Developing a building damage function using SAR images and post-event data after te Typoon Haiyan in Te Pilippines Bruno ADRIANO 1, Erick MAS 2 and Sunici KOSHIMURA 3 1 Member of JSCE, Graduate Student, Graduate Scool of Engineering, Tooku University (Aoba 468-1-E301, Aramaki, Aoba-ku, Sendai 980-0845, Japan) E-mail:adriano@geoinfo.civil.tooku.ac.jp 2 Member of JSCE, Assistant Professor, International Researc Institute of Disaster Science, Tooku University (Aoba 468-1-E301, Aramaki, Aoba-ku, Sendai 980-0845, Japan) E-mail:mas@irides.tooku.ac.jp 3 Member of JSCE, Professor, International Researc Institute of Disaster Science, Tooku University (Aoba 468-1-E301, Aramaki, Aoba-ku, Sendai 980-0845, Japan) E-mail:kosimura@irides.tooku.ac.jp A building damage function was developed using very ig-resolution syntetic aperture radar (VHR SAR) images and post-event building damage data after te 2013 Super Typoon Haiyan, obtained from Tacloban city, te Pilippines. Structural damage caused by te typoon was analyzed focusing on canges in te radar footprint signatures of affected buildings. Tese canges are classified using a pase-based coefficient calculated from pre- and post-event COSMO-SkyMed images. Te damage function is expressed as te damage ratio of structures wit regard to te pase-based coefficient. Te application of te estimated damage function in Tacloban city sows tat from a group of structures wit a POC coefficient of 0.2 or less, at least alf of tem are expected to be damaged. Key Words: damage function, cange detection, Typoon Haiyan, radar remote sensing 1. INTRODUCTION On November 8, 2013, te Super Typoon Haiyan it a portion of te Souteast Asia, in particular te sout region of te Pilippines. Te Joint Typoon Warning Center (JTWC) assessed te system as a Category 5 (super typoon on te Saffir-Simpson urricane wind scale). Te super typoon Haiyan traveled wit a sustained wind speed of 315 km -1 and made its first landfall wit a maximum wind speed of approximately 230 km -1. Te typoon made its first landfall over eastern Samar Island at 4:40 local time (LT), and its second landfall over Leyte Island at 7:00 LT (Fig.1a). As of April 17, 2014, reports on damage to people registered 6,300 dead, 28,689 injured and 1,061 missing 1). Spaceborne Eart observation is a valuable option in gatering information on te extension of damaged areas after occurrence of natural disasters. In particular, syntetic aperture radar (SAR) sensors are capable to observe te Eart's surface regardless te weater and dayligt conditions. Te new generation of SAR sensors suc as TerraSAR-X and COSMO-SkyMed (CSK) are able to provide detailed surface information for a single urban structure 2). Tis information can be used to evaluate te condition of building structures after natural disasters. In tis paper, we developed a building damage function for storm surge impact using SAR images and post-event building damage data after te Typoon Haiyan. Te structural damage is analyzed using te canges on te backscattering features of building footprints between pre- and post-event SAR images. 2. STUDY AREA AND IMAGERY DATA Te study area is Tacloban city, located approximately at 125 o 00''E, 11 o 15''N on te norteast of Leyte Island (Fig.1). Tacloban city, wit 221,174 inabitants 3), is te largest in population witin te Visayas region in te Pilippines. It was one of te ardest it places by te typoon, wic accounted for at least 2,500 of te approximately 6,300 deats I_1729
