover in local areas between the
two seasons, and misclassification errors from both images
(Plate 5C through 5E, and Plate 6), which resulted in only
small amount of difference.
The findings in our study are different from most previous
studies conducted in temperate regions. The previous studies
concluded that impervious surface extraction from sum-
mer images produced considerably higher accuracies than
those from images acquired in winter and other seasons (e.g.,
Wu and Yuan, 2007). The reasons for the different findings
between those previous studies and ours can be summarized
as follows. First, the previous studies used medium resolu-
tion images (e.g., Landsat
TM
/
ETM+
) (Wu and Yuan, 2007;
Weng and Hu, 2008; Hu and Weng, 2009; Weng
et al.,
2009;
Plate 6. The difference map of object-based impervious surface extraction between summer image and winter image in Tianjin study
area. 1, the area recognized as pervious surface from the summer image and recognized as impervious surface from the winter image;
2, the area recognized as impervious surface from the summer image and recognized as pervious surface from the winter image; 3, the
area with the impervious surface class label; 4, the area with the pervious surface class label. The grey ellipse shows some different
areas mainly caused by variations in the view angle and the illumination between the two season images.
346
May 2016
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
