You are on page 1of 7

IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 8, NO.

10, OCTOBER 2015 4935

A Soft Computing Method for Damage Mapping


Using VHR Optical Satellite Imagery
Babak Mansouri and Yaser Hamednia

Abstract—In this research, the feasibility of a method based high-resolution (VHR) optical data and damage detection algo-
on some soft computing algorithms for earthquake damage map- rithms were proved to be effective in estimating earthquake
ping is sought. The idea is to classify different patterns of change damages in urban areas [1]–[4].
associated with building footprints and to detect distinct damage
levels. A fuzzy inference methodology is employed to determine Pixel-based, object-based, or a combination of these image
the damage grade for individual building roofs by the means of processing approaches have been developed and applied in
evaluating the contribution of different patterns of changes. For urban classification and change/damage detection. In pixel-
implementation, satellite images of before and after the 2003 Bam, based schemes, image processing makes use of spectral, spatial,
Iran earthquake, are used in addition to some available ancillary or spectral–spatial hybrid features within the images. Huang
data. Building footprint pixels were extracted from pre- and pos-
timages using the ancillary building mask. Haralick second-order et al. [5] presented a hybrid feature classification and extrac-
textural features were computed for the building objects and an tion method where statistical features of the direction lines’
optimum set of such features was selected using genetic algorithm histogram were extracted from high-resolution multispectral
(GA). Considering optimal indices, different parts of roofs were imagery and after applying a dimension reduction step, some
classified into three change patterns as “low change,” “moderate suitable classifiers were presented. In object-based techniques,
change,” and “severe change” employing a support vector machine
(SVM) algorithm. For each building footprint, the contribution of the change/damage detection algorithm is applied to individ-
each class was calculated as the input of a fuzzy inference system ual urban scene components or objects, namely buildings, road
(FIS). Mamdani fuzzy engine was used to determine the damage segments, bridges, etc., as a whole. Since there is a tendency
grade of each building. The proposed algorithm was evaluated for change/damage detection algorithms to be fine-tuned for
by comparing the produced damage map with a reference dam- object contents (pixels representing the objects only), pixels
age map (ground truth). The results demonstrated the efficacy of
the method showing overall accuracies of 76% for detecting three outside the objects of interest are effectively left out, thus gener-
levels of structural damage (no to slight, moderate, and heavy to ally providing better results. Object-based methods can reduce
destruction) and 89% for determining binary damage levels (no- the effects of noises, minute shifts, or pixel misregistration of
collapsed, and collapsed) as suitable for such overall monitoring images leading to more accurate results [1], [6], [7]. It should
process. be noted that some pixel-object joint processing frameworks
Index Terms—Fuzzy logic, genetic algorithm (GA), image have been also developed. For example, Huang and Zhang
texture analysis, satellite applications, urban areas. [8] developed a support vector machine (SVM) approach for
the classification of high-resolution remotely sensed imagery
I. I NTRODUCTION where the spectral–spatial and semantic features are inte-
grated in the pixel and object levels to accurately classify the

