International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
computed using a simple arithmetic operator, such as the 
difference, correlation or block correlation. The location of 
building damage is displayed using a damage assessment map. 
Damage severity is established through building damage 
profiles, which demonstrate the general association between 
temporal changes in the remote sensing coverage and the extent 
of building collapse (as determined by field survey). 
  
  
(1) Remote sensing data 
Digital imagery of study site, acquired before and after the 
  
  
  
  
(2) Pre-processing 
Geo-reference and co-register images. For high resolution 
optical, perform edge enhancement and texture analysis 
Y 
(3) Change detection 
Identify changes between before and after images using 
arithmetic operators, such as difference and Correlation. 
Y 
(4a) Damage detection algorithms (4b) Ground truth data 
Locate urban building damage using - e&-] Validate damage maps and 
damage maps. Establish damage severity calibrate profiles using observed 
using damage profiles damage states 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
Figure 1. Building damage detection methodology 
3. IMPLEMENTATION 
3.1 1999 Marmara (Turkey) earthquake 
3.1.1 Method: Change detection algorithms were developed 
for the Turkish city of Golcuk, the most densely populated 
settlement in Kocaeli province. As illustrated in Figure 2, the 
city experienced extensive damage during the 17th August 1999 
magnitude 7.4 earthquake. According to Coburn ef al. (1999, 
cited in Rathje, 2000), 30-40% of structures collapsed, due 
mostly to pancaking and the soft first story effect. 
  
    
  
  
Figure 2 Building damage in Golcuk (courtesy of R. Andrews) 
As shown in Table 1, remotely sensed SPOT 4 panchromatic 
and ERS intensity and complex images were acquired ‘before’ 
and ‘after’ the earthquake. In terms of pre-processing, the 
respective datasets were geo-referenced and co-registered. The 
SPOT imagery was provided in a geo-referenced format, and 
the scenes co-registered within the ENVI processing 
environment. For the SAR coverage, ground control points 
were extracted using the ESA SAR Toolbox, with geo- 
referencing performed in ENVI. The co-registration was fine 
tuned by applying a template matching algorithm. 
  
  
  
  
  
  
  
Earthquake Sensor Spatial Date 
Resolution (m) 
Marmara SPOT 4 (pan) 10 Before: 7/15/99 
After: 8/20/99 
ERS 2 20 Before: 4/24/99 
ERSI After: 9/10/99 
Boumerdes Quickbird 0.6 and 2.4 Before: 4/22/02 
(pan and MS) After: 5/23/03 
  
  
Table 1. Remote sensing imagery specifications 
The performance of several change detection measurements 
was investigated. For the SPOT panchromatic data: (1) simple 
difference (dif); (2) sliding window-based (cor) correlation; and 
(3) modulated block correlation (bk cor) were computed 
between the ‘before’ and ‘after’ images. In the case of SAR 
intensity and complex data, indices comprised: (1) simple 
difference between intensity values; (2) sliding window-based 
correlation; (3) modulated block correlation; and (4) coherence 
(coh) or complex correlation between complex images. 
The next methodological step compares indices of change with 
‘ground truth’ damage data collected shortly after the 
earthquake by the Architectural Institute of Japan 
reconnaissance team (AlJ, 1999). The zone-based sampling 
strategy in Figure 3a was employed, using 70 administrative 
boundaries corresponding with the street network. Within each 
zone, damage states were recorded for a sample of buildings 
using the European Macro-seismic Scale (EMS98): 
e Grade 1: Negligible to slight damage 
° Grade 2: Moderate damage 
° Grade 3: Substantial to heavy damage 
* Grade 4: Very heavy damage 
° Grade 5: Destruction/collapse 
The corresponding damage state map in Figure 3b expresses the 
percentage of collapsed structures (Grade 5) as a function of the 
total number of observations (sum of Grade 1 through Grade 5). 
For analytical purposes, these percentages are divided into the 
follow categories: A (0-6.25% of buildings totally collapsed); B 
(6.25-12.5%); C (12.5-25%); D (25-50%); and E (50-100%). 
The additional ‘Sunk’ zone corresponds with an area in north- 
cast Golcuk experiencing significant subsidence. 
Finally, damage maps and building damage profiles were 
generated to determine the location and severity of urban 
damage. For the profiles, average change was computed for 
cach of the 70 zones, and these responses aggregated by 
damage state. The result is a central measure of tendency and 
standard deviation (plotted as error bars) for class A through 
class E. 
  
Esthnated fault fine on surface 
  
^ - x s — 
«o M, : Ye 
Ve Lancy Iynit Bay * Branch, 500 m 
ive ou i s > ( 
9^. e 4 M Sx \ 
be Se T€ ee n. 
       
    
€ 
  
  
être 
% Collapsed Bulldings div 
A. 0-6 25 
B: 62512.5 
124 N 
é HE D: 25.6 1 
{ € E 40-166 
Golcuk — : Sunk 
(a) (b) 
  
Figure 3. (a) Delineation of the 70 sample zones in Golcuk. (b) 
Ground observations of building collapse (AIJ, 1999) 
3.1.2 Damage location: Figure 4 depicts damage maps for 
Golcuk, obtained using the various measures of change. The 
difference scene in Figure 4a is classified to highlight regions of 
Golcuk exhibiting pronounced differences in reflectance. 
Changes are concentrated in the city center, with strongly 
negative values suggesting a marked increase in reflectance 
following the earthquake. With reference to the damage map in 
Figure 3b, these areas correspond with zones of severe building 
damage (categories D-E). This result suggests that debris 
612 
Interna 
associa 
return 
damage 
differer 
smoke 
upper à 
coastal 
widesp 
  
  
  
  
Figure 4 
and lo 
which fr 
dam 
Results f 
and Fig 
visualiza 
since the 
interest. | 
of low « 
Golcuk ( 
damage | 
widespre