Object-Based Change Detection with Sliver Removal
To minimize the overall discrepancy from the lack of ac-
curate geo-localization, image matching was carried out by
splitting the original images into four rectangular tiles with a
small overlapping area. This operation produces eight images
that can be matched by considering a threshold for outliers
of 23 m, that is, the expected geo-localization accuracy of
QuickBird data (CE90%). A final
RMS
of approximately 8 m
was found for the different tiles because the affine model is
not suitable for the whole tile. Obviously, the
RMS
values of
least squares estimation are not optimal in terms of metric
accuracy. However, the aim of this matching step is only the
extraction of corresponding points for the estimation of the
piecewise affine functions and classification resampling.
Shown in Figure 6c is the result after classification registra-
tion based on the piecewise affine functions estimated on the
triangulation from the detected set of corresponding points.
The initial poor geo-localization accuracy in classification
c
1
= [
V
1
È
S
1
È
O
1
] (Figure 6b) is significantly enhanced, obtain-
ing a new consistent set of polygonal files
c
1*
= [
V
1*
È
S
1*
È
O
1*
]
(Figure 6c). The spatial modification allows objects in
c
1*
to
be compared with
c
2
, notwithstanding results that are still af-
fected by spurious slivers.
As the goal was the detection of partially collapsed build-
ings, object-based change detection was carried out with an
algorithm developed for the city of L’Aquila. The basic assump-
tion is the use of shadow information, which becomes visible in
the post-earthquake image. The intersection between vegetation
and shadows in the pre-earthquake image and shadows after
the earthquake is estimated as
S
2
Ç
(
V
1*
È
S
1*
). These areas are re-
moved from data processing. Then, damaged buildings (
DB
) are
detected with the relationship
DB
= (
S
2
\
S
2
Ç
(
V
1*
È
S
1*
)
O
1*
. The
results still affected by slivers from the lack of edge correspon-
dence of the multi-temporal polygons are shown in Figure 6d.
Finally, the estimation of the density for the different poly-
gons forming layer
DB
can be used to remove small slivers.
A threshold of 1.6 for geometric density was used to separate
slivers from real-change polygons. The use of a triangulated,
irregular network allows one to obtain a very good correspon-
dence for object boundaries, for which the remaining spuri-
ous effect gives objects like filaments. Although the consid-
ered case study is very complicated (the medieval city center
of L’Aquila has very irregular buildings, monuments, squares,
etc.), the method was able to isolate real changes (damaged
buildings) from errors generated by mismatching boundaries.
Conclusions
This paper describes a procedure for sliver removal in object-
based change detection from
VHR
images. The method is a
two-step solution based on (a) piecewise affine functions for
a preliminary spatial registration of classification results, and
(b) a sliver removal for the resulting elongated features. The
method is robust for effects such as weak geo-localization ac-
curacy (e.g., the case of raw images), scale variations, shad-
ows, and fragmented segmentation.
The proposed case study is the city of L’Aquila. Damaged
buildings were detected using object-based image analysis
after the earthquake in 2009. However, the method was devel-
oped considering a wider use in object-based change detec-
tion where slivers are inevitable. The proposed solution can
overcome the limitation of Boolean algebra with vector clas-
sification layers without perfect edge-to-edge correspondence.
The lack of perfect boundary correspondence generates spuri-
ous slivers during basic and simple geometric operations on
vector layers.
The developed solution offers a large flexibility in the
preliminary phases of the work (e.g., image registration, pan-
sharpening, and object-based classification). These tasks can
be carried out with different commercial software, where the
users can modify the parameters of image analysis following
Figure 5. The triangulation for an image of the case study acquired on 08 April 2009.
166
February 2016
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
