PE&RS July 2016 Public - page 526

as transferring the patches from the base image to the target
image (as shown in Figure 1).
Change Analysis
Since the major reason for proposing the PWCR method in
this study, is to be able to use the multi-view-angle images for
change detection, it has to be proven that after the coregistra-
tion, changed patches can be detected by comparing the spec-
tral properties of the corresponding patches in the bi-temporal
images. Therefore, we examined the so-produced coregistered
patches in change detection. We tested two change detection
criteria: Image Differencing and MAD Transform. Image dif-
ferencing is a simple method and MAD transform is a more
advanced method for change detection. In this section, more de-
tails are provided regarding the used change detection criteria.
Image Differencing
The subtraction between coregistered multispectral images
provides a simple way to detect changes. With this crite-
rion, small intensity differences indicate no change, while
large positive or negative values indicate changes. Decision
thresholds specify the significance of change which are usu-
ally expressed in terms of standard deviation from the mean
difference value (Canty, 2014).
In this study, we use the mean value of the correspond-
ing patches in the Image Differencing algorithm to detect the
changed patches.
MAD Transform
MAD Transform, which can be considered as an extended
version of PCA (principle component analysis), is used as
another change identifier in this study. In PCA change detec-
tion, the primary component contains consistent information
between the bi-temporal datasets, whereas the secondary
component captures the difference (Chen and Hutchinson,
2007). However, as Nielsen et al. , (1998) state, PCA is depen-
dent on scale, number of channels, and forces the same gain
setting of intensity measurements (in the sensors), while MAD
is designed to compensate for the shortages of PCA. A MAD
Transform is linear scale invariant; which means it is insensi-
tive to linear radiometric and atmospheric effects (Nielsen et
al., 1998). Because of the linear model of the MAD Transform,
it can be inferred that this method is also capable of compen-
sating for the slight individual bandwidth dissimilarities (if
they are linear) in different sensors.
MAD Transform is based on measuring the difference of a
linear transformation of two multispectral vectors from bi-
temporal images, X and Y, in such a way that the variance of
the intensities is maximized (Canty, 2014):
D a X b YT T = − (7)
in which, a and b are coefficients calculated based on Canoni-
cal Correlation Analysis. As a result, the multispectral bands
are transformed into a new space (D) in which the changes are
highlighted.
Study Area
The proposed framework is implemented on four bi-temporal
satellite image datasets, from the city of Hobart, Tasmania,
Australia, and Fredericton, New Brunswick, Canada, each
covering at least a one square kilometer area. The coverage
areas contain typical urban structures with combinations of
small to large and low to moderately high elevated buildings.
Photographs of the study areas are presented in Figure 6.
Table 1 reports the specifications of the satellite images
used in this study. The bi-temporal combinations of the im-
ages and DSMs for change detection are presented in Table 2.
The datasets are selected from images with different view-
ing angles (both azimuth and off-nadir angles are different),
Figure 6. Photographs of the study areas: (a) a combination of houses and elevated condos, Fredericton, Canada (dataset ID: DT4); (b) a
combination of houses and industrial buildings in Hobart, Australia (dataset IDs: DT1 and DT2); and (c) a typical developed urban envi-
ronment in Hobart, Australia (dataset ID: DT3)
Table 1. Metadata of S atellite Images Used in this Study
Data IK62 IK74 IK75 Geo000 Geo001 WV2 WV2
Satellite name Ikonos Ikonos Ikonos GeoEye1 GeoEye1 Wordview2 Wordview2
City name Hobart Hobart Hobart Hobart Hobart Fredericton Fredericton
Country Australia Australia Australia Australia Australia Canada Canada
Date 2003-02-22 2003-02-22 2003-02-22 2009-02-05 2009-02-05 2011-07-20 2013-08-18
Sat. Az (deg) 293.74 329.42 235.74 193.69 53.48 169.9 226.6
Sat. Elev. (deg) 75.17 69.14 69.20 70.06 63.87 72.9 59.1
Solar Az. (deg) 47.1 47.21 46.9 59.58 59.58 154.8 167.9
Solar Zenith (deg) 41.2 41.23 41.1 40.62 40.62 23 33.4
Approx. GSD (m) 0.9 0.9 0.9 0.5 0.5 0.5 0.58
Table 2. B i-temporal Combinations of Satellite Images Used in this Study
Dataset
ID
Target
Image
Base
Image
Source
of DSM
DSM
Accuracy
DT1 IK62 Geo000 Geo000-Geo001
stereo imagery 0.5m
DT2 IK75 Geo000 Geo000-Geo001
stereo imagery 0.5m
DT3 IK74 Geo001 Geo000-Geo001
stereo imagery 0.5m
DT4 WV2-2013 WV2-2011 LiDAR data 0.5m
526 July 2016 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING