3D Data Collection and Automated Damage ...

Construction Research Congress 2014 ©ASCE 2014

1209

3D Data Collection and Automated Damage Assessment for Near Real-time Tornado Loss Estimation 1

2

Alireza G. KASHANI , Andrew GRAETTINGER , and Thang DAO3

Downloaded from ascelibrary.org by OREGON STATE UNIVERSITY on 07/06/15. Copyright ASCE. For personal use only; all rights reserved.

1

PhD Candidate, Department of Civil, Construction and Environmental Engineering, University of Alabama, Box 870205, Tuscaloosa, AL 35487-0205, United States; email: [email protected] 2 Associate Professor, Department of Civil, Construction and Environmental Engineering, University of Alabama, Box 870205, Tuscaloosa, AL 35487-0205, United States; PH (205) 348-1707; FAX (205) 348-0783; email: [email protected] 3 Assistant professor, Department of Civil, Construction and Environmental Engineering, University of Alabama, Box 870205, Tuscaloosa, AL 35487-0205, United States; PH (205) 348-0726; FAX (205) 348-0783; email: [email protected]

ABSTRACT Tornadoes cause great hardship and economic loss to US communities each year. The lack of quantitative information associated with real damage states of individual buildings results in inaccurate loss estimates and hampers effective decision making for mitigation, response and recovery. A near real-time tornado loss estimation tool is developed and tested as part of this work. The GIS-based damage assessment tool employs post-event point cloud data collected by terrestrial scanners and pre-event aerial images. The tool automatically calculates the percentage of roof and wall damage at the individual building scale, which is used as input to empirical or statistical loss estimation methods. An accuracy analysis through a set of controlled experiments indicated that for typical point cloud density (>25 points/m2), the tool results in less than 10% error in detection of pre- and post-event roof/wall surfaces. The GIS-based tool was validated with datasets collected after the 2013 Moore, OK tornado and produced detailed percentage of damage for buildings, which was not provided by infield inceptions. INTRODUCTION Annually, approximately 1000 tornadoes hit United States and cause huge economic loss to the nation, which underline the need for proper management of resources for tornado mitigation, response, and recovery. Currently, the average loss due to tornadoes in United States is estimated to be $1 billion per year (Changnon, 2009), and losses due to tornadoes are increasing every year. Based on an insurance study, tornado events causing total losses of $1 billion are becoming more frequent (Folger, 2011). In 2011, over 1600 tornadoes occurred in the United States resulting in more than $25 billion in property damage (Prevatt et al., 2012). Accurate loss estimation in aftermath of tornadoes is important because decision makers in different organizations; such as federal and local governments, insurers, and tornado researchers, require loss information for important decisions about disaster assistance programs, costs and pay-outs, and for evaluation of

