This document summarizes a study that used SAR (Synthetic Aperture Radar) data to detect parcel-level damage from the 2003 Bam, Iran earthquake. Researchers developed a parcel database from city maps and extracted building footprints. SAR images before and after the quake were analyzed at the parcel-level. A change detection index was calibrated using simulated radar cross-section curves for different building orientations. The calibrated SAR damage map correlated well with a visual damage interpretation from optical images, validating the parcel-based SAR approach for detecting earthquake damage.

1. 6th International Workshop on Remote Sensing for Disaster Applications
1
Parcel-based Damage Detection using SAR Data
Babak Mansouri, Kambod Amini-Hosseini,
and Reza Nourjou
Risk Management Research Center
International Institute of Earthquake Eng. and Seismology
Tehran, Iran
mansouri@iiees.ac.ir
Masanobu Shinozuka
Dept. of Civil and Environment Engineering
University of California, Irvine
California, USA
shino@uci.edu
Abstract— Remote sensing in general has the capabilities in
detecting some important phenomenon on the ground surface
with minimal knowledge of the study area. However, ancillary
site data help in reducing the errors and provide a base for result
validation and calibration. So far, pixel-based remote sensing
methods have been exploited by different research groups around
the world and the basic image processing schemes have been well
documented. Also some object-based (object-oriented) image
processing algorithms were developed for the purpose of
detecting and classifying objects on the ground. For example, in
urban areas where the physical changes to buildings are of
interest, in order to reduce the detection errors and minimize the
false alarms, it seems logical to apply any proper change
detection algorithms only to the patches that correspond exactly
to building parcels. This is even more crucial for the case of SAR
image processing because SAR returns are strongly sensitive to
the imagery geometry and features comprised within each pixel.
In this research, an urban database with parcel information is
developed from the city CAD files (for BAM). The parcels are
extracted from aerial photos (stereography processing) then
complemented and updated using very high resolution optical
data. The SAR change detection algorithm including the
calibration modeling are introduced and then applied to the
parcel layer. The results are then calibrated with the direct visual
damage interpretation data from VHR optical data.
I. INTRODUCTION
Remote sensing in general has the capabilities in detecting
some important phenomenon on the ground surface with
minimal knowledge of the study area. However, ancillary site
data help in reducing the errors and provide a base for result
validation and calibration. So far, pixel-based remote sensing
methods have been exploited by different research groups
around the world and the basic image processing schemes
have been well documented. Also some relatively new object-
based (object-oriented) image processing algorithms were
developed for the purpose of detecting and classifying objects
on the ground. For example, in urban areas where the physical
changes to buildings are of interest, in order to reduce the
detection errors and minimize the false alarms, it seems
logical to apply any proper change detection algorithms only
to the patches that correspond exactly to building parcels.
This is even more crucial for the case of SAR image
processing because SAR returns are strongly sensitive to the
imagery geometry and features comprised within each pixel.
The feasibility of change/damage detection using civilian
SAR satellites data with a ground resolution of about 20
meters (Envisat ASAR SLC image) is sought keeping in mind
that the rapid advancement in such technologies will deliver
much higher resolutions and better abilities to detect changes.
Very high resolution satellite SAR images are already
accessible through the Radarsat2 (3m in fine mode) and
Terrasarx (1m) systems and likely to be delivered vastly for
urban disaster applications.
Due to the remote sensing pre_ and post_event data
availability for the Bam earthquake, Envisat satellite data was
chosen. The sensor collected before- and after-event imagery
of the Bam, Iran earthquake that occurred on December 26th
,
2003. For this study, two sets of before and one after SAR
data are used. The change detection scheme evaluates these
results using orbital information to assess the levels of change
in different city parcels. Such damage maps can potentially
serve in disaster response/management and also in estimating
economic losses to urban settings. It is noted that in previous
researches [1] & [2] good results in identifying the regional
location of collapsed buildings were reported. Finally, a
damage map that was obtained from a direct visual damage
interpretation result is used to calibrate these findings at the
end.
II. METHODOLOGY
Figure 1 depicts the major steps involved in this research.
The parcel information are extracted from the aerial digital
maps and made GIS ready. Radar data for before and after the
event are coregistered and the SAR change index map is
extracted. All the data are then georeferenced. The change
index map is used in a way that only building parcels are taken

