Hurricane Katrina Damage Assessment on Lands Managed by the Desoto National Forest using Multi-Temporal Landsat TM Imagery and High Resolution Aerial Photography Renee Jacokes-Mancini Forest Service Southern Region Jess Clark & Kevin Megown USFS Remote Sensing Applications Center (RSAC) RS2006 Conference Salt Lake City, Utah April 26, 2006
Overview Background of Hurricane Katrina damage Need for damage assessment Method use of assessment Results Discussion
Hurricane Katrina made landfall in Louisiana and Mississippi. Landfall: August 29, 2005 Strong Category 3 at landfall Hurricane Katrina Sustained winds of 111 to 130 mph Costliest Atlantic Hurricane in U.S. history
Water Damage June 2005
Water Damage September June 2005 2005
Wind Damage
Initial Forest Damage Assessments Forest Inventory and Analysis (FIA) September 2005 Forest Health Protection (FHP) November 2005 Initial assessments provided little explanation of damage severity or were based on few samples. Different method needed to meet National Forest management information requirements.
New Damage Assessment New damage assessment performed by RSAC Utilizes multi-stage sampling and multiple scales of remotely-sensed imagery Provides a statistically-defensible approach to mapping damage potential
Methods - Landsat Multi-temporal temporal Landsat imagery acquired Pre-storm: Oct. 15, 2004 Post-storm: storm: Nov. 3, 2005 WRS Path/Row: 21/39 Converted to Reflectance Other path/rows acquired but sampling performed only on 21/39
Methods - Landsat Multi-temporal temporal Landsat imagery acquired Pre-storm: Oct. 15, 2004 Post-storm: storm: Nov. 3, 2005 WRS Path/Row: 21/39 Converted to Reflectance Other path/rows acquired but sampling performed only on 21/39 Desoto NF overlaid on path 21 row 39
Methods Multi-stage Sample Built-up, up, Urban, and Ag areas masked out of Landsat imagery Those land use types can easily confuse the NDVI change We are interested in damage to forest Pre-storm Post-storm storm
Methods Multi-stage Sample Built-up, up, Urban, and Ag areas masked out of Landsat imagery Those land use types can easily confuse the NDVI change We are interested in damage to forest Urban/Ag Masked Out
Methods - Landsat Normalized Difference Vegetation Index (NDVI) performed on pre- and post- storm Landsat Change in NDVI calculated as a continuous dataset ISODATA Unsupervised Classification performed on NDVI change product Output: 12 class thematic image
Methods - Photos The Desoto NF sent RSAC over 400 scanned photos flown by FHTET during October, 2005 1:15,840 scale Flown with Inertial Measurement Unit (IMU) Photos scanned 1/2 m pixels Data compiled into Block Files for Leica Photogrammetry Suite (LPS) Photos orthorectified using 30m DEM
Methods - Sampling 12-class raster NDVI change product converted to point coverage Each point represented center of pixel 10 points randomly selected from each class 120 primary sampling units (PSUs( PSUs) Assessment done with this dataset to estimate variance and sample needs for whole study
Methods - Sampling Distribution of samples based on variance. Based off initial assessment, to reach a standard error of 2%, we needed to sample the following Class # of Samples 1 4 2 19 3 5 4 23 5 5 6 25 7 48 8 32 9 37 10 110 11 99 12 16 Total: 423 Primary Sampling Units (PSUs)
Methods - Sampling 423 PSU locations chosen randomly Note: Top and west edges excluded because path 21 row 39 orbit excluded it from the Landsat acquisition.
Methods - Sampling 423 PSU locations chosen randomly Note: Top and west edges excluded because path 21 row 39 orbit excluded it from the Landsat acquisition.
Methods - Sampling Digital Mylar Image Interpreter (ArcMap extension) used to perform damage assessment 3x3 grid (9 points) centered around PSU All 9 points fit within one Landsat pixel Assessed for damage no damage Damage = obvious thrown trees
Methods - Sampling Size of Landsat pixel over digital photo Nearly 4,000 points interpreted for damage no damage
Methods - Sampling Size of Landsat pixel over digital photo
Methods - Sampling 1 = Primary Sample Unit = 9 points = 3807 samples 1 Primary Sample Unit 9 points 3807 Samples of Damage/Non-Damage
Methods - Analysis Strata Acres Damage (%) Std. Error (%) Sample provided basis for % damage estimates with standard errors for each of the 12 NDVI change classes Class 1 3,353.27 Class 2 38,260.84 Class 3 68,966.91 Class 4 57,618.99 Class 5 45,993.95 Class 6 76,533.46 Class 7 56,562.17 Class 8 40,710.96 Class 9 49,566.29 0 1.23 7.29 18.44.43 7.76 34.1 6.35 19.36 0.847 3.18 3.09.427 1.83 5.05 3.39 5.04 Class 10 36,150.08 13.2 7.19 Class 11 37,081.25 42.67 7.79 Class 12 27,080.15 12.44 4.63 537,878.32 Overall 77,681 Acres Attributed as Damaged
Methods - Analysis Student- Newman-Keuls (SNK) Test used to create 4 separable damage classes Strata Class 11 Class 7 Class 9 Class 4 Class 10 Class 12 Damage (%) 42.67 34.10 19.36 18.44 13.20 12.44 Acres Damaged 15,821.46 19,282.61 9,593.56 10,624.94 4,769.64 3,369.85 Class 6 7.76 5,936.70 Class 3 7.29 5,029.07 Class 8 6.35 2,584.74 Class 2 1.24 472.52 Class 5 0.43 196.39 Class 1 0.00 0.00
Methods - Analysis Student- Newman-Keuls (SNK) Test used to create 4 separable damage classes Strata Class 11 Class 7 Class 9 Class 4 Class 10 Class 12 Class 6 Damage (%) 42.67 34.10 19.36 18.44 13.20 12.44 7.76 Acres Damaged 15,821.46 19,282.61 9,593.56 10,624.94 4,769.64 3,369.85 5,936.70 4 3 2 Class Acres Damaged Class 4 35,104.07 Class 3 Class 8 7.29 6.35 5,029.07 2,584.74 Class 3 20,220.80 Class 2 21,690.00 Class 1 668.92 Class 2 Class 5 Class 1 1.24 0.43 0.00 472.52 196.39 0.00 1
Class 1 Lowest probability of damage Class 2 Class 3 Class 4 Highest probability of damage Results
Class 1 Lowest probability of damage Class 2 Class 3 Class 4 Highest probability of damage Results
Results Based on the data, less damage than originally assumed was present on the Desoto Ranger District Geographic Locations of Damaged Areas Reliable and Statistical Defensible Method Method can be applied to other events/locations Results in Weeks
Expected Time-Line Needs Acquisition of aerial photography: Variable days to weeks Orthorectification and reprojection of photography: 2 days Acquisition and processing of Landsat imagery: 2 days Sampling photography for damage/no-damage: damage: 3 days Data analysis: 1 day Hurricane Photography acquisition begins Scan images, compile into block files Orthorectify photos Begin photo sampling Day 1 8 10 14 18 22 25 Data Analysis Landsat acquired NDVI change products created
Discussion Data-driven damage assessment Multi-stage sampling design an effective way to assess damage probability Quick and efficient way to create damage assessment Product can be used to help direct salvage and other management decisions Potential application for future events
Questions For Further information contact: Renee Jacokes-Mancini rjacokes@fs.fed.us (404) 347-2588 Jess Clark jtclark@fs.fed.us (801) 975-3769
Questions