N ATURAL disasters are affecting the lives of millions of


people around the world each year. It is possible to avoid
or reduce the impact of disastrous events such as earthquakes
high-resolution imagery. Gusella et al. [2] provided an object-
oriented methodology using eCognition image processing soft-
ware [9] to carry out damage detection by supervised method
with effective disaster management strategies. An effective identifying “collapsed” versus “no-collapsed” buildings. Their
monitoring procedure is essential to evaluate the magnitude and reported producer accuracy (PA) was 67.4% for “collapsed”
the extent of urban damage for disaster management activi- and 74.4% for “no-collapsed” cases. Also, the user accuracy
ties. Remote Sensing and GIS technologies are progressively (UA) was 78.5% for “collapsed” and 62.5% for “no-collapsed”
advancing in disaster damage assessment and mapping. Very buildings. Mansouri et al. [10] produced and compared cross-
power and coherence maps for the 2003 Bam earthquake using
Manuscript received January 26, 2015; revised September 07, 2015; accepted
six ENVISAT/ASAR interferometric image pairs. The results
October 06, 2015. Date of publication November 22, 2015; date of current
version January 08, 2016. This work was supported in part by the International were aggregated into grids of 25 pixels × 25 pixels and the
Institute of Earthquake Engineering and Seismology internal contract #AM outcome was compared to a zonal damage map that showed
7-269 (15/5/90), and in part by the GEM-EMME (Earthquake Model of Middle an increasing trend of detection for higher levels of actual
East) project.
B. Mansouri is with the Department of Emergency Management,
damages. In another research using SAR imagery, earthquake
International Institute of Earthquake Engineering and Seismology, Tehran damage detection was performed on building parcels as objects
19537-14453, Iran (e-mail: mansouri@iiees.ac.ir). [11]. Hoffman [12] used interferometric coherence changes in
Y. Hamednia is with the International Institute of Earthquake Engineering ENVISAT/ASAR images as a measure to detect building dam-
and Seismology, Tehran 19537-14453, Iran (e-mail: y.hamednia@iiees.ac.ir).
Color versions of one or more of the figures in this paper are available online
ages in the 2003 Bam earthquake. A coherence change index
at http://ieeexplore.ieee.org. was quantitatively interpreted as to describe the building dam-
Digital Object Identifier 10.1109/JSTARS.2015.2493342 ages at block level. It was found that coherence index could
1939-1404 © 2015 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
4936 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 8, NO. 10, OCTOBER 2015