Construction Research Congress 2014



mitigation policies (Meade and Abbott, 2003; Okuyama and Chang, 2004). The government and insurers need to get accurate loss estimates as soon as possible to allocate and project funding resources needed for assistance programs and pay-outs. Trends and geographic patterns of losses are also used as a measure to evaluate the effectiveness of the past mitigation policies, engineering designs, and to determine future tornado research priorities (Rose, 2004; Downton and Pielke, 2005). Several studies demonstrated the inaccuracy of loss estimates after various natural disasters in the US (Downton and Pielke; 2005, Irving, 2008; Erdik et al, 2011). Comprehensive assessments of all damaged properties and refinement of inaccurate initial estimates are time intensive (Irving, 2008). Time needed for accurate loss estimates and data reliably issues highlight an evident need for efficient damage data collection and rapid loss estimation methods. Various limitations of field data collection and damage assessment lead to inaccurate loss estimates. Initial estimates for decision makers are prepared in the immediate aftermath of a disaster based on preliminary damage assessments and historical data related to the affected area. Forming inspection teams for damage assessment requires available funding and human resources. Resource limitations and access restrictions can hinder team arrival to damage sites. The shortage of time in preliminary assessments usually impedes comprehensive inspection of all damaged buildings. Therefore, comprehensive assessments are only performed on a limited number of sampled buildings, while for other buildings “windshield estimates” are often prepared by viewing damage from a car window (Downton and Pielke, 2005). Surveys on limited number of buildings increases the data granularity and negatively impacts on accuracy of the loss estimates. Automated assessment allows for surveying all damaged buildings and also reduces the bias made by the engineering reconnaissance team. Remote sensing is a solution to overcome limitations traditionally associated with the field data collection and damage assessment. Recently, airborne remote sensing technologies such as satellite and aerial imagery as well as airborne laser scanning have been implemented in aftermath of some recent disasters such as the hurricane Katrina (Dash et al., 2004; Eguchi et al., 2011; Li et al., 2008; Olsen et al., 2013; Ozisik & Kerle, 2004; Rehor, 2008; Womble et al., 2008). Spatial data is rapidly collected from large damaged sites and then manually or automatically processed to analyze disaster damage, which is appropriate for broad damage assessment required in the early hours and days after disasters. However, data collected by airborne sensors has limitations associated with loss estimation. The low resolution of data in comparison with ground-based data collection hinders the detailed damage assessment. In addition, since airborne sensors only capture the tops of objects; therefore, the level of damage to building walls and cladding is not collected. Terrestrial laser scanning technologies, including stationary or mobile scanners, have potential advantages compared to airborne scanning. Some studies have shown that such high point density is appropriate for structural deformation analysis and post-disaster damage assessment (Kashani et al., 2013; Olsen et al., 2010; Park et al., 2007). In addition, terrestrial scanners are able to scan building walls and cladding that are not detectable in airborne collected scans.


1210



This study formalizes and tests a novel approach for automatic tornado damage assessment and loss estimation employing post-event terrestrial laser scanning and pre-event aerial imagery. The automated damage assessment tool presented in this study includes custom Geographic Information System (GIS) models to overlay point datasets with image layers, automatically detect building component s, and calculate the percent of damage by comparing pre- and post-event geometry of roof and wall surfaces. Once damage of each individual building is quantified, the results can be used as input to empirical or statistical loss estimation methods (Vickery et al. 2006a and 2006b; van de Lindt and Dao, 2011). AUTOMATED GIS-BASED DAMAGE ASSESSMENT TOOL The inputs of the automated GIS-based damage assessment tool presented in this work are a post-event point cloud dataset and a pre-event aerial photograph. A point cloud dataset collected with a terrestrial laser scanner contains (x, y, z) coordinates of measured points, which is used to measure post-event geometry of buildings. An aerial photograph taken before the tornado is used to measure preevent geometry of buildings. Roof and wall damage percentages are calculated by comparing the pre- and post-area of roofs and length of walls.

Figure 1. Flow Diagram of the automated damage assessment tool The damage assessment tool includes three custom GIS models that automatically detect building roof and wall surfaces and calculate percent of damage. These GIS models are programmed using ArcGIS Model Builder and made of a combination of ArcGIS tools available in ArcGIS 10.0 (ESRI). Figure 1 shows the sequence of the GIS models as well as inputs and outputs. The first model detects roofs in the point cloud and generates post-event roof polygons. The second model detects roofs in the aerial image and generates pre-event roof polygons. Finally, the third model detects walls in the point cloud and generates post-event wall polylines. As shown in Figure 1, the results of each model are used as input for the subsequent model. These custom models are described in the following sections. Post-event Roof Detection Model Post-event roof detection model generates polygons representing the remaining portions of building roofs after damage. The top row of Figure 2 illustrates