2. 6th International Workshop on Remote Sensing for Disaster Applications
2
into account and the pixels corresponding to the rest of the
features are filtered out. Also the SAR index is calibrated
based on the geometry of imagery and considering the most
visible walls of each parcel. Because the SAR processed index
is highly affected by the random noise, another layer namely
the city block layer was introduced so that the computed
indices are averaged for the parcels comprised in the block.
Figure 1 – Major steps involved in the algorithm
A. Change/damage index
The basic assumption for change detection using a
repeat-pass interferometric technique (single antenna but two
image acquisitions) is that scene distances to the receiving
antennas are generally the same. The interferometric phase is
then mainly affected by changes in the scattering behavior of
the scene, or changes in the scene geometry. In here,
interferometric data are used for creating SAR change index
map. Table 1 lists the baseline information between the
interferometric pairs used in this research.
Table 1 - Interferometric data pairs used in this study
( )( )
*
1 2
1 2
* *
1 1 2 2
CC
Coherence(C,C ) =
CC C C
∑
∑ ∑
(1)
Equation (1) is defined as the coherence between two complex
images; its denominator is defined as the cross-power (Xp).
When the same image is used in the cross-power formula it is
called the self-power (Sp) of the image. The sigma is
evaluated within a window of the size 3 pixels (in range) by
15 pixels (in azimuth). Window computations allow for
compensation of minute mis registrations of the data pairs and
for the reduction of inherent noises, which often occur at the
expense of reducing data resolution. It is best to compare
before-before and before-after interferograms, coherence
maps and X-powers that have similar baseline correlation.
The use of a common “before” dataset serves as a baseline
image. Coherence maps reflect scene/object changes that are
independent of the locality, largely because of the
normalization. For cross-power, strong backscattering (i.e.
corner reflectors) changes are more pronounced and more
suitable for urban damage assessment.
Nevertheless, the presence of false assignments,
random objects (moving object such as cars) and also feature
changes observed in the nature are unavoidable. Since the
level of radar returns is not only city specific but also sensor
and building orientation specific, an additional step of
averaging is applied to help summarize the difference values
contained in each parcel.
Figure 2 - Cross power difference as computed in a 3 pix . by 15 pix. window
Pairs used: (Jun-11-03,Feb-11-04) and (Jun-11-03,Dec-3-03)
B. Ancillary data – Parcel Information
The scope of this research is to compile high resolution
city data with parcel level of details including the city
topography and building height information and other attribute
data. The parcel maps and building height information were
extracted from 1:2000 scale digital maps provided by the
National Cartographic Center (NCC) of Iran. These maps were
created by processing aerial stereo-photographs. The extracted
Sensor-target plane
(m)
Baseline information
Data pairs
Normal Parallel
June 11, 03 Dec. 3, 03 473.21 147.98
June 11, 03 Feb. 11, 04 476.12 133.22
Master SAR
complex data
Slave SAR
complex data
co-registration
SAR change index
difference in cross-power
(computed in parcel layer)
Parcel-based
damage assessment
Aerial stereography
3D parcel extraction
(parcel layer)
SAR Imagery
Georeferencing
Parcel calibration
from simulated
RCS curves
(coefficient map)
comparison
with direct visual damage
interpretation from
VHR optical data
=> Calibrated Damage Map
~ 3 km

3. 6th International Workshop on Remote Sensing for Disaster Applications
3
0.000
1.000
2.000
3.000
4.000
5.000
6.000
0 45 90 135 180
city parcel information have been processed and compiled
from different data sets that needed both spatial adjustments
and temporal change considerations.
The urban parcel information is entered in GIS for the city of
Bam. Figure 3 shows a portion of this data that has been GIS-
ready and comprises of city parcel records pronouncing the
building footprints and building heights.
Figure 3 - A portion of the 1:2000 urban digital map comprising of parcel data
(original scene: 1.6 km by 1.2 km)
C. RCS Simulation
Urban environments can essentially be represented by a
combination of different geometrical shapes, i.e., rectangular
plates. The Envisat SAR system is consistent with a
monostatic measurement/simulation, i.e., the transmitter and
the receiver are regarded as the same antenna and located at
the same position with respect to the scene. It is expected that
after a building collapses, the backscattering coefficient of the
image is reduced drastically. The Radar Cross Section values
of the objects are highly sensitive functions of the sensor-
object observation and object azimuth angles.
The RCS simulation is performed for VV polarization
according to a vertical dihedral corner reflector and for each 1
degree azimuth angle increment to cover a full range of
possibilities. As can be imagined such reflectors intercept the
radar beam effectively. The effective area intercepting the
beam is a function of the incident and azimuth angle and also
the wall-ground area. Figure 4 is the computed RCS value (in
square meters) with respect to the azimuth angles.
Figure 4 - VV polarization angle dependent RCS simulation curve
for vertical dihedral reflector
D. Implementation in GIS
In order to apply the method for each parcel, the database
(parcel records) was refined as to filter out all the buildings
that are obscured. Moreover, analyzing each building footprint
sides and corners, and considering different angles, an
automated process selects the most radar detectable walls of
the building. The corresponding azimuth angle is stored for
each parcel record as seen in Figure 5. Then, the dedicated
algorithm estimates the SAR signature based on the angle
dependent RCS values for each parcel then computes the
calibration mask.
Figure 5 - Geodatabase analysis: detection of the most visible walls of the
parcel
The azimuth angles are attributed to the related parcel
record. Figure 6 shows the entire city, the optical very high
resolution data as the base map and the color-coded parcels
reflecting the azimuth angle. Angles around 82 degrees are
close to the maximum radar reception in general since the
satellite orbit is about 98 degrees near polar and the images are
acquired in the descending pass.
Figure 6 - Parcel azimuth angle for the most visible walls
III. RESULT
Since the nature of the radar data used is noisy and also
coarse in term of resolution, a city block mask was also used
in averaging the change detection results. Therefore, two
RCS (sm)
Azimuth angle (degrees)
A sample region

