Yosuke Ito 
  
EXTRACTION OF DAMAGED REGIONS USING SAR DATA AND NEURAL NETWORKS 
Yosuke ITO’, Masafumi HOSOKAWA fT, Hoonyol LEE}, Jian Guo LIU} 
"Department of Civil Engineering, Takamatsu National College of Technology, Japan 
ito @takamatsu-nct.ac.jp 
tTEarthquake Disaster Section, National Research Institute of Fire and Disaster, Japan 
hosokawa@fri.go.jp 
tT. H. Huxley School of Environment, Earth Science and Engineering, 
Imperial College of Science, Technology and Medicine, U.K. 
hoonyol.lee@ic.ac.uk, j.g.liu@ic.ac.uk 
  
  
Working Group III/4 
KEY WORDS: Classification, Coherence, Decorrelation, Earthquakes, Interferometry, LVQ, Multi-temporal. 
ABSTRACT 
This paper presents classification results using neural networks based on INSAR coherence imagery data for evaluation of 
the damage of Kobe earthquake in 1995. Coherence derived from multi-temporal SAR data before and after an earthquake 
presents a temporal decorrelation in disturbed regions. As L-band SAR data is more robust than C-band SAR data for 
spatial and temporal decorrelation, we used multi-source and temporal SAR coherence images derived from interferomet- 
ric pairs of JERS-1 and ERS-1 single look complex images (SLCs). Hazard areas can be estimated by classifying two 
categories defined as the damaged regions and otherwise using set of the coherence images. A neural classifier was used 
because of requiring no assumption for probability distribution function of each category. A competitive neural network 
trained by the learning vector quantization (LVQ) was adopted to the neural classifier in consideration of generalization 
ability and convergence. Total five coherence images were produced using effective interferometric pairs derived from 
two JERS-1 and four ERS-1 SLCs. The average coherence of JERS-1 is higher and has significantly higher contrast than 
that of ERS-1 even though the spatial decorrelation and the temporal separation are nearly equal. A hazard survey map 
was used for assessing extraction results. The LVQ method generated 2346 higher kappa coefficient by adding the JERS-1 
coherence and produced better results than the maximum likelihood method from the view point of balance of the number 
of the correctly classified pixels. 
1 INTRODUCTION 
It is an important research topic to identify regions damaged by an earthquake. Synthetic aperture radar (SAR) data, in 
particular, is a valuable information source to define and classify damaged regions as it is not affected by cloud cover and 
contains the phase information highly sensitive to surface change. A strong earthquake can cause tremendous destruction 
in an urban area such as structure collapse and liquefaction. Furthermore, a conflagration can burn many buildings in 
urban circumstances. Land surface of a damaged region can be dramatically changed by the earthquake. 
Coherence derived from multi-temporal SAR data before and after the earthquake presents temporal decorrelation. An 
antenna pattern depending on SAR instrument and water vapor in the atmosphere does not affect the coherence. Coherence 
can thus be utilized to detect changes caused by the earthquake and to identify the damaged regions (Yonezawa and 
Takeuchi, 1999). Coherence as a local correlation can be affected by temporal, spatial and thermal decorrelation factors 
(Zebker and Villasenor, 1992). The theoretical spatial decorrelation is a function of SAR system parameters and a target 
location such as wavelength, radar bandwidth, slant range, perpendicular baseline and local terrain slope (Lee and Liu, 
1999). L-band SAR data is more robust than C-band SAR data for the spatial decorrelation because the wavelength of 
L-band (23.53cm) is much longer than that of C-band (5.66¢m). An L-band coherence image is more useful in a case of 
an interferometric pair with long temporal separation (Fujisawa and Rosen, 1998). Unfortunately, JERS-1 was terminated 
on 11%" October 1998. There are only limited interferometric image pairs available as the result of poorer orbit control 
than ERS. 
Hazard areas can be estimated by discriminating two categories defined as damaged regions and otherwise, using coher- 
ence images. In general, it is difficult to assume a probability distribution function (PDF) of the coherence values in each 
category since these categories are determined based on land survey results and include various objects on the surface. It 
is thus effective to employ a non-parametric classifier that does not require any assumption for the PDF (Ito and Omall, 
1997). 
  
156 International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B1. Amsterdam 2000. 
In th 
mult 
vecti 
(Kol 
The 
area 
burn 
Tabl 
byc 
micr 
cont 
scen 
expe 
Sea : 
road 
with 
scen 
on tl 
Tabl 
inter 
whe; 
Wave 
and | 
Simi 
is hi 
caus 
even