Identification of Tree Locations in Geographic Images
A thesis submitted in partial fulfillment of therequirements for the degree of Master of Science
in Computer Engineering
By
David Brown
Dr. Frederick C. Harris, Jr., Thesis Advisor
December, 2008
Committee
Dr. Frederick C. Harris, Jr.
Dr. Sergiu M. Dascalu
Dr. Timothy J. Brown
Overview
Purpose Background Methods Software Specification and Design Implementation and Results Conclusion and Future Work
Purpose
Purpose
To create an item placement utility for VFIRE (Virtual
Fire In Realistic Environments)
VFIRE is a virtual reality application for visualizing
wildfire simulations.
Current area of interest is Kyle Canyon in Southern
Nevada.
Purpose
Main uses of VFIRE:
Fire Training
Fire Planning
Fire Model Verification
Wildfire Visualization [25]
Purpose
The placement of items in the visualization should
correspond to their locations in the real environment.
This utility is intended to place large numbers of trees
with reasonable speed and accuracy.
It can also be used to place a small number of houses with
reasonable speed and accuracy.
Background
Background – Geographic Images
Photography
Standard Color
Panchromatic
Multispectral
Hyperspectral
Color Infrared (CIR)
Background – Photography
CIR images look different from most other types
True Color Image [5, page 45] False Color Image [5, page 45]
Background – Vegetation Maps
Vegetation data can be displayed as a map.
Vegetation maps have been created by LANDFIRE to
show various attributes.
Map of Vegetation TypeMap of Vegetation Cover
Background – Point Operations
Each output pixel is based on a single input pixel.
Changing the brightness of an image is a point operation.
Image After Increasing BrightnessOriginal Image
Background – Neighborhood Operations
Blurring an image is a neighborhood operation.
The blur filter is applied to each input neighborhood.
Original Image Blurred Image
Background – Edge Detection
The LoG (Laplacian of Gaussian) Filter is an edge
detection filter in which the level of detail can be
controlled.
Laplacian of Gaussian (LoG) filter [36]
Background – Template Matching
Can be used as the first step in image analysis
Used when finding the location of known item
Neighborhood operation where the filter mask is a
template of the desired item
Filtering produces a correlation image that can be scanned
for bright spots.
Background – Virtual Reality [2]
Requirements:
Virtual World
Immersion
Sensory Feedback
Interactivity
Background – HMDs [2]
Head Mounted Displays
100% Field of Regard
May cause dizziness
Only one person can view
at one time
A Head Mounted Display [2, page 14]
Background – Multi-Sided Projection Displays [2]
Field of regard depends on the
number of sides.
Wider field of view than HMDs
No dizziness
Many people can view at once
More bulky and expensive than
HMDsThree-Sided Projection Display [25]
Background – Head Tracking
• View must be adjusted for
location and orientation of
head.
• Stereoscopic display can be
used to create depth
perception. Head-Tracking Active Stereo Goggles [17]
Background – Input Devices
• A wand is a commonly used
input device for virtual
reality systems.
Virtual Reality Wand [17]
Background – Related Work
Applications
Plantation Management
Assessing Forest Health
Harvestable Lumber Estimation
Fuel Load Estimation
Background – Related Work
Culvenor [6]
Even-Aged Mountain Ash
(Eucalyptus)
NIR Selected from CIR
Identify Local Maxima
Identify Local Minima
Cluster Intermediate Pixels
Background – Related Work
Pouliot et al. [33]
Uniformly Spaced Spruce
Trees in an Arboretum
Absolute Difference of NIR
and Red Bands
Moving Window, Local
Maximum Filtering
Background – Related Work
Brandtberg and Walter [11]
80-Year-Old Stands of Scot pine,
Norway spruce, birch, and aspen
Perform Scale-Space Edge
Detection to Extract Tree Crown
Perimeters
Analyze Perimeter Curvatures to
Estimate Centroids
10-cm, CIR Brightness Scale Space Image
Estimating CentersAfter Edge Detection
Background – Related Work
Larsen [23]
Image of Norway spruce
Template Created from Ray-Traced
3D Tree Model
Model Incorporates Aircraft Position,
sun position, and Spiecies-Specific
Light-Scatering Parameters
Norway Spruce
3D Template Model
Background – Related Work
Image Analysis Software [4]
Will perform template matching
Background – Related Work
Image Analysis Software
Not likely to output locations in geospatial coordinates
Not likely to provide geospatially aligned overlays of
vegetation maps
Not likely to display vegetation map data for selected
locations
Not likely to make placements based on vegetation maps
Methods
Methods
Goals:
Achieve adequate tree-placement accuracy using whatever
images (if any) are available.
Enhance Accuracy using vegetation maps.
Make tree placements using vegetation maps alone if no
photographic image is available.
Methods
System:
Interactive (not fully automatic)
Template Matching (no image constraints)
Templates Created at Runtime (quickly create multiple
templates)
Vegetation maps provide information about terrain.
Placements can be made based on vegetation maps alone.
Methods
System:
The algorithm is not tailored to any particular image.
