Development of a Statistical Relationship Between GroundBased and Remotely-Sensed Damage in Windstorms Tanya M.
Download ReportTranscript Development of a Statistical Relationship Between GroundBased and Remotely-Sensed Damage in Windstorms Tanya M.
Development of a Statistical Relationship Between GroundBased and Remotely-Sensed Damage in Windstorms Tanya M. Brown Ph.D. Candidate Wind Science & Engineering Research Center Daan Liang J. Arn Womble February 17, 2010 Outline Need for Research Objective Hypothesis Damage Scales Damage Assessment Methodology “Super Tuesday” Tornado Ground-Based Datasets “Super Tuesday” Tornado Remote-Sensing Datasets Modifications to Dataset Developing Regression Models Validating Regression Models • Hurricane Katrina Remote-Sensing Datasets • Hurricane Katrina Ground-Based Datasets Future Work Acknowledgements Need for Research Tropical cyclones and tornadoes are responsible for 45.6% and 26.5% of catastrophic insurance losses 1988-2007, respectively. New technology • VIEWS • Satellite Imagery – Quickbird 61 cm panchromatic – WorldView 1 50 cm panchromatic • Aerial Photographs – NOAA hurricane surveys ~ 37 cm resolution – Pictometry 15 cm resolution Damage detection is not yet automated Ground surveys are costly and time-consuming Objective Provide a tool for the prediction of windstorm damage states at ground-level using remotesensing imagery, for site built one- or two-family residences (FR12) Hypothesis Windstorm damage evaluated from ground observations is statistically correlated to windstorm damage evaluated from remote-sensing imagery Damage Scales Table 1: Degree of Damage (DOD) states and wind speed parameters for FR12 structures from the Enhanced Fujita (EF) Scale Table 2: Womble’s Remote Sensing (RS) Damage Scale for Residential Construction Damage Assessment Methodology: “Super Tuesday” Tornadoes Select areas of both ground-based and remote- sensing damage data coverage Assign Degree of Damage (DOD) from EF Scale to each structure in ground-based dataset Assign Remote-Sensing Damage Scale (RS) from Womble (2005) to each structure Assign percentage of damage category for each RS Scale rating •0% •51-75% •1-25% •>75% •26-50% Ground-Based Datasets: “Super Tuesday” Tornadoes Tornado outbreak February 5, 2008 in Arkansas, Mississippi, Alabama, Tennessee, and Kentucky VIEWS deployed to capture high-definition ground- based photographs 32 hours of video captured; 8 hours processed to extract still, georeferenced images Each DI, DOD, and corresponding EF Scale value were evaluated for structures in photos • DIs other than FR12 are eliminated for this research Sample size: 271 FR12 structures Ground-Based Datasets: “Super Tuesday” Tornadoes <EF-0 EF-0 EF-2 EF-3 EF-1Damage Damage Damage Damage DOD<1 DOD DOD DOD6 284 Ground-Based Datasets: “Super Tuesday” Tornadoes Percentage of Ground-Based Damage for Madison County DOD 5 0% DOD 7 2% DOD 8 0% DOD 6 4% DOD 4 15% DOD 3 6% DOD <1 53% DOD 2 15% DOD 1 5% Remote-Sensing Datasets: “Super Tuesday” Tornadoes 26 square kilometers of satellite imagery in Madison County, TN February 8th and March 2nd, 2008 from QuickBird • Panchromatic 61 cm • Multispectral 2.44 m February 10th, 2008 from WorldView 1 • Panchromatic 50 cm Remote-Sensing Datasets: “Super Tuesday” Tornadoes QuickBird WorldView pan-sharpened QuickBird 1 imagery 61 cm resolution 50 Remote-Sensing Datasets: “Super Tuesday” Tornadoes RS Scale Evaluated from February 8th QuickBird Imagery 0% 0% 1% 0% 1% 0% 3% 3% 1% 3% RS Scale Evaluated from February 10th WorldView 1 Imagery 1% 0% 1% 0% 0% 1% A 0% 1% 2% 4% 1% A 0% B 1-25% B 26-50% 9% 39% B 1-25% 7% B 26-50% B 51-75% B 51-75% B >75% B >75% C 1- 25% C 1- 25% C 26-50% C 26-50% 53% C 51-75% C >75% C >75% D 1-25% D 1-25% D 26-50% D 51-75% D >75% 40% C 51-75% 29% D 26-50% D 51-75% D >75% Remote-Sensing Datasets: “Super Tuesday” Tornadoes Modification to Dataset: “Rerated” DODs Lots of structures rated DOD 4, with very slight damage seen in remote-sensing survey DOD 2 –Loss of roof covering material <20% DOD 4 –Loss of roof covering material >20% AND uplift of roof deck What about houses with more than 20% roof covering loss, but NO uplift of roof decking? What