International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
and transverse directions of the building. Figure 1 demonstrates
a small area with extracted shadows of buildings and the
aligned directions shown in FFT power spectrum.
o com b)
Figure 1. Example of dominant directions in focused area: a)
extracted shadow scene and b) FFT power spectrum
When earthquake occurs, the strong ground motion will cause
the damage of buildings. As a result, buildings might be fully
collapsed, partly collapsed or its orientation might be changed.
In those cases, the lengths of building's shadows along the
dominant directions in focused area will be changed. Thus,
length comparison between pre-event and post-event scenes can
be one of the indicators for damage detection.
As shown in Figure la, the complexity of urban area also
generates the complexity in extracted shadows. There are
considerably large-size shadows generated by buildings or tree
rows and small-size shadows generated by stand-alone trees,
cars, or tents. Lack of spectral information from shadows, shape
and size of shadows must be employed in discrimination
analysis. Scale-space analysis, which takes scale property of
objects into account, has been developed for decades
(Lindeberg, 1993). Shadow analysis in this study concerns the
changes of shadow lengths. Thus, the scale-space analysis
should well preserve the details of each considerably large
shadows but remove all small shadows. Non-linear scale space
analysis based on area morphology has been proved to satisfy
this requirement (Acton and Mukherjee, 2000). Briefly, area
morphology scale space can be illustrated as follow.
Let set S defined on domain 42 c Z^. Area open $ o s remove all
components of area less than s in the set S. Area close $ e s
remove all components of area less than s in the set S"
(complement of S). The scale space is constructed using AOC
(area open-close) or ACO (area close-open) operators. Let /, be
the image representation at scale s, AOC (area open-close)
scale space //} given by
1,(1)=1,(t-1)os(t)es(t) (1)
Given a scale space {I}, 1,(x,y) is intensity at position (x,y) and
scale s, to discriminate the shadows based on their size, we
cluster /,(x,y) across the scale space. In this study, we simply
apply K-mean clustering algorithm. Scale-space analysis
employed in this study is like a non-linear filtering. Figure 2
demonstrates a result of area morphology scale space filtering.
It shows a perfect preservation of object's boundaries.
608
Figure 2. Example of area-morphology scale space filtering: a)
Before filtering and b) After filtering
3. METHODOLOGY
Based on the above observations, we proposed an automatic
shadow analysis, which compared the extracted shadows from
pre-event and post event high resolution imageries to find out
the cue for damage detection. This section presents only the
core processing, which is fully automatic. There must be some
pre-processing steps, which depend on the in-hand high
resolution satellite imagery, before this automatic processing
can be employed. Thus, the proposed method can be employed
not only one kind of high-resolution satellite imagery.
Step 1: Extract both pre-event and post-event imageries in the
small equal portions. Separately processing each small portion
is not only to focus the analysis into specific dominant
directions but also to speed up the processing. While area
morphology scale space produces a perfect result (Figure 2), its
shortcoming is time consuming (Acton, 2000). Processing in a
small portion will be faster than a very big scene. Moreover,
parallel processing can be employed.
Step 2: Apply area morphology scale space filtering as
introduced in Section 2 for each portion.
Step 3: Concerning the boundaries of extracted shadows,
dominant directions of each portion are computed. Examining
the local window 3x3 of each pixel, the direction of each pixel
is assigned follow the rules in Figure 3, where direction is
represented in degree. Consequently, histogram analysis shows
the dominant directions for each portion.
X | | X x
TR KT X S ee eee Fh
X x | | ET N
0 90 45 135
X X | x
X x X ix ix
oerte ru ES x
EX X
XN T x. ] n. Xx
22.5 67.5 112.5 157.5
Figure 3. Distribution of boundary pixels in 3x3 local window
and assigned angle
Step 4: Compute the lengths of shadow along the above
specified dominant directions. Length comparison is carried out
afterwards. Let Dif/L, and Dif/L; be the differences in lengths
Inter
alon
foll
whe
Figui
Step:
amor
thres!
Step
assigi
The |
pan-s
Alger
areas
The f
2002)
23, 21
terrail
mapp
comp:
in the
Furthe
due tc
propo.
registe
Shado
classif
2900
sharpe
Detect
checki
really
(Figur
discrir
buildir
be wel