(a) (b) (c) 11 20'N Tacloban Samar Island Fig.4a Fig.4b Fig.4a Fig.4b Leyte Island 124 40'E 125 0'E [db] +7 [db] +7 (d) -39 0 250 m 0 250 m -43 (e) 0 250 m 0 250 m None Moderate Hig Total Survived Destroyed Fig.1 (a) Tacloban city. (b) Pre-event CSK data. (c) CSK data. (d) Building damage data provided by JICA in GIS-format. (e) Reclassified building damage used in tis study. Te blue dased line indicates te inundation boundary of te storm surge 7). Table 1 Number of building according to te damage level and reclassified damage level used in tis study. JICA Total 457 Hig 3 Moderate 1288 None 953 GTD Destroyed 460 Survived 2241 in te islands 4). Te Japan Society of Civil Engineers (JSCE) and te Pilippines Institute of Civil Engineers (PICE) conducted a collaborated field survey one mont after te disaster. Tey reported a maximum inundation of approximately 6 m at Tacloban city 5). A follow-up survey conducted by te International Researc Institute of Disaster Science (IRIDeS) team from Tooku University on mid January 2014, two monts after te event, reported at least 7 m of inundation dept at downtown Tacloban 6). Two SAR images from te CSK X-band sensor were used in tis study. A pre-event scene acquired on August 19, 2013 (Fig.1b) and a post-event scene acquired on November 20, 2013 (Fig.1c), approximately 12 days after Tacloban city was it by te typoon. After conducting pre-processing of te CSK images, tese were resampled at 0.2 m/pixel in a square size. (1) Building ground trut data In addition, te Japan International Cooperation Agency (JICA) provided a GIS file of building footprint data. Tese data include building damage level classification estimated troug visual interpretation of pre- and post-event ig-resolution optical satellite images. Te building damage level was classified in 4 categories: Total, wen structures were totally destroyed, wased or blown away. Hig, wen te roof ad been totally destroyed. Moderate, wen te roof ad been partially damaged. None, wen tere were not visible damage. In tis study, we selected te buildings inside te inundated area previously calculated in Adriano et al. 7). Finally, to construct te ground trut data (GTD), building damage levels were reclassified in 2 classes: Destroyed and Survived. Te spatial distribution of te reclassified damages levels, in te study area, as well as te number of buildings in I_1730
(a) Pre-event (g a ) (b) (g b ) Amplitude Amplitude +6-25 0 10 m (c) POC (ga,gb) = 0.12 (d) POC (ga,ga) = 1.0 1 1 0 0 Fig.2 Example of te calculated POC function. (a) Pre-event sample CSK data g b. (b) sample CSK data g a (c) POC function between g a and g b. (d) POC function between two identical sample data (g a ). eac category (Fig1d and Table 1) sow tat 80% of te buildings were not significantly damaged (survived), and were concentrated on te center of downtown Tacloban. Moreover, most of te destroyed buildings corresponded to te urban settlements located on te sout and nort of te study area. According to field surveys, tese areas were igly vulnerable due to non-engineering construction 7) 9). 3. METHODOLOGY (1) Pase-based cange detection coefficient Te pase-based detection metod used in tis study is te Pase-Only Correlation (POC) function 10), 11). Tis coefficient as sown good performance to detect damage areas using moderate-resolution satellite images 7). Te definition of te POC function is as follows. First, consider two images g a and g b. Let G a and G b denote te 2D Discrete Fourier Transforms (2D-DFT) of te two images, as sown by equations (1) and (2). { } G = F (1) a g a G b = F{ g b } (2) Te cross-spectrum between te 2D-DFT images is calculated by multiplying te element-wise G a and te complex conjugate G b (G * b), sown by equation (3), were denotes an element-wise product. Pre-event Fig.3 Scematic plots of te building footprint signature in a SAR image. Top and bottom figures sow cross-section and te 3D view of te radar backscattering. R G G * a b = (3) * Ga Gb Te POC function is given by te maximum value of te modulus of te 2D Inverse Discrete Fourier Transform (2D IDFT) of R, as sown by equation (4). POC = F 1 { R} (4) max Te most exceptional advantage of te POC-function compared to te ordinary correlation is its accuracy in image matcing 10), 11). In general, wen two images are similar, teir POC value gives a peak equal to 1.0. Tis fact can be interpreted as a low cange. On te oter and, wen two images present differences, te peak drops significantly. Tis can be interpreted as a ig cange. An example of te calculation of te POC values using a sample of te CSK data from te pre- and post-event images is sown in Fig.2. Te figure sows te calculated POC coefficient from a destroyed building and te details are sown in Fig.4b. Te resulting coefficient from te pre- and post-event samples indicates low correlation or ig canges (POC=0.11), wic is consistent wit te building damage level. In addition, te calculated POC value from two identical samples (POC=1.0) is also sown as a proof of concept. I_1731
