1C
m. The
ble and
er. The
tershed
uilding
ide the
orithm.
hadows
lapsed
essfully
rmance
building
e to its
rmanent
| spatial
tly used
of the
that the
and their
hniques.
iping of
ion and
he main
heir cast
"WS Was
shadows
s cast by
gs. Their
id corner
racing of
leses.
cases, to
the first
etect the
ymbining
ion, the
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
discrimination of the damaged areas from the non-damaged
could be done successfully. In the second step, aerial
photographs of the same area taken before and after the
earthquake were available. They matched the two images by
using the affine transform and also by hand. Then, the colors of
the corresponding pixels in each image were checked. Thus, the
areas having color differences were examined. As a
consequence, the two-case method were said to be fairly good
in determining the damaged areas. A similar study was
conducted by Mitomi er a/.,(2000) who detected the damaged
buildings by processing the aerial television images taken after
the 1999 Kocaeli, Turkey and Chi-Chi, Taiwan earthquakes.
The method was composed of defining the characteristics of
damage to wooden buildings based on hue, saturation,
brightness and the edge elements. Firstly, the damaged and non-
damaged pixels were classified. Then, texture analysis was
carried out and the damaged buildings were identified.
A near-real time earthquake damage assessment using the
integration of GIS and remote sensing was performed by
Gamba ef al., (1998). Their approach was comprised of two
phases. In the first phase, GIS side of the study was performed
via collecting and analyzing data about buildings and
infrastructures. In the second phase, the system receives near-
real time imagery of the suffered area to perform change
detection through shape analysis and perceptual grouping using
the pre and post-event aerial images.
Turker and San (2003) used pre- and post-event SPOT HRV
images to detect the Izmit earthquake induced changes. The
change areas were detected by subtracting the near-infrared
channel of the merged pre-event image from that of the post-
event image. The overall accuracy for the change areas were
found to be 83%.
In a recent study, Turker and Cetinkaya (in press), detected the
collapsed buildings caused by the 1999 Izmit, Turkey
earthquake using digital elevation models (DEMSs) created from
the aerial photographs taken before (1994) and after (1999) the
earthquake. The DEMs created from two epochs were
differenced and the difference DEM was analyzed on a
building-by-building basis for detecting the collapsed buildings.
The producer's accuracy for collapsed buildings was computed
as 84%. Further, Turker and San (in press) utilized the cast
shadows to detect the collapsed buildings due to Izmit, Turkey
earthquake. The available vector building boundaries were used
to match the shadow casting edges of the buildings with their
corresponding shadows and to perform analysis in a building
specific manner. Of the 80 collapsed buildings, 74 were
detected correctly, providing 92.5096 producer's accuracy.
In the present case the earthquake-damaged buildings are
detected from post-event aerial images using watershed
segmentation. This segmentation is based on the concepts of
watersheds and catchment basins, which are well known in
topography. In this approach, a gradient image can be regarded
as a topographic surface where the gray-levels of the gradient
image represent altitudes. (Figure 1). Therefore, the edges in the
image having high brightness values are considered as
watershed lines while the interior regions of the image having
low brightness values can be considered as catchment basins
(Sonka et al., 1998). The first step of the segmentation is finding
the minima (catchment basin) and piercing of it. Then, whole
relief is immersed into the water that causes the water flooding
into the areas close to the piercing points. As the relief goes
deeper in the water, some flooded areas tend to merge. In order
to prevent this, infinitely tall dams are placed along the
643
watershed lines. At the end, the resulting group of dams defines
the watersheds of the image (Shafarenko er al., 1997). Vincent
ef al,(1991) developed a fast and flexible algorithm for
computing watersheds in digital grayscale images. The
algorithm is based on an analogy of an immersion process. In
this algorithm, the flooding of the water in the image is
efficiently simulated using a queue of pixels. They applied the
algorithm in several fields with regard to picture segmentation
including MR imagery and digital elevation models.
Tm eel = oe Watersheds
_ -
—-—---=> Catchment
A [ 2-7 basins
|
—— 9
(b)
Figure 1. Watershed segmentation in one dimension: (a) gray-
level profile of the image data; (b) watershed segmentation —
local minima of gray level (altitude) yield catchment basins;
local minima define the watershed lines.
3. STUDY AREA AND DATA
3.1 Study Area
A selected urban area of the city of Golcuk was used as the study
area (Figure 2). A post-quake panchromatic aerial photograph
(Im resolution) of the region was obtained from General
Command of Mapping of Turkey. The study area consists of 284
rectangular shaped buildings. Of the 284 buildings, 79 were
collapsed and the remaining 205 buildings were un-collapsed.
The vector building boundaries were available from a previous
study conducted by San (2002) in our department. The data
contains for each building the Cartesian coordinates of the edge
points.
Figure 2. Study Area
