Yosuke Ito 
  
  
0 
Osaka Day ı — iim 
Figure 1: Three look amplitude image of J1 in Kobe, Japan 
3 COHERENCE IN DAMAGED REGIONS 
Coherence at each pixel is estimated by 
F,F; 
em D : Q) 
(1211?) / (Es?) 
where E, and E» are SLC values, * and (-) denote complex conjugation and ensemble averaging, respectively. After 
applying the earth flattening to an initial interferogram, ^/ was computed between master and slave SLCs based on 16 
pixels taken in the azimuth direction and 4 pixels in the range direction in consideration of the standard deviation (SD) 
required for the classification. All coherence pixels were resampled to approximately 30m squared area after being 
converted from slant range to ground range geometrically. The coherence can be modelled as 
Y= Pn Pt Ps 3) 
where p, is the thermal noise of the SAR system and p; the temporal decorrelation (Zebker and Villasenor, 1992). We 
can assume p,, = 1 since the signal-to-noise ratios for both SAR systems are high enough for this study. If the coherence 
image is derived from a SLC pair across the earthquake event, p; is affected by the changes caused by the earthquake and 
other temporal effects and therefore becomes lower in the damaged regions than in the undamaged regions. 
The SLCs of two interferometric pairs: J1 and J2; E2 and E3, were acquired under almost the same weather and time 
conditions except B , and SAR system parameters. Figures 2 (a) and (b) show the coherence images of JC1-2 and EC2-3 
for comparison of radar wavelength effects. Histograms of JC1-2 and EC2-3 are presented in figure 3. The coherence of 
JC1-2 is considerably higher and has significantly higher contrast than that of EC2-3 as shown in figures 2 and 3 even 
though they have nearly equal p; and D. Robustness for L-band SAR coherency is thus convincingly demonstrated. 
On the other hand, as EC1-2 (figure5 (a)) is derived from two SLCs before the earthquake, it has no relation to the earth- 
quake and includes land use information only. As shown in figure 5 (b), the histogram of EC1-2 indicates higher coherence 
than those of JC1-2 and EC2-3. The extraction approach using the decorrelation by surface change can essentially apply 
to regions with the high coherence only. This will eliminate the areas subject to temporal decorrelation not relevant to the 
earthquake. Hence, the study area is reduced such that ^; > 0.6 in ECI-2. 
Figure 4 shows a hazard map surveyed by Ministry of Construction and Architectural Institute of Japan in 1995. The map 
was rasterised to 30m x 30m to correspond to the coherence images in pixel by pixel. Black dots in the map indicate 
burned or completely collapsed structures. These black dots are expressed as the damaged category (w;) and the others 
are defined as category (w») in the map. The coherence of the damaged regions tends to fall down in figures 2 (a) and 
(b). The surface changes caused by the earthquake evidently bring about this phenomenon. To assess the distribution 
for w, and v» in all the coherence images, mean values and SDs of the coherence value for each category are calculated 
as illustrated in figure 6. Here, | denotes the mean value and the both side ranges are +SD. AII coherence images have 
  
158 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part Bl. Amsterdam 2000. 
Pixel count 
à sil 
ism 
This 
over 
are t 
dete 
4 ] 
The 
of re 
Well 
denc 
the 1 
the 1 
dist: 
to Ze 
cate 
Ac 
The 
Koh 
men