0 30 m Poto Poto 0 30 m (a) 0 25 m (a) Pre-event (2) Radar remote sensing analysis Te canges of te backscattering features in damaged buildings, suc as its layover area, are strongly correlated to te extent of structural destruction 12). A scematic plot sowing te caracteristics of a non-damaged and damaged building radar footprint is sown in Fig.3. Te top figures sow a cross-section wit te radar backscattering properties and te reduction of te layover lengt ( ) in damaged structures. Te bottom figures sow te 3D view of te relation between te radar footprint or layover area and te building eigt (). Previous studies ave employed tis relation to estimate building eigts using ig-resolution SAR images and GIS data 13), 14). Based on Fig.3, te extent of te layover lengt is calculated using equation (5); wic is a function of te building eigt and te incidence angle () of te SAR image. = tan (5) In tis study, we estimated te layover area of non-damaged structures from te pre-event SAR scene. Ten tese areas are used to construct a mask applied in te post-event scene to conduct te cange detection analysis. (b) Pre-event Poto Fig.4 Pre- and post-event images and potos in-situ of eac damage classification. (a) Survived building (b) Destroyed building. Te solid polygons sow te initial GIS footprint, and te dased polygons sows te estimated layover-template. Te dark red arrows indicate te range and azimut direction of te radar sensor. Te layover area for a single building is estimated by sifting its original GIS-footprint in te direction of te sensor, as sown in Fig.4. Te number of times in te sifting process depends on te building eigt and te location of te building wit respect to oter buildings and streets. For instance, Fig.4a sows a 2-story building; in tis case, te initial footprint was sifted 2 times. Conversely, Fig.4b sows a wareouse of approximately 15 m eigt; ten, in tis case te initial footprint was sifted 4 times. Ten, to construct te mask of layover areas, te estimated layover is spatially added to te original footprint, as sown by te solid and dased polygons in Fig.4. Finally, te mask, or layover-template, is used to evaluate te canges on te radar backscattering caracteristics between te preand post-event SAR images. Te cange detection analysis is based on te POC function tat is calculated from te pixels inside te mask. Examples of te cange analysis are sown in Fig.4. For instance, a survived building in Fig.4a results on a POC value of 0.8 (low canges); tis is consistent wit te damage observed in te post-event poto. Conversely, te POC value of a destroyed building in Fig.4b resulted in 0.11 (ig canges); tis is also consistent wit te damage level interpreted in te poto. (3) Damage function for building damage Building damage function, traditionally, as been developed to identify structural vulnerability against strong ground motion witin te risk analysis of building structures. An extension of tis concept was apply to introduce building damage functions or tsunami fragility functions for tsunami impact 14). Tis function gives te probability of structural damage wit regard to te ydrodynamic features of tsunamis, suc as inundation dept and flow velocity. In tis study, we construct a damage function to evaluate te building damage ratio based on te canges on te radar footprint signature, wic is defined by te POC coefficient. To construct te building damage function, we assume tat te cumulative probability P of occurrence of te damage is given as equation (6). x P(x) =1 (6) were is te standardized normal distribution function, x is te POC coefficient; μ and are te mean and standard deviation of x, wic are obtained from equation (7). x = 1 + (7) I_1732