not be interpreted at the pixel level. Chini et al. [13] exploited TABLE I
SAR and VHR optical satellite images to quantify damage H ARALICK T EXTURAL F EATURES F ROM GLCM M ATRIX
caused by the 2003 Bam earthquake. They studied the sensi-
tivity of different parameters as extracted from remotely sensed
images to the levels of actual damage as obtained from ground
survey map. Using VHR optical data, building extraction was
realized by the application of morphological operators com-
posed of opening and closing operations and employing an
unsupervised Isodata classifier. Damage mapping was at pixel
scale according to normalized difference of the radiances. There
method indicated that for different levels of damage, signifi-
cant increase (52% or more) in damage detection defined as
rate of damaged pixels was observed when the building mask
was utilized. Chini et al. [14] combined visual inspection and
automatic classification tools using QuickBird panchromatic
images related to 2005, Kashmir, Pakistan earthquake to detect
damages, ruptures, and landslides. For building damage detec-
tion, they used the same method as described in [13]. Tiede P [i, j]: co-occurrence probability of pixel pairs with i
et al. [15] developed rulesets in an object-based image anal- and j.
ysis (OBIA) environment [cognition network language (CNL)
by Definiens] for automatic building damage extraction using
illumination situation of both images in overall sense), eight
VHR data for the 2010 Haiti earthquake. The damage assess-
second-order histogram Haralick [18] textural features were
ment algorithm was based on the shadow indicator as casted by
extracted according to definitions listed in Table I, and an
buildings before and after the earthquake. Also, a vegetation
optimum set of three such features were selected using the
mask has been created to delineate the results from vegeta-
combination of GA and SVM as will be explained in upcom-
tion shadows for both images. Corbane et al. [16] provided a
ing parts. The classification of each building footprint (roof)
comprehensive report on the analysis of building damage in the
was completed for three distinct classes describing three dam-
2010 Haiti earthquake using VHR satellite and aerial imageries.
age patterns. These patterns of change (low change, moderate
This study mainly focused on the use of remotely sensed images
change, severe change) represent subregions for individual
and some techniques involved for visual damage interpretation
building footprint as determined by a supervised classification
and damage mapping by a number of researchers. Such idea
method employing SVM. Finally, a fuzzy inference methodol-
was supported through the GEOCAN consortium (2010) for the
ogy is used to determine the damage grade of each building
main purpose of setting some global standard for rapid dam-
(footprint) by means of evaluating the contribution of distinct
age mapping utilizing crowd sourcing through the web. Chen
change patterns. As follows, the details of the proposed method,
and Hutchinson [17] have presented a probabilistic classifica-
the implementation steps, and the results are reported for the
tion framework by means of multiclass classifier for damage
2003 Bam earthquake as our case study.
mapping of buildings. Three damage levels were assigned to
individual structures where the train data were visually selected
and the remaining structures were predicted using the relevance
vector machines (RVM) classification procedure. II. M ETHODOLOGY-DAMAGE M APPING
Although the study of advanced method for building extrac- The flowchart of the proposed method is shown in Fig. 1.
tion is very important, we supposed that this is feasible either In this study, “before” and “after” QuikBird VHR data of the
by some sophisticated automated or semi-automated meth- 2003 Bam earthquake and a building mask are considered. The
ods or by compiling actual building data (i.e., geo-databases) initial step involves the preprocessing of the satellite images
whenever available. In this research, we mainly focused on and consequently a difference image was made by subtracting
the development of a change/damage detection scheme as a the after from the before panchromatic data. Also, a building
combination of some soft-computing methods [genetic algo- mask is made using a vector map as ancillary data.
rithm (GA), SVM, and fuzzy inference system (FIS)] trying Haralick textural images were created and the textural dif-
to present a systematic and potentially rapid damage detec- ference images were produced by subtracting textural image of
tion scheme using VHR optical satellite imagery. However, “after” from “before” in eight textural bands. Then, a set of
for general applicability, the robustness and the timeliness of three optimum features was determined using a combination of
such processes depend mainly on the amount and the qual- GA and SVM procedures. Using train and test data, the building
ity of building data, amount of manual processes, computation roofs are represented within three distinct change levels using
facilities, prior validations, and calibration of the results consid- the SVM supervised classifier. Finally, by considering the con-
ering building attributes (i.e., roof types, building heights, roof tribution of pixels representing different patterns of change, an
materials, etc.) with respect to the imagery details. FIS is employed to determine the overall damage for each build-
After preprocessing steps needed to coregister before and ing object. Further details of the proposed method are described
after satellite images and histogram matching (to match the in the following sections.
MANSOURI AND HAMEDNIA: SOFT COMPUTING METHOD FOR DAMAGE MAPPING USING VHR OPTICAL SATELLITE IMAGERY 4937

Fig. 2. Combined GA and SVM for optimal features selection.

Fig. 3. Change patterns in an arrangement of building footprints.

by different object-based computer codes such as eCognition as


reported in [2] or a method presented in [13] as beneficial for
semiautomated building extraction cases.
Fig. 1. Flowchart of the proposed method.

