Enhancing FDS for Modeling WUI Fires
Download
Report
Transcript Enhancing FDS for Modeling WUI Fires
Extracting Building Footprints
from LiDAR and Aerial Imagery in
the Wildland Urban Interface
(WUI)
Derek McNamara
GIS Analyst
Coeur D’Alene Tribe
Presentation Overview
Goals & Objectives
Available Data
Examined Methods
Accuracy Assessments
– Methods
– Results
Discussion (Limitations & Challenges)
Future Work
Background
Catastrophic Fires in WUI
– ~9,000 homes destroyed 1985-1994 (NFPA)
Gain Understanding WUI Fire Behavior
Few Physics-Based Models of WUI Fires
Cooperative Project
– CDA Tribe & National Institute for Standards & Technology
– Modification of Fire Dynamics Simulator (FDS) for WUI
CDA Tribe Provides Model Inputs
–
–
–
–
Structure Information
Tree Stem Locations (Crown Width, Height, Height to Live Crown)
Vegetation
Fire Barriers
Goals & Objectives
1) Extract Building Footprints
– Entire CDA Tribe Reservation
– Inputs to FDS for testing
2) Compare Methodologies
– Conclusions on Feasibility
– Identify/Develop Robust Methodology
Available Data
LiDAR Coverage (Entire Reservation)
Multispectral Imagery (NAIP)
Structure Locations
– Surveyed NFPA 1144 Assessment Form
Examined Methods
Building Extraction
– Many Methods Conducted in Urban Areas
– Fewer studies in Rural Areas.
4 Methods Examined in WUI
Initial LiDAR Processing
All Return
LiDAR
Point Data
Bare Earth
LiDAR
DEM
Extract
Values
To Points
Calculate
Surface
Height
Last
Return
Threshold
Surface
Height
Threshold
Derive
nDSM
Bare Earth DEM
– AML: Evans (2005, RMRS)
– Create TIN Ground Points
– TIN to Raster (1 meter)
Last Return Points
– Threshold (2 – 15 meters)
Points Outside Threshold
– Set to Zero
– Create TIN from Heights
– TIN to Raster (1 meter)
Texture Measure Extraction
Common Method in Literature
– Maas (1999); ICREST (2001); etc..
Local Variations of Height
Edge Detectors, Texture, Slope, Curvature,
Etc…
– Similar Results
Present Results Texture Variance
– Incorporate Curvature & Slope
Texture Variance Extraction
Workflow
nDSM
LiDAR
Texture
Variance
(3X3)
Texture
Threshold
(Binary)
Vectorization
Aggregate
Polygons
Vectorization
– Binary Image Raster to Polygon
– Delete Largest Polygon
Aggregate Polygons
Square
– Feature Analyst
Polygons
– Remove Isolated Polygons
Square Polygons
Area
Thickness
– Feature Analyst
Threshold
Threshold
– Douglas-Peucker
Thickness
– Zonal Geometry
Curvature
Slope
Threshold
Threshold Curvature & Slope
(Rottensteiner & Briese, 2002)
Final
Building
Polygons
– Apply Threshold (reclassify)
– Count Pixels in Polygon
– Delete polygons w/ > 50%
Height Threshold Workflow
Hewett (2005 ESRI UC)
nDSM
LiDAR
Aggregate
Polygons
Convert
To
Integer
Square
Polygons
Reclassify
To
Binary
Area
Threshold
Thickness
Threshold
Vectorization
Curvature
Threshold
Slope
Threshold
Final
Building
Polygons
Integer Conversion
– Removes Interpolation Errors
Object Oriented Image Classification
(Ibrahim, 2005)
Related Pixels part of objects.
Assigns Relationships Related Pixels
– Not a Pixel-by-Pixel Approach
– Appropriate for Urban Classification
(Wikipedia, 2005)
Implemented in Feature Analyst
Multispectral & LiDAR Height
LiDAR Intensity & Height
Multispectral & LiDAR Workflow
PCA 1
NAIP
Imagery
Curvature
nDSM
Apply
Height
Mask
Texture
Variance
nDSM
Training
Examples
Slope
nDSM
Set-Up
Learning
Image
Classification
First PCA
– Reduce Bands
Set-Up Learning
– Input Representation
– Spatial Context
Manhattan 7X7
Aggregate
Polygons
Square
Polygons
Thickness
Threshold
Final
Building
Polygons
Area
Threshold
Height Mask
– Improves Classification
LiDAR Intensity Workflow
LiDAR
nDSM
Intensity
Curvature
nDSM
Apply
Height
Mask
Texture
Variance
nDSM
Training
Examples
Slope
nDSM
Set-Up
Learning
Intensity Not PCA
Image
– Measure of Signal
Strength
– Often contains
noise
Classification
Aggregate
Polygons
Square
Polygons
Thickness
Threshold
Final
Building
Polygons
Area
Threshold
Accuracy Assessment Methodology
(Song & Haithcoat, 2005)
10 measures Described by Song & Haithcoat
(2005)
Use all except Shape Similarity
Most Measures Calculated on Correctly Extracted
Building Polygon
Average Across all Correctly Extracted Building
Polygon
Reference Data
– Manually Digitized
NAIP, LiDAR, Structure Photos.