0.7 (a) GTD 0.6 In-situ poto in Fig.7 Damage ratio 0.5 0.4 0.3 = 0.1583 = 0.3522 0.2 None Moderate Hig Total 0.1 0 250 m 0.0 0.0 (b) Model 0.2 0.4 POC coefcient Fig.5 Damage function for building destruction, in terms of te canges of te footprint backscattering features between te pre- and post- event SAR images, wic is expressed by te POC function. Te circles indicate te distribution of te damage ratio. 0.6 0.8 1.0 In-situ poto in Fig.7 To calculate te mean and standard deviation in equation (7), te buildings are sorted in increasing order according to te POC coefficient values. Ten, we divide into groups of 90 buildings in eac sample. Next, a damage ratio is calculated by dividing te number of destroyed buildings over te total number of buildings in eac sample. Te corresponding cange coefficient for eac group is given by te average value of te POC coefficient witin te range of te sample. Finally, te two statistical parameters of damage function (μ and ) are obtained by plotting x and te inverse of 1 on normal probability papers, and performing te least-square fitting of tis plot. Trougout te regression analysis, te parameters are determined as sown in Fig.5. Te proposed function does not sow damage ratios greater tan 70% because tere are few buildings in te wole sample. Neverteless, tis curve suggests tat at least 50% of a group of structures may be damaged wen teir corresponding POC coefficient are less tan 0.2. Te proposed function can be used as a measure to assess te damage due to te potential storm surge and waves impact. Finally, to verify te applicability of te proposed damage function to conduct damage detection, te building damage were classified into 4 damage levels following te ground trut data (GTD) provided by JICA: (i) 0.00 to 0.15 for None damage, (ii) 0.16 to 0.30 for Moderate damage, (iii) 0.31 to 0.50 for Hig damage, and (iv) 0.51 to 1.00 for Total damage (Fig.6). In general, te spatial distribution of building damage using te damage function (Model) correlates wit te damage levels Damage ratio 0.51-1.00 0.31-0.50 0 250 m 0.16-0.30 0.00-0.15 Fig.6 Application of te estimated damage function. (a) GTD used in tis study. (b) Building damage ratio estimated using te proposed damage function. Te wite ellipse sows te area were te damage function cannot reproduce te interpreted damage levels. Tis migt be due to te rapid and informal reconstruction of tis area. Fig.7 In-situ potos from te areas tat were informally reconstructed few days after te event, wic migt cause te false detection from a post- SAR image, acquired almost two weeks after te event. observed in te GTD. In particular, te damage levels at te center of downtown Tacloban are consistently lower in te Model and te GTD; tis was also observed in post-event field surveys 6), 8). Conversely, te area inside te wite ellipse in Fig.6 sows low damage ratio, wic is not consistent wit te GTD. Tis migt be because footprint sizes of ouses located at tis area are relatively small to be well represented in te SAR images. In addition, based on te post-event field survey 6), 8), te ouses at tis area were rapidly reconstructed few days after te I_1733
event, wic migt cause te false detection from a post- SAR image acquired almost two weeks after te event wen several ouses were informally reconstructed, as sown in Fig.7. 4. CONCLUDING REMARKS A building damage function was constructed using te ig-resolution SAR data (COSMO-SkyMed images) and post-event building damage data for te 2013 Typoon Haiyan tat it te city of Tacloban. Te pase-based cange detection metod was used to conduct te cange analysis. Te damage function sows te damage ratio of structures wit common POC coefficient values. Te damage function can be used as a measure to assess te possible extent of damage due to typoon impacts. However, note tat te damage function developed in te present study is from te typoon tat it Tacloban city, wic includes storm surge and wave impacts. Terefore, tey migt not be applicable as a universal measure of typoon impact or damage and is expected to serve as a tool for damage estimation of future events in Tacloban city. ACKNOWLEDGMENT: Tis researc was supported by te J-RAPID program from te Japan Science and Tecnology Agency (JST) and IRIDeS Researc Grant (A-1). REFERENCES 1) National Disaster Risk Reduction and Management Council: Effects of Typoon Yolanda ; SitRep: No108-03APR2014. 