B. Textural Feature Extraction


A. Data and Preprocessing Extracting and processing textural features from optical
Panchromatic band of QuickBird satellite images was used satellite images showed merits in detecting built environ-
for acquisition dates of September 30, 2003 and January 3, ment change/damage in hard-hit areas after earthquakes. Two-
2004 for “before” and “after” Bam earthquake, respectively. dimensional gray-level co-occurrence matrices (GLCM) are
Considering the desert environment and the acquisition dates, widely used in textural analysis because they are able to detect
no substantial vegetation (dominated by palm trees) changes the spatial dependency of gray-level values within an image
were observed. Moreover, the building mask is not (or very [18]. The mathematical representations for eight second-order
minimally) overlapping with the vegetation canopy. Only the Haralick features are shown in Table I. These textural indices
panchromatic band of the temporal images was used because are computed considering a GLCM of 5 pixels × 5 pixels with
of its high spatial resolution and also because minimal num- 0◦ , 45◦ , 90◦ , and 135◦ orientations considering one pixel off-
ber of datasets was intended to be used. The preprocessing step set. A final eight-band textural difference image was produced
involved histogram matching (to match the illumination situa- by subtracting temporal textural images for the average of the
tion of both images in overall sense) and geo-referencing with four aforementioned directions.
about fifty ground control points (i.e., tie points) spread evenly
within the images. Also, a difference image was created by sub-
C. Optimal Feature Selection-Change Pattern Classification
tracting the post-event from the pre-event images. Observing
the “difference image” of the temporal panchromatic bands, GA is widely used as an optimization method. The opti-
three main patterns are conveniently observed that can be inter- mum feature selection is completed employing GA and SVM
preted as areas of low changes, moderate changes, and severe classification in tandem. An optimal set of three (out of eight
changes. An alternative method is comparing before and after candidates) textural features was selected by GA where the fit-
images directly as done in [19]. We are convinced that visual ness function was set as the classification accuracy in the SVM
inspection of the difference image is more convenient and faster procedure (Fig. 2).
to process for supervised classification. SVM is a powerful supervised classification method in which
Building footprints were considered for change evaluation relatively accurate results can be obtained using limited num-
and damage assignment. Therefore, extracting image features ber of training data. SVM procedure is used as to delineate
associated with building footprints was essential. For this pur- between three classes for different change patterns, namely
pose, a building mask was generated initially from an ancillary “low change,” “moderate change,” and “severe change” for roof
dataset where building footprints were extracted by visual pixels. Radial basis function (RBF) kernel is utilized in this
inspection from VHR satellite images [19]. This information study as widely used in land cover classification by satellite
was complemented by some ground truth and 1:2000 digital images as it has produced better results as compared with other
map of the region [1] noting that both processes were visual kernels [20]. Fig. 3 depicts an arrangement of some neighboring
and manual. It is also noted that similar masks can be generated buildings where individual roofs are represented with patches of
4938 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 8, NO. 10, OCTOBER 2015

TABLE II
B UILDING DAMAGE C LASSIFICATION ACCORDING TO EMS-98 [21]

Fig. 4. (a) Pre-event and (b) post-event images of Bam (QuickBird).

