You are using an outdated browser that does not fully support the intranda viewer.
As a result, some pages may not be displayed correctly.

We recommend you use one of the following browsers:

Full text

The 3rd ISPRS Workshop on Dynamic and Multi-Dimensional GIS & the 10th Annual Conference of CPGIS on Geoinformatics
Chen, Jun

ISPRS, Vol.34, Part 2W2, “Dynamic and Multi-Dimensional GIS”, Bangkok, May 23-25, 2001
Yukio KOSUGI*, Munenori FUKUNISHI*,
Mitsuteru SAKAMATO**, Wei LU** and Takeshi DOIHARA**
*Tokyo Institute of Technology (kosuqi@pms.titech.ac.jp )
**Asia Air Survey Co., Ltd. R&D Department (ta.doihara@aiiko.co.jp )
We propose a method of detecting dislocation and slope failures by evaluating a pair of high-resolution aerial photographs,
taken before and after the earthquake. In our approach, we compared the two images in an adaptive nonlinear mapping
mechanism, to find out the discontinuous distribution of shifting vectors required for adjusting the two images.
1. Introduction:
For analyzing and prediction of dislocation earthquakes, as
well as for preventing landslip disasters, it is important to
detect automatically the geographical changes such as
dislocation and slope failures, from a pair of high-resolution
images taken before and after the earthquake [1-2]. Simple
subtraction of two images may give some information
regarding the land dislocations. In most cases, however, it
is difficult to get a pair of images of the same photographic
conditions, even after the geometric compensation using
affined transformation.
In our system, we nonlinearly map an aerial photo image to
another one observed at different time, of the same area
[1]. During the nonlinear mapping, we generated a set of
shifting vectors require for the nonlinear mapping and
evaluated the discontinuity in the vector distribution, to find
out the dislocation and landslips.
2. Nonlinear Mapping Algorithm:
As shown in Fig. 1, our nonlinear mapping algorithm
consists of the competition procedure to find out the best
matched shift of sub-areas and the consensus operation
among shifting vectors of near-by sub-areas for
maintaining continuous transformation. After the consensus
operation, image A will be deformed a little bit according to
the set of shifting vectors. After several times of iteration, a
nonlinear mapping from A to B will be formed, and the
spatial derivatives of shifting vectors will demarcate the
3. Choice of Consensus Operations:
In the algorithm mentioned above, the neighborhood zone
shape in the consensus operation significantly influences
the failure detection ability. In our former approach [1], we
defined a squared neighborhood, as indicated in dotted
line of Fig.2. However, when the zone of that shape
covers both sides of the sheer border S1-S2, the
consensus operation dilutes the difference of shifting
vectors at the boundary to be detected in the spatial
a) Strip zone approach: In our first trial, we defined a
strip in parallel with the border, as illustrated in double
solid line in Fig.2. In the strip zone approach, however,
we have to hypothesize the sheer border orientation, or
repeat the whole algorithm for several possible
orientations to find out the most clear result.
Imaae A
Competition among shifting candidates
Consensus among shifting vectors
Deformation of image A
According to the shifting vectors
Spatial derivation
Indication of discontinuous shifts
Fig.1 Sheer detection algorithm

* *


X *i *
■ ^ ^
X *
S 2
Fig.2 Neighborhood zone shapes
b) Bi-cluster approach: If we can classify all the shifting
vectors within an area into two groups according to
their vector orientations, we can make the consensus
operation to each of the groups. For doing this
approach, we plotted the shifting vectors on the Px-Py
2D space, where Px and Py are horizontal and