1211



the analysis procedure of the post-event roof detection model. The input is the postevent point cloud data. First, the point cloud data is converted to a 2D raster dataset in which each cell saves the minimum Z value of the points within that cell. The minimum cell assignment allows for filtering trees and light poles from the datasets. Next, raster cells representing building roof surfaces are identified by a slope-based filtering analysis (Vosselman, 2000). In the slope-based analysis, a window, as shown in the top row second column in Figure 2, is slid on the raster dataset, and the slope of vectors connecting each raster cell to the other cells in the window is compared with a threshold value set larger than the surrounding ground surface slope. Therefore, roof cells are identified in the raster where the vector has a steeper slope than the threshold value set based on the surrounding ground surface slope and typical roof pitches in the area. In typical residential subdivisions a threshold slope of 30% results in good classification, since the ground slope and the street grade is often smaller than 30% while typical roof pitches are 4/12 – 8/12 (larger than 30%). The minimum raster cell assignment in previous step may filter roof overhangs. Points located in area around the identified roof cells are checked, and if a roof overhang is detected, that area is added to the roof segment. The result is a binary raster classified into roof cells and none-roof cells as shown in the top row third column of Figure 2. Finally, the detected roof raster cells are converted to a polygon in the last step. Pre-event Roof Detection Model Pre-event roof detection model generates polygons representing building roofs before damage. The second row of Figure 2 illustrates the analysis procedure of the pre-event roof detection model. The inputs are an aerial image showing buildings before damage and the post-event roof polygons generated in the previous GIS model. The set of pixels in the aerial image that overlapped the post-event roof polygon is used as a sample to determine Red, Green, and Blue (RGB) and intensity values of roof pixels. The model employs the GIS buffering tool to generate a polygon that is a 1.5 meter band around the post-event roof polygon as shown in the second column of the second row in Figure 2. Pixels under the buffer are used as another set of training samples to establish a range of the RGB values for non-roof pixels. Since the remaining portions of damaged roofs are typically connected to an external wall, the new buffered polygon will overlap many non-roof pixels (ground surface) in the image. Using the RGB samples, the model identifies whether or not the image contains a roof pixel within a rectangular neighborhood as shown in the second row third column of Figure 2. Then, the RGB samples are updated based on the new detected roof pixels, the rectangular neighborhood is enlarged, and the image classification procedure iterates until the roof polygon overlaps the entire area of the roof in the image. The output is the pre-event roof polygon. Post-event Wall Detection Model Post-event wall detection model generates polylines representing the remaining external walls after damage. The third row of Figure 2 illustrates the analysis procedure of the post-event wall detection model. The inputs are the postevent collected point cloud data and the pre-event roof polygons generated in the


1212



previous GIS model. First, the point cloud data is converted to a 2D raster dataset in which each cell saves the difference between the maximum and minimum Z values and the number of points within that cell. The outline of the pre-event roof polygon generated in the previous GIS model is used to isolate the probable locations of wall raster cells. Threshold values, set based on the fact that values of the raster cells along remaining walls are considerably higher than at other locations, are used to identify wall cells in the isolated area. Finally detected wall raster cells are converted to polylines as shown in the last row last column of Figure 2.

Figure 2. Automated GIS damage assessment models including the post-event roof detection model (top row), the pre-event roof detection model (middle row) and the post-event wall detection model (bottom row) EXPERIMENTAL VALIDATION OF GIS MODELS The performance of the GIS models was evaluated through a set of controlled experiments. These experiments were designed and conducted to determine how input- and algorithm-related factors such as the density of points in the input point cloud, the raster cell size in the GIS models, and the extent of damage of scanned buildings influence on the accuracy of results. Synthetic point cloud datasets of a typical pitched roof building with controlled conditions such as point density and extent of damage were created and used in experiments. To prepare a point cloud dataset, first a solid 3D model of a building was created based on the footprint and dimensions extracted from the building aerial image. Then, the 3D model was divided into planar segments, and points were randomly added on the segment until the point density reached to desired values. Finally, the point cloud of the building was manually “damaged” by removing


1213



points on the building roof and walls. Figure 3 shows six sets of building point clouds prepared for experiments. Each point cloud represents a degree of damage (DoD) defined by the Enhanced Fujita scale (EF-scale) for one and two-family residences from a building with no damage shown in Figure 3(a) to total destruction shown in Figure 3(f).