Accuracy Assessment Methodology
Cont.
Detection Rate = Producer’s Accuracy
Correctness = User’s Accuracy
Matched Overlay (Correct Buildings) =
overlapping building area
referencebuilding area
totalnumber correctbuildings
Area Omission Error (Correct Buildings) =
nondetected building area
referencebuilding area
totalnumber correctbuildings
Accuracy Assessment Methodology
Cont.
Area Commission Error (Correct Buildings) =
incorrectly detectedbuilding area
referencebuilding area
totalnumber correctbuildings
RMSE (Correct Buildings) =
d
^
2
# cornerscorrectbuilding
totalnumber correctbuildings
Accuracy Assessment Methodology
Cont.
Corner Difference (Correct Buildings) =
detectedbuilding corners referencebuilding corners
totalnumber correctbuildings
Area Difference (Correct Buildings) =
detectedbuilding area - referencebuilding area
referencebuilding area
totalnumber of correctbuildings
Perimeter Difference (Correct Buildings) =
detectedbuilding perimeter- referencebuilding perimeter
reference
building
perimeter
totalnumber of correctbuildings
Accuracy Assessment
Can Not Discern Statistical Differences
Should Look at Methods Building Verse
Building
Identify Common Extracted Building Between
Methods
Examine statistical differences among methods.
Accuracy Assessment
Completeness Measures
MEASURE
Texture
Extraction
Height
Multispectral
Extracted Building (Multispectral)
Extraction
Extraction
LiDAR Intensity
Extraction
Reference Building
Detection Rate (%)
Average Matched
Overlay (%)
69.7
16.9
80.6
73.5
19.0
83.6
72.3
28.0
79.0
66.7
12.4
79.5
Average Area
Omission Error (%)
19.5
16.4
21.0
20.1
Average Area
Commission Error (%)
19.2
19.3
11.3
13.1
Correctness (%)
Accuracy Assessment Geometric Accuracy
MEASURE
Average RMSE (m)
Average Corner
Difference (#)
Texture
Extraction
2.02
1.4
Height
Extraction
1.90
1.59
Multispectral
Extraction
2.03
1.51
LiDAR Intensity
Extraction
2.40
2.01
Treed Area
Reference Buildings
Extracted Buildings (Intensity)
Last Return Intensity Brightness Values
0 - 74
74.00000001 - 123
123.0000001 - 160
160.0000001 - 198
198.0000001 - 255
Accuracy Assessment
Shape Similarity
MEASURE
Texture
Extraction
Height
Extraction
Multispectral
Extraction
LiDAR Intensity
Extraction
Average Corner
Difference (#)
1.4
1.59
1.51
2.01
Average Area
Difference (%)
19.7
22.0
19.4
20.1
Average Perimeter
Difference (%)
11.1
14.2
12.6
13.0
Texture Variance Measure
– Lowest Corner Difference
– Lowest Perimeter Difference
Best Shape Representation (?)
Accuracy Assessment
MEASURE
Correctness (%)
Texture
Extraction
16.9
Height
Extraction
19.0
Multispectral
Extraction
28.0
LiDAR Intensity
Extraction
12.4
Slope & Curvature Filters
– Show Promise
Remove ~ 40% of Incorrect Polygons
– After Area & Thickness Thresholds
– Not Removing Buildings Under Trees
– Coarse Approach
More Advanced Approach Better (?)