4. 6th International Workshop on Remote Sensing for Disaster Applications
4
masks namely the parcel layer and the block layer were used
in this research. As mentioned, the parcel layer reflects the
calibration coefficients and the building block layer reflects
the averaged SAR change index values. Figure 7 is the results
of a statistical classification of the calibrated values (sensor-
target and object orientation) of the change index as computed
using both mentioned layers.
Figure 7 –SAR change index calibrated with the parcel RCS coefficients
Yamazaki et al. (2005) [3] have created a damage map for
Bam by visual interpretation of the VHR Quickbird optical
data as shown in Figure 8. They have used the EMS-98
damage grades and the process of assigning different building
damage grades was fully manual. Table 2 summarizes their
results in addition to the assumed equivalent damage factor
ranges according to the ATC13 report. The ATC13 damage
factor values were used in order to quantify the results.
Table 2 – Visually interpreted damage grades and ATC13 damage factor
Figure 8 – Spatial distribution of visually interpreted damage grades [3]
0.00
0.10
0.20
0.30
0.40
0.50
0.60
0.70
0.80
0.90
0.00 20.00 40.00 60.00 80.00 100.00
SAR change index db( )
Damagefactorbyvisualinterpretation
Figure 9 – Scatter plot of damage factors versus visual interpretation and SAR
change index (trend line shows relatively good correlation)
Figure 10 – SAR change detection calibrated with the visual interpretation
Figure 9 is the scatter plot of the values read for each parcel
from the visual interpretation and the calibrated SAR damage
index. The data shows good correlation. The associated trend
line shown is used as the calibration curve in order to calibrate
the results shown in Figure 7 (statistical classification) with
the actual damage data (scaled with actual damage). The
calibrated damage map is computed and depicted in Figure 10.
ACKNOWLEDGMENT
IIEES is acknowledged for supporting the research
project # 327-8302. Also the support of the University of
Pavia, MCEER, EERI, UCI, IUSS, EUCENTRE is
appreciated. Envisat ASAR data was provided by the
European Space Agency. Professor Yamazaki is
acknowledged for providing the authors with the visual
damage interpretation data for the Bam earthquake.
REFERENCES
[1] B. Mansouri, M. Shinozuka, C. Huyck, B. Houshmand, “Earthquake-
Induced Change Detection in Bam, Iran, by Complex Analysis Using
Envisat ASAR Data”, Special Issue 1, Volume 21, Dec. 2005, S275,
Earthquake Spectra, Earthquake Engineering Research Institute (EERI),
Oakland, CA.
[2] B. Mansouri, and M. Shinozuka, “SAR image calibration by urban
texture: Application to the BAM earthquake using Envisat satellite
Assumed equivalent damage
factor in ATC13
centralrange
# of buildings
interpreted
Damage grade
assigned
5%1%-10%1597Grade 1&2
20%11%-30%3815Grade 3
45%31%-60%1700Grade 4
80%60%-100%4951Grade 5
1% - 10%
11% - 30%
31% - 60%
61% - 100%
Damage levels
SAR change index (db)
Grades1&2
Grade 3
Grade 4
Grade 5

5. 6th International Workshop on Remote Sensing for Disaster Applications
5
ASAR data”, 3rd International Workshop on Remote Sensing for Post-
Disaster Response, 12th
and 13th
September 2005, Chiba, Japan.
[3] F. Yamazaki, Y. Yano and M. Matsuoka, “Visual Damage Interpretation
of Buildings in Bam City Using Quickbird Images Following the 2003
Bam, Iran, Earthquake”, Earthquake Spectra, Special Issue 1, Vol. 21,
S329, December 2005.