Figure 3. Point cloud of model buildings with different degrees of damage Each of the prepared point cloud datasets shown in Figure 3 with various point densities were combined with the aerial image of the building and run through the roof and wall detection models. Also, various cell sizes were used in the GIS models. Figure 4 (a) - (f) show the results of the input point clouds shown in Figure 3 (a) – (f) with given point density of 25 points per meter and raster cell size of 40 cm. The pre-event roof polygons are shown in light gray, the post-event roof polygons are shown in dark gray, and post-event wall polylines are shown in black. The area and length of the resulting roof polygons and wall polylines in each run were compared with the actual area and length determined manually to determine the accuracy of results.

Figure 4. Generated roof polygons and wall polylines for the building with different degrees of damage


1214



The analysis indicated that for typical point cloud densities greater than 25 points per square meter, a cell size value of 40 – 50 cm results in an optimum performance of both the post-event roof detection model and the post-event wall detection model with errors lower than 10% and 5%, respectively. The analysis also indicated that the pre-event roof detection model results in roof polygons with errors lower than 5% for all buildings with different amounts of damage except for a building totally collapsed, which is shown in Figure 3 (f) and 4 (f). That extent of damage is when most of the roof and walls are collapses and only a few internal walls remain. The area of the post-event roof polygon in this was very small (0.1% of the original roof surface); therefore, the pre-event roof detection model only detected half of the pre-event roof in the aerial image as shown in Figure 4 (f). CASE STUDY: DAMAGED BUILDINGS AFTER THE 2013 MOORE TORNADO The developed GIS models were tested on an actual case study of a damaged residential area in the aftermath of the 2013 tornado that struck Moore, OK. The residential neighborhood was at Eastmoore Ct. located to the north of the centerline of the tornado path. Figure 5 (a) and (b) are the aerial photographs of the site taken before and after the tornado. The extent of damage increased from north to the south, which was closer to the center of the tornado path. Point cloud data was collected with a Leica c10 scanner. Ten full 360 degree scans were collected from different locations in order to cover the entire site. Scans were then registered manually. The process of data collection in the field and point cloud registration took approximately one day. To reduce the data collection and registration time and achieve near-real time assessment, mobile scanners mounted on vehicles can be used. Mobile scanners collect point cloud data while driving down streets; therefore, mobile scanners eliminate the need for setting up the scanner in different locations. Also, the point cloud data collected by mobile devices are rapidly registered based on the scanning trajectory and locations. The GIS models were run with the point cloud data along with the aerial image as inputs, and percents of roof and wall damage were calculated. Figure 5 (c) illustrates the polygons and polylines generated by the GIS models, while Table 1 presents the calculated percent of damage values. Building number 10, shown with hatch area in Figure 5 (c), had been cleaned up before the data collection. Therefore, the GIS models did not identify any roof or wall surfaces associated with building 10. A comparison between the model results in Figure 5 (c) and visual inspections by an assessment team indicates that the roofs and walls have been properly detected by the GIS models. Minor discrepancies were seen in a few areas of the generated postevent roof polygons and wall polylines. In a few buildings (e.g. building number 2), the roof sheathing is gone but the remaining roof rafters or debris were detected by the GIS tool as remaining intact roof surface. Also in a few areas (e.g. south wall of building number 9), gaps in the point cloud data due to the scanner occlusion were detected as wall damage. Overall, the error in results caused by these discrepancies is negligible. In addition, the model produced detailed percentage of damage for buildings, which was not provided by infield inceptions.


1215



1216

Figure 5. Case study of a damaged residential area after the 2013 Moore, OK tornado: (a) Pre-event aerial photo; (b) Post-event aerial photo; (c) generated roof polygons and wall polylines with the GIS models

Table 1. Percent of roof and wall damage for buildings of the case study PostPre-event PostPre-event Roof event Roof event Wall Building Damage Wall Area Roof Length (%) Length (m2) Area (m2) (m) (m) 1 246.27 117.6 52% 73.99 56.4 2 335.34 314.08 6% 93.87 90.8 3 299.55 178.08 41% 76.08 62.4 4 270.6 192.8 29% 74.31 54 5 302.62 179.68 41% 80.54 71.05 6 205.49 86.4 58% 65.82 33 7 281.63 127.36 55% 74.48 70.49 8 249.84 115.2 54% 73.45 68.4 9 285.51 116.48 59% 81.53 40.8