Objected-Oriented Approach (Slope & Curve Filters)
Remove ~ 21% of Incorrect Polygons (Intensity)
Remove ~20% of Incorrect Polygons (Multispectral)
Limitations of All Methods
Lose of Data (Interpolate Point Cloud to Raster)
Only Rectangular Buildings
Smaller Structures Not Discernable
Height Thresholds Vary Over Different Areas
Poor Job of Removing Trees
Last Return DOES NOT Detect Building Edge
Last Return Necessary
(Buildings Surrounded by Trees)
First
Last Return
Return
Trees Filtered
Trees Present
Building Extraction Results
Entire Reservation
Utilizing All Methods
Database Of Over 11,000 Footprints
Structure Point Locations (NFPA Surveys)
– Remove Noise
Required Manual Clean-up
– Smaller Structures
– Densely Canopied Areas
– Modification of Extracted Footprints
Large Area Extraction
Challenges
Software Limitations
– Large Dataset
– Vectorization Routines
Registration of NAIP to LiDAR
– Difficult in Forested Areas
Varying Height Thresholds
Different Techniques (First Return Vs Last Return)
– Opened Versus Treed Areas
Trees!!!
Conclusions
LiDAR Feasible
Did Not Quantify Difference Between Methods
Multispectral Removes Most Noise
– Does Not Discern Buildings Under Trees
LiDAR Intensity Too Much Noise (?)
Height Easy/Good Results
Texture Best Shape Similarity (?)
Slope & Curvature Filters Show Promise
Object-Oriented Approach
– No Filter Thresholds
All Methods Useful in Open Areas
– Easy to Apply
Sensor Limitations
– Small Structures Not Discernable
Feature Analyst
– Easy to Use
– Hierarchical Learning Not Examined
Future Work
General
Better Software
– Handle Large Point Cloud
Development of LiDAR Standards
– WUI Work
Incorporates Many Extraction Scenarios
– Good Test Bed
Future Work
CDA Tribe
Segment Man-made Objects From Vegetation!!!
– Slope & Curvature (Point Data)
Building Extraction Point Data
– Plane Fitting
Other Building Information
– Roof Type
– Height
Accuracy Assessments
– Incorporate Shape Similarity
– Statistically Test Differences Between Methods
Acknowledgements
NIST: Ruddy Mell, et al. (Funding)
Jeremy Adams, Noel Sanyal (Building
Clean-up)
Berne Jackson (Systems Manager)
Frank Roberts (GIS Manager)
James Twoteeth, Heather Fuller (CDA GIS)
Josh Arnold (NFPA Surveys)
USGS (Feature Analyst)
Questions?
Light Detection and Ranging
(LiDAR)1
Flight Height
~1829 meters
Coverage Area
129,500 hectares
Field of View
25 degrees
Vertical Accuracy
15 centimeters
Horizontal Accuracy
10 centimeters
Returns Per Pulse
5
Line Spacing
1,862 Feet
Maximum Along Track Spacing
1.8 meters
Maximum Cross Track Spacing
2.6 meters
Nominal Post Spacing
2.0 meters
~Number of Elevation Points
347,000,000
Number of Basestation Locations
1
1Flown
by Horizons Inc., South Dakota
Accuracy Assessment
Cont.
Height Misses
Duplicate Structure
Outliers
Small Structure
References
United States Department of Agriculture. Forest Service. Rocky Mountain
Research Station. Fuels Planning: Science Synthesis and Integration.
Environmental Consequences Fact Sheet: 3. Structure Fires in the Wildland-Urban
Interface. 2004-09. Research Note RMRS-RN-23-3-WWW.
Evans, J.S., and A.T. Hudak. A Progressive Curvature Filter for Identifying Ground
Returns from Discrete Return LiDAR in Forested Environments. (Submitted IEEE
Transactions on Geoscience and Remote Sensing).
Haithcoat, T., and W. Song, J. Hipple. Automated Building Extraction and
Reconstruction from LIDAR Data. R&D Program for NASA/ICREST Studies
Project Report. September, 2001.
Hewett, M. Automating Feature Extraction with the ArcGIS Spatial Analyst
Extension. 2005 ESRI International User Conference Proceedings.
Mass, G-H. The Potential of Height Texture Measures for the Segmentation of
Airborne Laserscanner Data. Presented at the 4th Airborne Remote Sensing
Conference and Exhibition, Ottawa, Ontario, Canada, 21-24 June 1999.
Rottensteiner, F. and C. Briese. A New Method for Building Extraction in Urban
Areas from High-Resolution LIDAR Data. IAPRSIS, Vol. XXXIV/3A, Graz, Austria,
pp. 310-317
Song, W., and T.L. Haithcoat. Development of Comprehensive Accuracy
Assessment Indexes for Building Footprint Extraction. Geoscience and Remote
Sensing, IEEE Transactions on. 43:2. February, 2005.