do we rate these? DOD 3 pertains to glass breakage, so its inappropriate Settled on DOD 2.5, to indicate worse than DOD 2 damage, but definitely less than DOD 4 damage—all homes originally rated DOD 4 were investigated to see if they should actually be DOD 4, or the new category of DOD 2.5, some ended up being DOD 3 because of glass breakage Suggests that the DOD categories in the EF Scale for FR12 structures may need to be adjusted in future revisions Modifications to Dataset Percentage of Ground-Based Damage for Madison County DOD 5 0% DOD 4 2% DOD 7 2% DOD 8 0% DOD 6 4% DOD 3 12% DOD 2.5 7% DOD <1 53% DOD 2 15% DOD 1 5% Developing Regression Models Table 3: Numerical representation scheme for the alphabetic and alphanumeric RS Scales used in the regression. Developing Regression Models 4 satellite datasets • • • • QuickBird WorldView 1 combined averaged 2 RS Scales • • alphabetic alphanumeric 3 regression types • • • linear exponential quadratic Original & “Rerated” DODs 4 satellite datasets • • • • QuickBird WorldView 1 combined averaged 1 RS Scale • alphanumeric 3 regression types • • • linear exponential quadratic “Rerated DOD” only 4 statistical transformations • • • • logarithmic exponential square root squared 3 variables transformed • • • x-variable (RS) alone y-variable (DOD) alone both x- and y-variables =192 models!!! Developing Regression Models re-rated DOD vs. alphanumeric RS Scale 10 8 8 6 6 4 QuickBird alphanumeric comparison 2 re-rated DOD re-rated DOD re-rated DOD vs. alphanumeric RS Scale 10 4 0 -1 0 0 1 2 3 4 5 -1 -2 0 2 3 4 5 -2 RS Scale rating re-rated DOD vs. alphanumeric RS Scale DOD vs. alphanumeric RS Scale 10 10 8 8 6 6 4 combined satellites alphanumeric comparison 2 4 averaged satellites alphanumeric comparison 2 0 -1 1 RS Scale rating DOD ratings re-rated DOD WorldView 1 alphanumeric comparison 2 0 0 1 2 -2 3 4 5 -1 0 1 2 -2 RS Scale rating RS Scale ratings 3 4 5 Developing Regression Models Alphanumeric RS Scale always outperforms alphabetic RS Scale “Rerated” DOD dataset always outperforms original WorldView 1 imagery generally outperforms QuickBird imagery Average satellite dataset generally outperforms QuickBird, WorldView 1 & combined dataset Quadratic regression models generally outperform linear & exponential models Statistical transformations generally increased model performance slightly DOD 2.256* RS 1.143* RS 0.3565 Developing Regression Models The five best models were selected for validation • • • All used the alphanumeric RS Scale All used the averaged satellite damage states All used the “rerated” DOD Eq. 4-5—Quadratic model with square root transformation of ‘RS’ (R2=0.6863) • DOD=2.256*RS-1.143*√RS+0.3565 Eq. 4-6—Quadratic model with natural logarithm transformation o f ‘RS+1’ (R2=0.6859) • DOD=2.737*(ln(RS+1))2-0.2846*ln(RS+1)+0.3465 Eq. 4-7—Quadratic model (R2=0.6795) • DOD=0.1656*RS2+1.070*RS+0.2654 Eq. 4-8—Exponential model, natural logarithm transformation of ‘RS+1’ (R2=0.6767) • DOD=0.4224*exp(1.772*ln(RS+1)) Eq. 4-9—Quadratic model, squared transformation of ‘RS’ (R2=0.6740) • DOD=-0.03206*RS4+0.8830*RS2+0.4650 Validating Regression Models Test the performance of the models against a Hurricane Katrina dataset • Evaluate the wind damage ONLY • Do not evaluate the surge damage Methodology: • Assign alphanumeric RS Scale rating to each FR12 structure • Predict the ground-level damage state using the developed models • Assign DOD from EF Scale to each structure in ground-based dataset • Compare the predicted ground-level damage state to the actual ground-level damage state obtained by the ground-based survey Remote-Sensing Datasets: Hurricane Katrina Aerial imagery obtained in Harrison, Hancock, and Jackson Counties in Mississippi September 6th-11th, 13th, and October 9th, 2005 Pictometry images • 15 cm resolution • Sample size: 517 FR12 structures August 30th-31st, and September 2nd, and 4th, 2005 NOAA images • 37 cm resolution • Sample size: 1008 FR12 structures Averaged aerial imagery dataset • Sample size: 505 FR12 structures Remote-Sensing Datasets: Hurricane Katrina NOAA 37 cm resolution Pictometry 6 in. resolution Remote-Sensing Datasets: Hurricane Katrina RS-scale Evaluated