we generally have omitted to take into account the influence of


the amount of debris around buildings. Also, we have supposed
that observing the roof pixels only, and not the debris spread
around the footprints, D4 and D5 are practically undistinguish-
able, so they have been combined together as well. It is noted
that three ranges of structural damage as “no to slight,” “moder-
ate,” and “heavy to destruction” or alternatively “no-collapsed”
and “collapsed” (binary damage levels) have been widely used
three different patterns for change. Each change pattern is clas-
in disaster management and disaster damage mapping.
sified according to the SVM approach considering the training
data obtained from the temporal difference image. The main
idea is to estimate the final building damage grade by assessing
the contribution of the above-mentioned three change patterns III. I MPLEMENTATION , R ESULTS , AND D ISCUSSION
for every building object. This final phase of damage detection In this study, the panchromatic bands of pre-event and post-
is performed by FIS as it is described in the next part. event QuickBird satellite images were used. A part of the
bitemporal images is shown in Fig. 4. Considering the build-
ing geo-database and the derived building mask of Bam, 11294
D. Damage Detection–FIS building footprints were present as studied.
Fuzzy inferencing is the process of mapping some input vari- After extracting building objects from the mask and comput-
ables to one output variable using fuzzy logic. Generally, an ing the difference image, eight Haralick textural features were
FIS is comprised of three principal steps as fuzzification, infer- extracted from temporal images where an eight-band textural
encing, and defuzzification. Fuzzification involves division of difference image was produced. Applying the GA optimization
input feature space into fuzzy subspaces where each subspace is procedure in tandem with the SVM algorithm, three optimal
specified by a fuzzy membership function. Fuzzy rules are gen- features, i.e., “mean,” “dissimilarity,” and “second moment,”
erated by comparing the contribution of every change pattern were selected. The GA parameters were set according to pop-
with the actual damage grade for some representative buildings. ulation size of 50, crossover rate of 80%, and mutation rate of
Inferencing requires the assessment of the strength of each rule 5%. The fitness function of the GA algorithm was set as the
which is being triggered. Finally, defuzzification aggregates all difference between successive Kappa coefficients (1). The con-
triggered rules and generates a nonfuzzy output. vergence criteria for Kappa coefficient was taken below 1%
After completing the change pattern classification step, each difference as computed from the SVM classification for each
building object (roof) is represented by three distinct classes generation. The Kappa coefficient of the selected chromosome
(change patterns) as fed to the FIS for final building dam- was computed as 0.85 which shows high degree of observer
age detection. The membership functions for input and output agreement. In SVM implementation, the total numbers for train
variables were adjusted for some validated cases containing and test data were 500 and 1500 pixels per each class, respec-
820 buildings spread within the images. The fuzzy rules were tively. Each pixel represented spatially a small part of building
devised by comparing the contribution of different change objects attributed with a specific change pattern. These pixels
patterns as adjusted according to 25 reference buildings per were visually selected and labeled from the difference image
actual damage grades. This damage labeling is according to (panchromatic bands). In this study, RBF kernel was used for
the European Macroseismic Scale EMS-98 damage grades [21] classification with Gamma of 0.2 and Penalty Parameter of 50.
as defined in Table II. In a previous research, a damage map Using the SVM classifier and the selected three optimum
was generated by visual interpretation of similar VHR tempo- features, the textural difference image was classified into three
ral data [19] that is considered as our reference map. In that, change pattern classes as “low change,” “moderate change,” and
D1 and D2 classes were combined together as it was impossi- “severe change.”
ble to delineate between them. For D2, some damages could be As shown in Fig. 5, each building object is represented by
observed on the façade but not detectable when roof objects are possible three change patterns as distinctive parts. As shown
observed from above. Observable debris around the buildings in Table III, the contribution factor of each patch is tabulated
tended to show from D3. However, by using a building mask, for some example footprints. To produce the desired damage
MANSOURI AND HAMEDNIA: SOFT COMPUTING METHOD FOR DAMAGE MAPPING USING VHR OPTICAL SATELLITE IMAGERY 4939

Fig. 5. Change pattern classification with SVM.

TABLE III
R ELATIVE P IXEL P OPULATION FOR T HREE PATTERN C LASSES

Fig. 6. Membership functions for input variables (a: very low, b: low, c:
medium, d: high, and e: very high).

Fig. 7. Membership functions for output variable or damage (a: negligible to


slight, b: moderate to heavy, c: very heavy, and d: destruction).
map, the contribution of such patterns is regarded as input to
the FIS. The membership functions used for fuzzification and TABLE IV
defuzzification processes are depicted in Figs. 6 and 7. S OME S AMPLE F UZZY RULES FOR E STIMATING B UILDING DAMAGE
In order to derive the fuzzy rules, 25 building footprint were
selected randomly per each damage class as reference. Some
of these rules are shown in Table IV. Mamdani [22] fuzzy
engine was employed for this calculation for its generality of
application considering that the output membership functions
were supposed to be fuzzy sets. After the aggregation process,
a fuzzy set was obtained for each output variable that needed
defuzzification.
The output of the FIS is a damage label assigned to each
building object. The confusion matrix was calculated using the
results of damage detection as compared to a reference dataset
[19] as ground truth (Table V). Using this method, it was impos- D4D5) to obtain three prevalent damage states. Using the con-
sible to differentiate between D1 and D2; thus, D1 and D2 fusion matrix, a set of statistical indicators can be derived for
damage grades were combined together as D1D2. D1D2, D3, the analytical evaluation. The overall accuracy (OA) is simply
D4, and D5 damage grades represent “negligible to moderate computed by dividing the pixels that are correctly classified by
damage,” “substantial to heavy damage,” very heavy damage,” the sum of number of pixels in the confusion matrix. The pro-
and “destruction” according to EMS98 damage classification ducer’s accuracy (PA), showing the probability of a reference
(Table II). Furthermore, D4 and D5 were combined together (as sample being correctly classified by the classifier algorithm, is
4940 IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING, VOL. 8, NO. 10, OCTOBER 2015

