il 2004
pure
d 2003,
ped in
ntation
gitizing
arly an
in this
can be
uted by
F linked
lifferent
precise
/ in the
1g from
sise co-
here the
by the
re class,
authors
2002)
ation by
u...
yuildings
; Right:
indings.
ssible to
nly roof
ake into
ish if a
ding of
' can be
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
performed, showing a significant increasing in reflectance on
the damaged structures (Figure 2, right).
3.2 Automatic change detection
Since polygons are the principal geometric features in urban
areas, it is possible to consider building as objects, identified in
a map or in a satellite image: a change detection can be carried
out by updating a database of polygons, in this case from the
pre-event situation to the post-event situation. However, after
an earthquake event, particularly in a developing country, a
cartographic base is often unavailable or not accessible: remote
sensing imagery represents therefore one of the most important
information sources.
Using only the imagery, there are different approaches,
principally pixel-based and object-based. In pixel-based
approaches, images are processed as they are, and the
registration is the very first result to acquire. In object-based
change detection algorithms, imagery are firstly divided in
meaningful regions, to simulate the abstraction done by a
human interpreter.
It is necessary however to preliminary define what is a change,
how the change is perceived and then to identify which
algorithm is suitable to explain the perception of change. For
the human eye, is very simple to recognize similar objects in
two images which are changed of not changed, and also to
define what is a building and what isn’t, and how is changed.
Conversely, in an automatic classification process, there is an
operation sequence:
l. image registration
2. define the object of interest
3. findachange detection index
4. classify the image.
Image registration is the first result to acquire, however there
are lots of trouble in the geometry of VHR imagery. In our
case, satellite information represents the only “cartographic”
information: no map, DEM or orthorectified products are
available. Features represented are very different in geometry
and illumination, causing problems in automatic registration
(lack of coherence in high resolution images is largely
evidenced in literature). Moreover, being the view angles
different, else if ground objects like streets are perfectly
registered, the same doesn't happen to buildings, because of
relief displacement.
For human analysis, is very simple to classify, comparing
objects and surroundings: as seen in paragraph 3, a difference in
reflectance in a single building roof is an index but this is not
definitive.
In the following paragraphs, some tests are shown, carried out
using pixel-based and object-based approaches.
4. PIXEL BASED CLASSIFICATION
Change detection module implemented in Erdas Imagine 8.6
package was applied. It computes the difference between pre-
and post-event images and highlights the changes in brightness
as a percent that exceed a user-specified threshold. This method
permits to identify large damaged areas but lots of “false
alarms” still remain. The results can be refined using a decision
tree and considering other variables (such as ISODATA
unsupervised classification and NDVI).
The pixel-based classification was conducted using Erdas 8.6
Knowledge Engineer module. This is a classifier that uses
hypotheses, rules and variables to create classes of interest and
695
to make a rule-based classification through an implementation
in a decision tree. Our aim is to identify a class representative
of damaged areas.
Marmara Earthquake: Landsat images and IRS August image
were registered to the IRS September post-event image. Two
methods were used: the Rubber Sheeting geometric model for
IRS images registration and a polynomial transformation for
Landsat. In the Rubber Sheeting model, a triangulated irregular
network (TIN) is formed over all the control points, then the
image area covered by each triangle in the network is rectified
by first (linear) order polynomial. This model is useful when
the geometric distortion is severe and there are a lot of CPs
(Control Points). In the IRS imagery, lots of CPs (50) were
identified to correct geometry (the imagery aren't
orthorectified) and to register the images.
The change detection module was applied between pre- and
post-event IRS images. In order to consider only the urban area
of Golcuk, two variables were introduced in the decision tree:
the class representative of buildings derived from ISODATA
classification of pre-event IRS image, and limit values in the
band 2 and 4 of post-event Landsat imagery, in order to
separate urban areas from vegetated areas.
Boumerdes Earthquake: IRS images and QuickBird June
image were registered to the QuickBird April image. There
were problems in registering QuickBird imagery to each other
because of buildings geometry and presence of shadows caused
by differences in sun azimuth angle and off-nadir view angle.
A first registration was made using 116 CPs with a polynomial
transformation and a rubber-sheeting model. In both cases the
results weren’t satisfactory because there was a shift in
buildings position between the registered and master image.
It was decided to register the image using 2494 CPs,
corresponding to the barycenters of buildings (in pre- and post-
event QuickBird images) whose coordinates derive from the
visual classification of Boumerdes urban area (described in
paragraph 2), and the Rubber Sheeting method. Therefore, the
registered image obtained and the master image present
minimum relief displacement effects between buildings.
The change detection module was applied between pre- and
post-event QuickBird images. In order to remove false changes
caused by presence of shadows, ISODATA classification of
pre-event image was used: a class for shadows was identified
and removed by its implementation in the decision tree. Other
two variables were used in order to select the urban area:
ISODATA classification of pre-event image and NDVI of pre-
and post-event images.
Figure 3. Boumerdes, change detection results
by pixel-based classification.