2014. 2) Ferro A., Brunner D., and Bruzzone L.: Automatic Detection and Reconstruction of Building Radar Footprints From Single VHR SAR Images, IEEE Trans. Geosci. Remote Sens., vol. 51, no. 2, pp. 935 952, 2013. 3) National Statistics Office: Population and Annual Growt Rates for Te Pilippines and Its Regions, Provinces, and Higly Urbanized Cities; 2010 Census and Housing Population. 2010. 4) Maar A., Lagmay F., Agaton R. P., Allen M., Baala C., Brianne J., Briones L. T., May K., Cabacaba C., Vincent C., Caro C., Dasallas L. L., Anne L., Gonzalo L., Ladiero C. N., Pillip J., Teresa M., Mungcal F., Victor J., Puno R., Marie, Ramos A. C., Santiago J., Kennet J., and Tablazon J. P.: Devastating storm surges of Typoon Haiyan, International Journal of Disaster Risk Reduction, vol. 11, pp. 1 12, 2014. 5) Tajima Y., Yasuda T., Paceco B. M., Cruz E. C., Kawasaki K., Nobuoka H., Miyamoto M., Asano Y., Arikawa T., Ortigas N. M., Aquino R., Mata W., Valdez J., and Briones F.: Initial Report of Jsce-Pice Joint Survey on te Storm Surge Disaster Caused By Typoon Haiyan, Coast. Eng. J., vol. 56, no. 01, p. 1450006, 2014. 6) Mas E., Bricker J., Kure S., Adriano B., Yi C., Suppasri A., and Kosimura S.: Survey and satellite damage interpretation of te 2013 Super Typoon Haiyan in te Pilippines, Nat. Hazards Eart Syst. Sci., vol. 15, pp. 805 816, 2015. 7) Adriano B., Gokon H., Mas E., Kosimura S., Liu W., and Matsuoka M.: Extraction of damaged areas due to te 2013 Haiyan typoon using ASTER data, in IGARSS 2014 and 35t CSRS, IEEE Geoscience and Remote Sensing Society (GRSS) and te Canadian Remote Sensing Society (CRSS), pp. 2154 2157, 2014. 8) Mas E., Kure S., Bricker J., Adriano B., Yi C., Suppasri A., and Kosimura S.: Field survey and damage inspection after te 2013 Typoon Haiyan in Te Pilippines, in Coastal Engineering Conference of te Japan Society of Civil Engineers Vol. 70, No. 2, p. I_1451 I_1455, 2014. 9) Bricker J., Takagi H., Mas E., Kure S., Adriano B., Yi C., and Roeber V.: Spatial variation of damage due to storm surge and waves during Typoon Haiyan in te Pilippines, in Coastal Engineering Conference of te Japan Society of Civil Engineers Vol. 70, No. 2, p. I_231 I_235, 2014. 10) Takita K., Aoki T., Sasaki Y., Higuci T., and Kobayasi K.: Hig-accuracy subpixel image registration based on pase-only correlation, IEICE Trans. Fundam. Electron. Commun. Comput. Sci., vol. E86-A, no. 8, pp. 1925 1934, 2003. 11) Ito K., Nakajima H., Kobayasi K., Aoki T., and Tatsuo H.: A fingerprint matcing algoritm using pase-only correlation, IEICE Trans. Fundam. Electron. Commun. Comput. Sci., vol. E87-A, no. 3, pp. 682 691, Nov. 2004. 12) Gokon H., Post J., Stein E., Martinis S., Twele A., Muck M., Geiss C., Kosimura S., and Matsuoka M.: A Metod for Detecting Buildings Destroyed by te 2011 Tooku Eartquake and Tsunami Using Multitemporal TerraSAR-X Data, IEEE Geosci. Remote Sens. Lett., vol. 12, no. 6, pp. 1277 1281, 2015. 13) Yamazaki F., Liu W., Mas E., and Kosimura S.: Development of building eigt data from ig-resolution SAR imagery and building footprint, in Safety, Reliability, Risk and Life-Cycle Performance of Structures & Infrastructures Deodatis, Ellingwood & Frangopol (Eds), pp. 5493 5498, 2013. 14) Liu W., Yamazaki F., Adriano B., and Mas E.: Development of Building Heigt Data in Peru from Hig-Resolution SAR Imagery, J. Disaster Res., vol. 9, no. 6, pp. 1042 1049, 2014. 15) Kosimura S., Oie T., Yanagisawa H., and Imamura F.: Developing fragility functions for tsunami damage estimation using numerical model and post-tsunami data from Banda Ace, Indonesia, Coast. Eng. J., vol. 51, no. 3, pp. 243 273, 2009. (Received Marc 18, 2015) I_1734