RKS
yaluation of
earthquake
\R data for
terferomet-
sifying two
er was used
ral network
neralization
rived from
ontrast than
survey map
the JERS-1
the number
\R) data, in
d cover and
destruction
buildings in
elation. An
Coherence
nezawa and
tion factors
and a target
ee and Liu,
velength of
in a case of
; terminated
rbit control
1sing coher-
lues in each
e surface. It
and Omatu,
Yosuke Ito
Table 1: Summary of JERS-1 and ERS-1 data acquisition
Acquisition | Precipitation | SLC
Satellite date (mm) name
JERS-1 9 Sep. 1992 0.0 J1
(L-band) | 6Feb. 1995 0.0 J2
1 Nov. 1992 1.5 El
ERS-1 12 Sep. 1993 0.0 E2
(C-band) | 27 Feb. 1996 0.0 E3
7 May 1996 0.0 E4
Table 2: Interferometric pairs of SAR data
SLC name Temporal Perpendicular Spatial Coherence
Satellite | Master | Slave | separation (days) | baseline (m) | decorrelation name
JERS-1 J1 J2 881 225 0.96 JC1-2
El E2 315 140 0.84 EC1-2
pe
E4 13 44 : 1
E4 : 2
In this paper, we propose an approach of neural network (NN) classification for extracting the damaged regions using
multi-source and temporal SAR coherence images. Among several known NN structures, we employed the learning
vector quantization (LVQ) as the NN for its merits of generalization ability, learning efficiency and good convergence
(Kohonen, 1997).
2 SAR DATA DESCRIPTION
The study area is densely inhabited districts in Kobe, Japan. The Hyogoken Nanbu earthquake, 7.2 magnitude, hit this
area on 17'^ January 1995. About 200,000 structures collapsed in the study area and 5.2% of the collapsed area was
burned. In a reclaimed land, subsidence caused by liquefaction resulted in destabilised structures and deformed roads.
Table 1 is a summary of JERS-1 and ERS-1 SAR data acquisition. Single look complex images (SLC) were produced
by compressing SAR raw data. To generate the precision SLCs of JERS-1, we applied low pass filtering to eliminate
microwave interference from ground radar systems and compensated the sensitivity time control and the automatic gain
control. The path number of both SAR scenes is 72. The row numbers are 242 for JERS-1 and 243 for ERS-1. All
scenes were acquired in a descending mode. Total six scenes of the multi-temporal JERS-1 and ERS-1 were used in the
experiment of methodology. Figure 1 shows the study area in a three look amplitude image of J1 where mountain and
sea areas are masked off. There is a densely built-up area in the central part of figure 1. Several dark lines indicate major
roads. Port-Island is one of the reclaimed lands. We choose the SAR data such that the earthquake event is included
within temporal separation of an interferometric pair which has an adequate baseline to produce a coherence image. The
scenes acquired before the earthquake are J1, E1 and E2 and after the earthquake are J2, E3 and E4. Daily precipitation
on the dates of data acquisition shown in table 1 were observed by the Kobe marine meteorological observatory.
Table 2 shows temporal separation (D), perpendicular baseline (B ) and theoretical spatial decorrelation (p;) for all the
interferometric pairs. Here, the spatial decorrelation p, is calculated as
C
Nr Br OH ay (1)
p.17
Where 05 is the nominal incidence angle of the radar on the ellipsoidal earth, a the local terrain slope, A the radar
wavelength, r the slant range, B,, the frequency bandwidth of the transmitted chirp signal, and c the velocity of light (Lee
and Liu, 1999). As the study area is almost flat, we assume o — 0?. JC1-2 is the coherence image derived from J1 and J2.
Similarly, EC1-2, ECI-3, etc., denote the coherence images derived from the relevant ERS-1 SLC pairs. The p, of JC1-2
IS higher than that of EC1-4 despite JC1-2 has longer B, . The coherence images that contain information of the damage
caused by th earthquake are JC1-2, EC1-3, EC1-4, EC2-3 and EC2-4. EC1-2 is the coherence image derived from before
event pair.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B1. Amsterdam 2000. 157