TABLE V
C ONFUSION M ATRIX AND R ESULTED PA AND UA

Fig. 9. Modeled damage map of Bam.

detected. Previously, Chen and Hutchinson [17] and Gusella


et al. [2] have developed damage maps for Bam and Banda
Aceh. In the first study, a damage map was produced according
to extracted dissimilarity-based features on part of Bam (with
468 structures) and no quantitative performance of their model
was reported for Bam. However, for the case of 2004 tsunami of
Banda Aceh, Indonesia, the OA of 77.2% with 15.5% precision
was reported for “moderate damage.” In that scheme, the train
and test and the ground truth were selected visually from the
VHR data. The second study indicated that the supervised dam-
age detection (380 training samples) resulted in a PA of 67.4%
for collapsed and 74.4% for no-collapsed buildings. The UA
Fig. 8. Reference damage map of Bam, courtesy of [19]. was 78.5% for collapsed and 62.5% for no-collapsed buildings.
In order to reflect the results of the present study according
calculated by dividing the number of correctly classified pix- to binary damage states (collapsed and no-collapsed), D1 and
els in each class by the total number of pixels in that class D2 and D3 were combined together as D1D2D3 (no-collapsed)
(derived from the reference samples). The user’s accuracy (UA) and D4D5 (collapsed). The new error matrix indicated an OA
expresses the probability that a pixel on the image represents of 89%, PA of 86% (no-collapsed) and 91% (collapsed), and
that class on the ground and is computed by taking the ratio UA of 89% (no-collapsed) and 88% (collapsed). In this case,
between the number of correctly classified pixels in a class and OA, PA, and UA are all nearly the same and reasonably high.
the total number of classified pixels in that class. Kappa coef-
ficient (K) is a measure of actual agreement between the error
matrix (classification and the reference data as major diagonal) IV. C ONCLUSION
and the chance agreement indicated by row and column totals In this study, a novel method for damage mapping is
(marginal sums) as defined as proposed by extracting eight Haralick textural features from
r r panchromatic VHR optical satellite images associated with
N i=1 xii − i=1 (xi+ .x+i )
K=  r (1) building footprints. GA and SVM algorithms are used in tan-
N 2 − i=1 (xi+ .x+i )
dem to select an optimum feature set composed of Mean,
N is the total number of pixels, r is the number of rows of the Dissimilarity, and Second Moment to produce patterns of
confusion matrix, xii is the number of pixels in row i and col- change for building footprints in earthquake-hit areas. The FIS
umn i, and x+i and xi+ are the marginal sums of row i and provides a means to produce damage maps by converting tex-
column i, respectively. tural change patterns into some predefined damage levels. The
These values were calculated as shown in Table V. The ref- process was applied to the urban area of Bam considering the
erence damage map is shown in Fig. 8 and the derived damage 2003 earthquake data. In this research, a total of 11 294 build-
map is depicted in Fig. 9. The statistics of damage mapping ing objects (roof) were evaluated and their associated damage
indicates that the PA for D1D2, D3, and D4D5 are 62%, 58%, grades were determined. The proposed algorithm was evaluated
and 91%, respectively, where the UA accounts for 58%, 66%, by comparing the produced damage map with a reference data
and 86%, respectively. Considering PA and UA values for as ground truth showing overall accuracies of 76% for detecting
D1D2 and D3 classes, it is noted that the corresponding values three levels of structural damage (i.e., no to slight, moderate,
fall below the OA, indicating that this detection is less accurate and heavy to destruction) and 89% for determining binary dam-
as compared to the D4D5 cases. This seems logical as more age levels (no-collapsed and collapsed) as suitable for such
drastic feature differences (or changes) are more likely to be overall monitoring process. It is important to emphasize that
MANSOURI AND HAMEDNIA: SOFT COMPUTING METHOD FOR DAMAGE MAPPING USING VHR OPTICAL SATELLITE IMAGERY 4941

