Plate 3A-2 shows a deciduous tree covered street in sum-
mer where the impervious surface underneath deciduous tree
canopies is hidden. Plate 3B-2 shows that in winter all these
deciduous trees have lost their leaves, exposing the street
underneath where impervious surface is mixed with
NPV
and
its shadow.
NPV
, impervious surface, and shadow all contrib-
ute to the spectral feature of the area (Plate 3B-2). Therefore, a
new class, which is a mixture of impervious surface,
NPV
, and
shadow was defined for the winter image as
INS
(abbreviation
of mixed impervious surface,
NPV
and shadow). Because
INS
mainly consists of impervious surface, with little
NPV
and
shadow, the
INS
class is considered impervious surface in the
final impervious surface extraction result.
The common land cover classes in the images of both
seasons include tree, grass, road, rooftop and other impervi-
ous surface, water, bare land, and shadow. Besides,
INS
is only
included in the classification scheme for the winter image. In
addition, two classes: shaded impervious surface and shaded
pervious surface were selected in the shadow classification.
Object-based Impervious Surface Extraction
The object-based classification method was adopted for imper-
vious surface extraction from both summer and winter images.
As a prerequisite step of the object-based classification, image
segmentation was first conducted to generate appropriate im-
age objects. The hierarchical multilevel image segmentation
method proposed by Li et al. (2011) was adopted. The method
provides a simple and direct way for hierarchical segmenta-
tion of multispectral images and has been successfully used in
urban impervious surface extraction and damage assessment
(Li
et al.,
2011; Zhang
et al.,
2012). Multichannel watershed
transformation (Li and Xiao, 2007) is first applied to produce
an initial segmentation result, while the dynamics of water-
shed contours (Najman and Schmits, 1996) are then used
to set the value of each watershed line (or contour) with its
saliency. After the image of the contour dynamics is produced,
different threshold levels can be applied to the dynamics im-
age to produce multiple hierarchical segmentation results. The
adjustment of threshold values allows the analyst to explore
the various possible segmentations. It should be mentioned
that other multi-level image segmentation methods, such as
the multi-resolution segmentation method implemented in
eCognition
®
(Baatz and Schape, 2000; Benz
et al.,
2004), can
also be used to produce multi-level segmentation results.
The classification of the whole image and the classification
of the shaded area were conducted using segmentation results
at two different levels (e.g., Li
et al.,
2011). For the classifica-
tion of the entire image area, a coarse level of segmentation
result was used, where each shaded area was assumed to be
segmented as a single object. After the classification, the non-
shaded area was masked out and only the shaded area was fur-
ther classified. The shadow classification was conducted at a
fine level of segmentation result, at which a shaded area could
be further partitioned into several segments with different
classes. To determine optimal segmentation levels (i.e., contour
dynamics thresholds), we first manually delineated polygons
(regions) of different land cover types on the images and ex-
tracted dynamics values of watershed lines (contours) between
adjacent regions. We then calculated the average of contour
dynamics as the reference values for optimal thresholds. The
optimal segmentation levels for different season images were
determined by simultaneously considering reference threshold
values and visual analysis of segmentation results. For both
study areas, the coarse and fine level segmentation results were
generated using contour dynamics thresholds of 18 and 3 for
the summer images, and 15 and 3 for the winter images.
After image segmentation, the average values of each
spectral band for each image object from both coarse and fine
level segmentation results were calculated. The object average
images of the two segmentation levels were then used in ob-
ject-based classifications of the whole images and the shaded
areas identified in the previous step, respectively. The sup-
port vector machine (
SVM
) method (Vapnik, 1995) was used
as the classifier, as it has been widely used in land cover clas-
sification of both hyperspectral and multispectral images (e.g.,
Melgani and Bruzzone, 2004; Zhang
et al.,
2012). The
SVM
is
Plate 3. A demonstration of deciduous tree covered area. A portion of aerial image acquired in summer (A) June 2012, and its two parts
(B-1 and B-2) of winter aerial image (February 2012) in Beijing study area. B-1 shows the area mainly with grass and bushes underneath
the deciduous trees (NPV), as well as shadow. B-2 shows the street area with impervious surface, deciduous trees (NPV) and shadow.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
May 2016
339