The user-defined templates are tailored to the image.
The algorithm is tailored to the correlation image
produced using the templates.
Methods
Data for Kyle Canyon
4-Meter Photographic GeoTIFFs Red, Green, Blue, NIR
1-Meter Photographic GeoTIFF Panchromatic
5-Meter Vegetation Maps Sampled from 30-Meters Vegetation cover, vegetation type, vegetation Height
Methods
1-Meter Panchromatic Image
Trees look like blobs.
Species, size, shadow, and density are different in
different parts of the image.
Software Specification and Design
Software Specification and Design
Use Cases
Software Specification and Design
System consists of five
groups of global functions
using two existing
libraries.
Implementation and Results
Implementation and Results
Detection Process
• The user selects a tree to
use as a template.
• The tree is the gray blob.
• The shadow is the dark,
elongated region. Cite of First Template (Zoomed In)
Implementation and Results
Detection Process
• The user draws a
highlighting mark over the
tree and shadow.
Template Defined by User
Implementation and Results
Detection Process
• Area Near Template,
Zoomed Out
Area Near Template (Zoomed Out)
Implementation and Results
Detection Process
• Correlation Image Stored
in Red Buffer of
Workspace Image
• Other buffers are used for
intermediate processing. Correlation Image
Implementation and Results
Detection Results
• Detection Results Using a
Single Template
Result Using One Template
Implementation and Results
Detection Process
• User controls tuning
parameters for tree
detection .
Tuning Parameter Window
Implementation and Results
• Detection Process
• User specifies data for trees
associated with each
template.
• Locations, types, etc are
then written to file.
Preparation to Create Output File
Implementation and Results
Detection Process
• Entire Image of Kyle
Canyon (8km × 6km)
Entire Kyle Canyon Image
Implementation and Results
Detection Process
• Vegetation Map of Same
Area (8km × 6km)
Entire Vegetation Map
Implementation and Results
Detection Process
• Vegetation Map As
Overlay onto Image
(8km × 6km)
Overlay of Vegetation Map onto Photographic Image
Implementation and Results
Detection Process
• Text Output Resulting
When User Clicks on
Image
Text Output from Clicking on Image
Implementation and Results
Partially Random Placement
• Tree Placements Made
According to Map of
Vegetation Coverage,
Without Using Image
Partially Random Placements
Implementation and Results
Partially Random Placement
• User can control how
placements are made when
no image is available
Options for Partially Random Placements
Implementation and Results
• Detection Accuracy
• Approach is to balance false positive and false negative
errors.
• This can be done by using multiple templates.
• Trees used as templates may differ only in the shadows
they cast.
Implementation and Results
Detection Process
• The user selects a second
tree to use as a template.
• It is similar to the tree in
the first template, but the
shadow is different. Addition of Second Template
Implementation and Results
Detection Process
• Adding a second template
reduces the number of
false negative errors and
slightly increases the
number of false positive
errors.Addition of Second Template
Implementation and Results
Detection Results
• Approximately 15% Poor
• Approximately 70%
Adequate
• Approximately 15% Good
Addition of Second Template
Implementation and Results
Detection Results
• Area near templates is an
example of “adequate”
detection accuracy.
Addition of Second Template
Implementation and Results
Detection Results
• Burnt trees may be
mistaken for living trees.
• Small Number of False
Positive Errors
Burnt Trees Incorrectly Marked
Implementation and Results
Detection Results
• Small trees or shrubs
mistaken for the medium-
sized trees being sought
• About ten percent of the
image has similar terrain. Apparent Shrubs Incorrectly Marked
Implementation and Results
Detection Results
• Trees incorrectly placed on
top of houses
• These false positive errors
are very noticeable.
Rooftops Incorrectly Marked
Implementation and Results
Detection Results
• Trees incorrectly placed on
roads
• These false positive errors
are very noticeable.
Roads Incorrectly Marked
Implementation and Results
Detection Results
• Area with few errors
• Good separation between
trees makes them easier to
detect.
Good Result
Implementation and Results
Detection Results
• Areas hidden by shadow
• Accuracy is unknown
Unknown Accuracy in Hidden Areas
Conclusions and Future Work
Conclusions
• The purpose of this utility is to achieve adequate tree
detection results using any available image.
• This goal cannot always be accomplished.
• Using the 1-meter panchromatic image of Kyle Canyon
produced at least “adequate” results in approximately
85% of the area of interest.
Conclusions and Future Work
Conclusions
• The process took a few hours, including the time required
to manually delete false positive errors located in
conspicuous areas.
Conclusions and Future Work
Future Work
• Ability to delete large numbers of trees at once
• Ability to change tuning parameters for the same
template in different regions
• Minimum Item Radius tuning parameter appears to have
little effect and may not be necessary.
Conclusions and Future Work
Future Work
• Ability to transfer templates from one group to another
• Ability to change the highlighting colors
• Ability to scroll image without momentary pauses and
without greatly increasing memory consumption
• Ability to detect artificial structures
Questions?
Top Related