from NOAA Imagery RS-scale Evaluated from Pictometry Imagery 0% 0% 2% 0% 0% 0% 0% 1% A 0% 3% 8% 0% 21% 0% 0% B 1-25% B 26-50% 0% 2% 0% A 0% 0% B 1-25% B 26-50% 2% B 51-75% 10% 25% B 51-75% 4% 0% B >75% B >75% C 1-25% C 1-25% C 26-50% C 26-50% 19% C 51-75% C 51-75% C >75% C >75% D 1-25% D 1-25% D 26-50% D 26-50% D 51-75% D 51-75% D > 75% 43% D > 75% 60% Ground-Based Datasets: Hurricane Katrina Made landfall at Buras, Louisiana at a Category 3 hurricane on August 29th, 2005 • Wind damage • Surge damage VIEWS deployed to capture high-definition ground- based photographs Video captured in Hancock, Harrison, and Jackson Counties in Mississippi; processed to extract still, georeferenced images Ground-Based Datasets: Hurricane Katrina EF-1 EF-2 Damage Damage DOD DOD2.5 6 Validating Regression Models Predicted Ground-Based Damage States Determined from Averaged Satellite Imagery and Eq. 4-5 DOD 5 1% DOD 6 0% DOD 7 0% DOD 5 1% DOD 8 0% DOD 4 3% DOD 2.5 6% Predicted Ground-Based Damage States Determined from Averaged Satellite Imagery and Eq. 4-6 DOD 6 0% DOD 4 4% DOD <1 11% Predicted Ground-Based Damage States Determined from Averaged Satellite Imagery and Eq. 4-7 DOD 7 0% DOD 8 0% DOD 5 0% DOD 4 3% DOD <1 11% DOD 3 DOD 2.5 4% 6% DOD 3 7% DOD 1 23% DOD 2.5 DOD 3 7% 4% DOD 1 23% Predicted Ground-Based Damage States Determined from Averaged Satellite Imagery and Eq. 4-8 DOD 5 0% DOD 6 0% DOD 4 2% DOD 2.5 8% DOD 7 0% DOD 8 0% DOD 2 52% Predicted Ground-Based Damage States Determined from Averaged Satellite Imagery and Eq. 4-9 DOD 5 2% DOD 3 4% DOD 2.5 6% DOD 7 0% DOD 8 0% DOD <1 11% DOD 3 4% DOD 1 23% DOD 2 52% DOD 6 1% DOD 4 4% DOD <1 11% DOD 7 0% DOD 8 0% DOD <1 11% DOD 1 23% DOD 2 51% DOD 2 49% DOD 6 0% DOD 1 24% DOD 2 48% Validating Regression Models Actual Ground-Based Damage States for the Homes in the Averaged Aerial Dataset DOD 5 0% DOD 6 2% DOD 7 0% DOD 8 0% DOD 4 5% DOD 3 4% DOD <1 33% DOD 2.5 17% DOD 2 24% DOD 1 15% Validating Regression Models Comparison of Actual Ground-Level Damage States vs. Predictions from Eq. 4-6 Using Averaged Aerials Comparison of Actual Ground-Level Damage States vs. Predictions from Eq. 4-7 Using Averaged Aerials 10 10 8 8 8 6 6 6 4 4 4 2 2 Eq. 4-5 predictions 1:1 ratio 0 -2 Predicted DOD 10 Predicted DOD Predicted DOD Comparison of Actual Ground-Level Damage States vs. Predictions from Eq. 4-5 Using Averaged Aerials 0 2 4 6 8 2 Eq. 4-6 predictions 1:1 ratio 0 Linear10(1:1 ratio) -2 0 0 2 4 6 8 Linear10(1:1 ratio) -2 -2 -2 Eq. 4-7 predictions 1:1 ratio 0 2 -2 Actual DOD Actual DOD Actual DOD Comparison of Actual Ground-Level Damage States vs. Predictions from Eq. 4-9 Using Averaged Aerials 10 10 8 8 6 6 Predicted DOD Predicted DOD Comparison of Actual Ground-Level Damage States vs. Predictions from Eq. 4-8 Using Averaged Aerials 4 4 2 2 Eq. 4-8 predictions 1:1 ratio 0 -2 0 2 4 6 8 Linear10(1:1 ratio) Eq. 4-9 predictions 1:1 ratio 0 -2 0 2 4 -2 -2 Actual DOD 4 Actual DOD 6 8 Linear10(1:1 ratio) 6 8 Linear10(1:1 ratio) Validating Regression Models Frequency analysis: Given an actual DOD rating, with what frequency does the model predict it exactly? To within one DOD category? To within two DOD categories? Models can predict DOD<1 through DOD 3 to within two DOD categories for at least 86% of the observations Results are not as good for DOD 4 and higher, but data are more limited at these levels Future Work Inclusion of datasets with higher levels of damage • Lots of DOD<1 through DOD 4 damage • Less than 7% of structures from “Super Tuesday” were rated > DOD 4 • Less than 2% of structures from Hurricane Katrina were rated > DOD 4 Include additional variables • • • • • Tornadoes—proximity to track Hurricanes—H*Wind wind speed Presence of debris Condition of surrounding structures Resolution of remote-sensing imagery Investigate economic recovery, by utilizing surveys from different dates Expand the study to include additional DI’s (mobile homes, apartments, shopping malls, schools, etc.) Acknowledgements ImageCat Inc. Anneley McMillan Paul Amyx Ron Eguchi & Beverley Adams Texas Tech University Amber Reynolds Rich Krupar NSS & Associates NSF (IGERT DGE-0221688) MCEER Contact Information: [email protected]