Wall Damage (%) 24% 3% 18% 27% 12% 50% 5% 7% 50%



CONCLUSION This study developed a novel automated tool for tornado damage assessment through combining terrestrial laser scanning technology, geometrical computing, and GIS modeling. Laser scanning and automatic analysis of collected data reduce the time and human resources needed for post-tornado investigations. The developed damage assessment tool rapidly produces quantitative damage information at individual building scale, which can be used as input to empirical or statistical loss estimation methods. The contributions of this study are summarized in the following points:  This study introduced a new GIS-enabled building detection method. Incorporation of post-event point cloud data and pre-event aerial images in the developed GIS models improved the building detection procedure and allows for the automatic calculation of the percent of roof and wall damage.  The result accuracy of the developed damage assessment tool was validated through controlled experiments. The analysis determined that for typical point cloud densities of greater than 25 points per square meter, a raster cell size within the range of 40 – 50 cm results in less than 10% error in post-event roof and wall detection. The analysis also indicated that the pre-event roof detection model results in roof polygons with errors lower than 5% for buildings with different degrees of damage except for a building totally collapsed.  The performance of the developed damage assessment tool was validated through a case study on actual damage data collected after the 2013 Moore, OK tornado. A comparison between the results and visual inspections indelicate that the roof and wall damage have been properly detected by the developed tool. Gaps in point cloud data, and remaining rafters and debris on roofs may cause small errors in damage calculation. LIST OF REFERENCES Changnon, S. A. (2009). Tornado Losses in the United States. Natural Hazards Review, 10(4), 145–150. Dash, J., Steinle, E., Singh, R. P., & Bähr, H. P. (2004). Automatic building extraction from laser scanning data: An input tool for disaster management. Advances in Space Research, 33(3), 317-322. doi:10.1016/S0273-1177(03)00482-4 Downton, M.W. & Pielke, R. A. (2005) How Accurate are Disaster Loss Data? The Case of U.S. Flood Damage. Natural Hazards 35:2, 211-228 Eguchi, R. T., Gill, S. P., Ghosh, S., Svekla, W., Adams, B. J., Evans, G., & Spence, R. (2010). The January 12, 2010 Haiti earthquake: A comprehensive damage assessment using very high resolution areal imagery. In 8th International Workshop on Remote Sensing for Disaster Management Erdik, M., Sesetyan, K., Demircioglu, M., Hancilar, U., and Zulfikar, C. (2011) Rapid earthquake loss assessment after damaging earthquakes, Soil Dynamics and Earthquake Engineering, V 31, I2, 247-266 Folger, P. (2011). Severe Thunderstorms and Tornadoes in the United States. DIANE Publishing. Irving, S. (2008). Disaster Cost Estimates: FEMA Can Improve its Learning from Past Experience and Management of Disaster-Related Resources. DIANE Publishing.


