PE&RS July 2016 Public - page 533

reliable source for this purpose (e.g., DT4). A
by stereo imagery can be also used (e.g., dataset DT1 to DT3)
provided that a proper matching technique is used which
results in high precision elevation model generation.
Regarding the
specifications, since the
is projected
to the base
image to transfer the patch
to the corresponding
place in the target
image, it is necessary that the
and one of
the images, which is considered as the base
image, are acquired
at the same time, or if they are not synchronized, there must
be no change between them. In this study, in datasets DT1 to
DT3 the
is generated from a stereo image set from which
one image is selected for change detection. Therefore, there
is no change between the base image and the
. However,
for dataset DT4, the
is generated using lidar data which is
separate from the bi-temporal imaging sources. In this case, the
is checked to ensure no structural differences exist.
On the other hand, in cases that
is generated from
stereo imagery (DT1 to DT3), since the
was generated
based on the same
, projecting it back to the associated
image space did not require the compensation for bias errors.
However, if the source of
is different (dataset DT4), bias
compensation is required for both bi-temporal images.
Change Detection Discussion
We tested two change criteria, Image Differencing and
Transform, on four different urban datasets. The shape of
curves of accuracy assessment step (Figure 12) and the
s (Figure 13), which are over 90 percent across
all four datasets, confirm that the change criteria used are
suitable for binary change detection. It also confirms that the
corresponding patches generated by the
method can be
used for change detection even with a simple change detec-
tion method such as Image Differencing.
As the main purpose of this study was to present a coregis-
tration method useful for off-nadir multi-view angle satellite
images, we did not test more complicated change criteria
available in the literature such as the ones presented by Ye
and Chen (2015) and Gueguen
et al.
, (2011) that can be used
for future studies.
By solving the coregistration problem, the
provides the opportunity to utilize a wide range of imagery
for change detection applications even when those images
were not useful for change detection using the conventional
methods. However, in selecting the bi-temporal image sets,
we still need to consider the illumination and seasonal ef-
fects. Also, if across-sensor images are used, the spectral band
widths of the images have to be similar.
The bi-temporal image sets used in this study have approx-
imately the same spectral properties. Also, by having similar
solar angles and close to anniversary acquisition times, the
illumination conditions of the bi-temporal sets are similar.
Based on Lambert’s cosine law (Riaño
et al.
, 2003), regard-
less of atmospheric effects, as soon as the solar angles of the
bi-temporal images are not highly different, the illumination
condition in the images should be similar. Therefore, a linear
radiometric normalization method was sufficient to attenuate
the radiometric differences of the images used in this study
since they were already corrected for atmospheric effects.
The suggested method for coregistration in this study is
based on patches (object-based); however, it is not limited to
object-based methods. As the method transfers the patches from
one image to the other one using their pixel components, one
can use change detection using pixel-based work as well, con-
sidering that it is not going to be a one-to-one transformation
and a post process is required in order to compensate for mul-
tiple projections. Also, the change detection accuracy could be
higher if the patches were meaningful object borders by using a
pre-generated urban map such as a building border
Finally, since this framework promotes the use of off-nadir
images and due to occlusions the study area will be limited
to the visible objects, multiple imagery can also be used for a
more complete view of the urban objects. Even though we only
investigated the use of bi-temporal imagery, the
is capable of being extended to use multiple images as well.
Almost all
satellite images are taken from a different
viewing angle, because of the agile ability of the satellite
sensors to quickly take images of interest within an off-nadir
angle of up to 45°. The angle difference creates significant
challenges in coregistration of images with elevation varia-
tions, especially in urban areas, so that it is very difficult to
achieve accurate change detection results. To avoid the chal-
lenges caused by viewing angle difference, most research pa-
pers either used nadir images or selected flat areas for change
detection. The vast majority of existing
satellite images
cannot be used for change detection.
To overcome the challenges of using images from different
viewing angles for change detection, this paper introduced
an effective patch-wise coregistration (
) method for
accurately coregistering individual patches (segments) in
bi-temporal images at a local level, so that accurate change
detection results can be achieved regardless of the viewing
angle difference. The
method utilizes the
of the bi-
temporal images and the elevation information of one corre-
to guide the co-registration of individual image
patches in the two images. Therefore, the relief displacements
of elevated objects in the images can be taken into account in
. Through the integration of a multispectral change
detection analysis method, such as the Image Differencing or
Transform methods, the change between the correspond-
ing patches can be identified by measuring the difference of
their spectral properties.
Experiments with four bi-temporal image sets acquired by
Ikonos, GeoEye-1, and Worldview-2 satellites demonstrated
that the coregistration accuracy of the
method reached
at the range from 80 percent to 95 percent in terms of over-
lapped areas (Area Ratios) of corresponding patches; whereas
the accuracy of the traditional coregistration method reached
the wide range from 5 percent to 90 percent depending on the
height of the object. The change detection accuracy (
) of
incorporated change detection framework reached
over 90 percent for all of the four datasets.
The success of the
method for coregistration of off-
satellite images and the high accuracy of change de-
tection in urban environments open up the potential of using
the widely available
satellite images for detailed change
detections, regardless of the angle difference of the images.
This will significantly increase the efficiency and lower the
cost of change detection.
This research has been funded by the National Sciences
and Engineering Research Council (
) of Canada. We
acknowledge the City of Fredericton and the Department of
Public Safety of the Province of New Brunswick for provid-
ing the lidar data of Fredericton. We also acknowledge Space
Imaging LLC and DigitalGlobe for providing the Ikonos and
GeoEye satellite images.
AL-Khudhairy, D., I. Caravaggi, and S. Glada, 2005. Structural
damage assessments from IKONOS data using change detection,
object-oriented segmentation, and classification techniques,
Photogrammetric Engineering and Remote Sensing
, 71(7):825–837.
July 2016
447...,523,524,525,526,527,528,529,530,531,532 534,535,536,537,538,539,540,541,542,543,...582
Powered by FlippingBook