Geography Reference
In-Depth Information
This study combined change vector analysis (CVA) and PCC to detect LULC
changes based on the object level. CVA is a widely used unsupervised change detec-
tion method that uses multichannel images (Malila
1980
). CVA can process any
number of image channels and can produce detailed change detection information
based on the channel change vector obtained by subtracting corresponding image
channels of two images acquired at different times. In this study, CVA was used to
detect changed objects based on selected features instead of pixel values of image
channels. Given that object-oriented CVA can process any number of features, it
was suitable for object-oriented change detection using PolSAR images. In object-
oriented CVA, change detection is based on feature change vectors (FCVs), which
are obtained by subtracting corresponding feature vectors of an image object in two
images acquired at different dates. Two co-registered images, image (
t
1
)andimage
(
t
2
), are assumed to be acquired over the same area at different times
t
1
and
t
2
.If
k
features are extracted from an image object, the feature vectors of the image object
in the two images are given by
X
D
(
x
1
,
x
2
, :::,
x
k
)
T
and
Y
D
(
y
1
,
y
2
, :::,
y
k
)
T
respectively, the FCVs is defined as
0
@
1
A
x
1
y
1
x
2
y
2
x
k
y
k
G
D
X
Y
D
(19.1)
where
G
includes all the change information between the two images for a given
image object, and the change magnitude
k
G
k
is computed with
q
.x
1
y
1
/
2
C
.x
2
y
2
/
2
CC
.x
k
y
k
/
2
k
G
k D
(19.2)
The higher the
k
G
k
is, the more likely that changes take place. Unsupervised
classifiers or threshold methods are commonly applied on the change magnitude to
identify changes.
Two aspects are essential for object-oriented CVA: one is the selection of
appropriate features to calculate FCVs, and the other is the determination of
a suitable threshold to identify changed objects. CVA has been applied on the
backscattering matrix of PolSAR data to detect the extent of change caused by
an inundation (Shen et al.
2007
). Given that the coherency matrix provides more
information than the backscattering matrix, this study applied CVA on the coherency
matrix to detect LULC changes. A widely accepted assumption is that the statistical
distribution of the pixels of change and no-change areas in the change magnitude can
be approximated as a mixture of Gaussian distributions (Bovolo and Bruzzone
2007
;
Camps-Valls et al.
2008
). Therefore, the expectation-maximization (EM) algorithm
was applied on the change magnitude to identify changed objects. EM is frequently
used for data clustering in machine learning and computer vision because it finds
clusters by determining a mixture of Gaussians that fit a given data set (Moon
1996
).
The Weka 3.6 software (Witten et al.
2011
) was used to implement EM algorithms
to determine the threshold to identify changed objects.
Search WWH ::
Custom Search