many known and unknown sources of error and uncertainty [14] M. Chini, F. R. Cinti, and S. Stramondo, “Co-seismic surface effects
are involved in different stages of the processes as many were from very highresolution panchromatic images: The case of the 2005
Kashmir (Pakistan) earthquake,” Nat. Hazards Earth Syst. Sci., vol. 11,
technically unavoidable. This scheme is partly based on visual no. 3, pp. 931–943, 2011.
interpretation but benefits from some robust classification and [15] D. Tiede, S. Lang, P. Fureder, D. Holbling, C. Hoffman, and P. Zeil,
inferencing techniques as suitable for a systematic damage “Automated damage indication for rapid geospatial reporting. An oper-
ational object-based approach to damage density mapping following
mapping. the 2010 Haiti Earthquake,” Photogramm. Eng. Remote Sens., vol. 77,
pp. 933–942, 2011.
[16] C. Corbane et al., “A comprehensive analysis of building damage in the
12 January 2010 Mw7 Haiti Earthquake using high-resolution satellite
R EFERENCES and aerial imagery,” Photogramm. Eng. Remote Sens., vol. 77, pp. 997–
1009, 2011.
[1] B. Mansouri, M. Mousavi, and K. Amini-Hosseini, “Parcel-based damage [17] Z. Chen and T. C. Hutchinson, “Probabilistic urban structural damage
detection using VHR optical data,” in Proc. 6th Int. Workshop Remote classification using bitemporal satellite images,” Earthquake Spectra,
Sens. Disaster Appl., Pavia, Italy, Sep. 2008, pp. 1–4. vol. 26, no. 1, pp. 87–109, Feb. 2010.
[2] L. Gusella, B. J. Adams, G. C. Bitelli, C. K. Huyck, and A. Mognol, [18] R. M. Haralick, K. Shanmugam, and I. Dinstein, “Textural feature for
“Object-oriented image understanding and post-earthquake damage image classification,” IEEE Trans. System Man Cybern., vol. 3, no. 6,
assessment for the 2003 Bam, Iran earthquake,” Earthquake Spectra, pp. 610–621, Nov. 1973.
vol. 21, no. S1, pp. 225–238, 2005. [19] F. Yamazaki, Y. Yano, and M. Matsuoka, “Visual damage interpretation
[3] B. J. Adams, B. Mansouri, C. K. Huyck, R. T. Eguchi, and M. Shinozuka, of buildings in Bam city using QuickBird images following the 2003Bam,
“Application of high-resolution optical satellite imagery for post- Iran earthquake,” Earthquake Spectra, vol. 21, no. S1, pp. 329–336,
earthquake damage assessment: The 2003 Boumerdes (Algeria) and Bam 2005.
(Iran) earthquakes,” Research Progress and Accomplishments: 2003– [20] J. Keuchel, S. Naumann, and M. Heiler, “Automatic land cover analy-
2004, Multidisciplinary Center for Earthquake Engineering Research sis for Tenerife by supervised classification using remotely sensed data,”
(MCEER), Univ. Buffalo, State Univ. New York, Buffalo, NY, USA, Remote Sens. Environ., vol. 86, pp. 530–541, 2003.
2004, pp. 173–186. [21] G. Grünthal, “European macroseismic scale 1998 (EMS-98),” Cahiers
[4] R. T. Eguchi, C. K. Huyck, S. Ghosh, and B. J. Adams, “The application du Centre Européen de Géodynamique et de Séismologie 15, Centre
of remote sensing technologies for disaster management,” in Proc. 14th Européen de Géodynamique et de Séismologie, Luxembourg, 99, 1998.
World Conf. Earthquake Eng., Beijing, China, Oct. 12–17, 2008. [22] E. H. Mamdani, “Advances in the linguistic synthesis of fuzzy con-
[5] X. Huang, L. Zhang, and P. Li, “Classification and extraction of trollers,” Int. J. Man-Mach. Stud., vol. 8, pp. 669–678, 1976.