1217



Kashani, A. G., Biswas, S., Crawford, P., Graettinger, A., Grau, D. (2013). Automated Tornado Damage Assessment and Wind Speed Estimation Based on Terrestrial LiDAR Data. Journal of Computing in Civil Engineering, ASCE, (in review) Li, M., Cheng, L., & Gong, J.Y. (2008), Post-earthquake assessment of building damage degree using LiDAR data and imagery, Science in China Series E: Technological Sciences, 51, 133-143 Meade, C. and Abbott, M., (2003) Assessing Federal Research and Development for Hazard Loss Reduction, RAND, Santa Monica, CA. Okuyama, Y., & Chang, S. E. (Eds.). (2004). Modeling spatial and economic impacts of disasters. Springer. Olsen, M. J., Kuester, F., Chang, B. J., & Hutchinson, T. C. (2010). “Terrestrial laser scanning-based structural damage assessment.” Journal of Computing in Civil Engineering, 24(3), 264-272. Ozisik, D. and Kerle, N. (2004). Post - earthquake damage assessment using satellite and airborne data in the case of the 1999 Kocaeli earthquake, Turkey. In: ISPRS 2004: proceedings of the ISPRS congress: Geo-imagery bridging continents, 1223 July 2004, Istanbul, Turkey. Comm. VII, TS-PG: WG VII/5. pp. 686-691. Park, H. S., Lee, H. M., Adeli, H. & Lee, I. (2007). “A new approach for helth monitoring of structures: terrestrial laser scanning.” Computer-Aided Civil and Infrastructure Engineering, 22, 19-30. Prevatt, D. O., Roueche, D. B., van de Lindt, J. W., Pei, S., Dao, T., Coulbourne, W., Graettinger, A. J., Gupta, R. & Grau, D. (2012). “Building Damage Observations and EF Classifications from the Tuscaloosa, AL, and Joplin, MO, Tornadoes.” Structures Congress 2012, ASCE, Chicago, IL. Rehor, M., Bähr, H.-P., Tarsha-Kurdi, F., Landes, T., and Grussenmeyer, P. (2008). Contribution of Two Plane Detection Algorithms to Recognition of Intact and Damaged Buildings in LiDAR Data. The Photogrammetric Record, 23 (124), 441-456. Rose, A. (2004) Economic principles, issues, and research priorities in hazard loss estimation. Modeling Spatial and Economic Impacts of Disasters, 13-36. van de Lindt, J. W., & Dao, N. T. (2011). Loss analysis for wood frame buildings during hurricanes. II: Loss estimation. Journal of Performance of Constructed Facilities, 26(6), 739-747. Vickery, P. J., Lin, J., Skerlj, P. F., Twisdale Jr, L. A., & Huang, K. (2006b). HAZUSMH hurricane model methodology. I: hurricane hazard, terrain, and wind load modeling. Natural Hazards Review, 7(2), 82-93. Vickery, P. J., Skerlj, P. F., Lin, J., Twisdale Jr, L. A., Young, M. A., & Lavelle, F. M. (2006b). HAZUS-MH hurricane model methodology. II: Damage and loss estimation. Natural Hazards Review, 7(2), 94-103. Vosselman, G. (2000). Slope based filtering of laser altimetry data. International Archives of Photogrammetry and Remote Sensing, 33(B3/2; PART 3), 935-942. WISE Research center (2004) A recommendation for Enhanced Fujita Scale, Submitted to National Weather Service and other interested users, Texas Tech University Wind Science and Engineering Research Center, Lubbock Womble, J.A., B.J. Adams, S. Ghosh, & C.J. Friedland (2008). Remote sensing and field reconnaissance for rapid damage detection in Hurricane Katrina, Proceedings of the. ASCE/SEI Structures Congress, April, 2008, Vancouver.


1218

3D Data Collection and Automated Damage ...

3D Data Collection and Automated Damage ...

Suggest Documents

automated 3d data collection (a3ddc) for 3d ... - Semantic Scholar

automated crystal screening and remote data collection

SUPPLEMENTARY DATA Automated 3D structure

Data collection and 3D reconstruction - GDMC

Automated data collection in fragmentation warhead ...

Survey Data Collection Using Complex Automated Questionnaires

AUTOMATED COLLECTION OF PEDESTRIAN DATA USING ...

Automated data collection using bar code | SpringerLink

AUTOMATED DAMAGE DETECTION AND ... - CiteSeerX

Using Twitter for tasking remote-sensing data collection and damage ...

Automated Spyware Collection and Analysis

Automated collection and dissemination of ionospheric data from the ...

Automated collection of imaging and phenotypic data to ... - Frontiers

Automated Crystal Mounting Technical Advance and Data Collection ...

Truck cycle and delay automated data collection system in surface ...

Contract: Xerox for Automated Data Collection and ... - Berkeley

Contract: Xerox for Automated Data Collection and ... - Berkeley

Automated collection of imaging and phenotypic data ... - BioMedSearch

High Data Output and Automated 3D Correlative LightâElectron ...

Automated Trash Collection Brochure

a multisource geodatabase for collection of flood-related damage data

Optimization of data collection taking radiation damage into account

Optimization of data collection taking radiation damage into account

On-site data collection of damage caused by flash floods

3D Data Collection and Automated Damage ...