An Automated Way to Extract Wing Beat Frequency and Flap

Behavior Signatures of Birds
An Automated Way to Extract
Wing Beat Frequency and FlapGlide Patterns from Thermal
Imagery
Val Cullinan, Ph.D.
Corey Duberstein, M.Sc.
Shari Matzner, Ph.D.
1
Acknowledgements
Signature Discovery Initiative: PNNL National Security
Directorate
Wind and Water Power Program: DOE Office of
Energy Efficiency & Renewable Energy
2
Presentation Overview
Problem Statement
Wing Beat as a Signature
Data Extraction
Wing Beat Position
Conclusions
3
Problem Statement
How do we assess risk to “flying animals” offshore?
Assess risk
BACI, BA, Impact-Reference Design, Response-Gradient
Design, Resource Selection Function, etc…
Presence, abundance
Behavior: flight height, avoidance
“Flying Animals”
Identification
Bird vs. bat
Common vs. rare spp
Endangered vs. least concern
Offshore
Remote
Inhospitable
Dynamic
4
Wing Beat as a Signature
Pennycuick 1996, 2001 allometric model
F = mass3/8acceleration1/2wingspan-23/24 wing area-1/3 air density-3/8
Bruderer et al. 2010
Measured 155 species, compiled 45 species (Europe)
4 flight types:
Continuous flapping: wading birds, waterfowl, auks, gulls,
terns
Soaring: storks, pelicans, lg raptors
Dynamic soaring: albatrosses, shearwaters
Flap-glide: passerines, gulls, terns
Conclusion: Pennycuick pretty reliable for continuous
flapping flight
5
Data Extraction
Wings
Up
6
Wings
Down
Wings
Up
Wings
Glide
Data Extraction
Pixel intensity values output
Centroid
Centroid of pixel mass output
Calculate “hot
Hot spot”
Spot in pixelated data
7
Wing Position
UP: w = 41, h = 15
Centroid
Hot Spot
NEUTRAL: w = 40, h = 9
DOWN: w = 38, h = 14
8
Wing Position
Discriminant Analysis
Hot spot relative to centroid
Frame height
Hot spot relative to height
and width
1st Root 74% of VAR
91% Correct with crossvalidation
Determine up/down cycle
and wing beat frequency
Root 1 vs. Root 2
5
4
Neutral
3
2
Down
Root 2
1
0
-1
-2
Up
-3
-4
-5
-6
-5
-4
-3
-2
-1
Root 1
9
0
1
2
3
4
Up
Glide
Down
Wing Beat Frequency
(Hz)
Automated vs. Manual
6
5
Gull
Est. WBF (Hz)
Mean (st. dev)
Observer Est. WBF
1 (n = 1)
2.99 Hz
2.5 Hz (n = 1)
2 (n = 3)
3.92 Hz (0.31)
3.34 Hz (0.16)
(n = 2)
4 (n = 3)
4.33 Hz (0.63)
3.70 Hz (n = 1)
3.2 ±1.5 Hz
4
3
2
1
0
Observer
10
3.7 ± 1.7 Hz
Modeled
Conclusions
Automated extraction possible
Advantages = Simplicity
Color agnostic
~Range agnostic
~Wind agnostic
Acoustic agnostic
Disadvantages
Approach or aspect specific?
Allometric data for classification
Frequency overlap and specificity
11
Future Direction
Model N.A. pelagic/coastal species
Robustness: more data (species, aspect, n, etc.)
Shape Analysis
PNNL Signature Discovery Initiative
preliminary work shows promise
Combine all attributes for signature
specificity
12
½W
Range Limitations
Size in Pixels
Camera Specifications
AGD FieldPro 5X
13
Spectral range:
Pixel array:
Pixel pitch:
FOV:
Focal length:
Frame rate:
Distance from Camera (m)
3-5 microns
320 x 256
0.03 mm
6 x 4 deg.
30 mm
30 Hz
W