3. Istanbul 2004
9 model
8
on Coincident
[ebbian rule for
his model has a
echanism called
shown in Fig.5.
n process.
rich consists of
rer). We define
as Xj, yj using
| value of these
ber as k. Then
1ed as Eq.6.
(6)
(7)
to yij,
area in
htness or other
feature vectors.
arching for the
n by Eq6 is
(8)
he position
mum value
executed for
this process is
rs generated in
int of mutual
odels to realize
hat emphasizes
s is realized by
1 the consensus
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
area R that has arbitrary shape and size around the target
position.
a” e median(d^ R) (9)
where median() = a function which returns median value
(3) Feedback process
The convergent characteristics of mapping varies due to not
only the evaluation value but also the size of sub-area, shape
and size of consensus area and so on. After several competition
and consensus process, these parameters are updated and these
processes are repeated. Feedback process is terminated when
all shifts in mapping are below a predefined value or the
iteration number of process reaches the preset one.
Output Layer
T +
Y 7 Search
2X /
5 ZZ Area
LT
Unit Element
i | * Sub-a rea a (Image Pixel)
S Output Layer OOO 000000
Input Layer Correlation Input Layer OOO OO000
: hl RES 4- Operation Processing
| Unit | | Hon "998000909
C,
LUS
OL,
Mapping Layer
Consensus Area
Figure 5. Conceptual diagram for adaptive nonlinear mapping
Adaptive
3 nonlinear
à mapping
34 Evaluation of Change Detection Ability by ROC
For evaluating of the automatic change detection ability, we use
ROC (Receiver Operating Characteristic) chart, which has been
widely used in evaluating the performance of radar and sonar
systems (Vantree, 1968). For our application, we redefined the
equations as follows.
PD = Nr/ Nc
PF = Nf /(Nt — Ne)
Nf = Na— Nr
(10)
where PD = Probability of detection
PF = Probability of false alarm
Nr = Number of buildings which are correctly detected
of the changes
Nc = Number of buildings which are actually changed
Nf = Number of buildings which are incorrectly
detected of the changes
Nt = Number of targeted buildings
Na = Number of buildings which are automatically
detected as the changed buildings
As in general, when we adjust the threshold level that specifies
the changed and non-changed area separately so as to get higher
PD value, at the same time the ratio of false change detection
increase, which means higher PF. Therefore when we plot the
ROC chart with PF value as x-axis and PD as y-axis, an ideal
curve should rise rapidly in upper left direction.
The number of buildings is used when vector maps of buildings
are available. However when digital maps are not available in
2D image matching method, pixel number is used for the
evaluation.
4. CHANGE DETECTION BY 3D IMAGE MATCHING
METHOD
4.1 Process Flow
The flow of 3D image matching is shown in Fig.7. In this
approach, we assume that stereo models with exterior
orientation before and after the earthquake and latest digital
map data managed by national or local government are
available.
In the first step, matching between buildings in digital map and
stereo models is executed. If the digital map is 2D, each
altitude of building will be calculated by supposing that all
nodes of the polygon have the same height, making all
polygons of buildings 2.5D. This process is conducted by
matching raster image and vector data with consideration of
geometric constraint based on exterior orientation parameter.
Namely under the condition that x-y coordinates of a polygon
building is fixed, z coordinate (altitude) is gradually shifted and
projected photo coordinate as raster image for evaluating
matching criterion in each supposed z value. Finally the most
suitable height is selected where the best matching criterion is
given. The cross correlation coefficient is used as evaluation
function in this study. In the same way, z-direction polygon
matching is executed in both stereo models before and after the
earthquake.
If there are changed areas in the image after earthquake,
inconsistency of altitude between stereo models will occur.
Therefore change detection by z-difference is performed in the
next step.
The decision of whether a building has changed or not is carried
out by the following judgement.
A = Z er 2 dns : changed (1 ] )
others: not changed
where Zbefore = Estimated altitude of building in before
earthqugke
Zier = Estimated altitude of building in after
earthquqke
Zihres = Threshold of altitude difference
[n the next step, decision of change detection by texture
analysis is carried out so as to solve the case that change has
occurred with minor height change. The following image
texture indexes are applied in this study. Indexes show
"constant (C)", *angular second moment (A)", "entropy (E)"
and “mean (M)” respectively (Weszka et al., 1976).