spatial features in urban areas using high-resolution multispectral
imagery,” IEEE Geosci. Remote Sens. Lett., vol. 4, no. 2, pp. 260–264,
Apr. 2007. Babak Mansouri was born in Tehran, Iran, in 1965.
[6] C. K. Huyck, B. J. Adams, S. Cho, H.-C. Chung, and R. T. Eguchi, He received the B.S. degree in applied physics
“Towards rapid citywide damage mapping using neighborhood edge dis- from Tehran Polytechnic, Tehran, Iran, in 1993 and
similarities in very high-resolution optical satellite imagery—Application the M.S. degree in structural engineering, the M.S.
to the 2003 Bam, Iran earthquake,” Earthquake Spectra, vol. 21, no. S1, degree in electrical engineering, and the Ph.D. degree
pp. 255–266, 2005. in civil engineering from the University of Southern
[7] F. Yamazaki, D. Suzuki, and Y. Maruyama, “Detection of damages due California (USC), Los Angeles, CA, USA, in 1997,
to earthquakes using digital aerial images,” in Proc. 6th Int. Workshop 2000, and 2002, respectively.
Remote Sens. Disaster Appl., Pavia, Italy, 2008, pp. 1–6. He joined ImageCat Inc., Long Beach, CA, USA,
[8] X. Huang and L. Zhang, “An SVM ensemble approach combining as a Remote Sensing Senior Scientist for about 2
spectral, structural, and semantic features for the classification of high- years and served as a Visiting Researcher at the
resolution remotely sensed imagery,” IEEE Trans. Geosci. Remote Sens., University of California Irvine, Irvine, CA, USA. Since 2005, he has been with
vol. 51, no. 1, pp. 257–261, Jan. 2013. the International Institute of Earthquake Engineering and Seismology (IIEES),
[9] Definiens Imaging, e-Cognition 4.0, User Guide. Munich, Germany: Tehran, Iran, as a Faculty Member of Risk Management Research Center and
Definiens Imaging GmbH, 2004. is the Head of the Emergency Management Department. His research interests
[10] B. Mansouri, M. Shinozuka, C. K. Huyck, and B. Houshmand, include seismic risk assessment, loss estimation and damage mapping using
“Earthquake-induced change detection in Bam, Iran, by complex analysis advanced technologies and intensive computation.
using envisat ASAR data,” Earthquake Spectra, Earthquake Engineering
Research Institute (EERI), Oakland, CA, USA, Special Issue 1, 2005,
vol. 21, pp. S275–S284. Yaser Hamednia was born in Oroumieh, Iran, in
[11] B. Mansouri, M. Mousavi, and K. Amini-Hosseini, “Parcel-based dam- 1981. He received the B.Sc. degree in applied physics
age detection using SAR data,” in Proc. 6th Int. Workshop Remote Sens. from K.N.Toosi University of Technology (KNTU)
Disaster Appl., Pavia, Italy, Sep. 2008, pp. 1–5. University, Tehran, Iran, in 2012 and the M.Sc. degree
[12] J. Hoffman, “Mapping damage during Bam (Iran) earthquake using in civil engineering from the International Institute
interferometric coherence,” Int. J. Remote Sens.,” vol. 28, no. 6, pp. 1199– of Earthquake Engineering and Seismology (IIEES),
1216, 2007. Tehran, Iran, in 2015.
[13] M. Chini and N. Pierdicca, “Exploiting SAR and VHR optical images His research interests include risk assessment,
to quantify damage caused by the 2003 Bam earthquake,” IEEE Trans. algorithm development, and GIS and remote sensing
Geosci. Remote Sens., vol. 47, no. 1, pp. 145–152, Jan. 2009. applications in urban disasters.

You might also like