20
Comparison of Full-Spectrum and Zero-Crossing Automated Bat Call Classifiers Donald Solick, Matthew Clement, Kevin Murray, Christopher Nations, and Jeffery Gruver Western EcoSystems Technology (WEST), Inc.

Comparison of Full-Spectrum and Zero-Crossing Automated Bat Call Classifiers Donald Solick, Matthew Clement, Kevin Murray, Christopher Nations, and Jeffery

Embed Size (px)

Citation preview

Page 1: Comparison of Full-Spectrum and Zero-Crossing Automated Bat Call Classifiers Donald Solick, Matthew Clement, Kevin Murray, Christopher Nations, and Jeffery

Comparison of Full-Spectrum and Zero-Crossing Automated Bat Call

Classifiers

Donald Solick, Matthew Clement, Kevin Murray, Christopher Nations, and Jeffery Gruver

Western EcoSystems Technology (WEST), Inc.

Page 2: Comparison of Full-Spectrum and Zero-Crossing Automated Bat Call Classifiers Donald Solick, Matthew Clement, Kevin Murray, Christopher Nations, and Jeffery
Page 3: Comparison of Full-Spectrum and Zero-Crossing Automated Bat Call Classifiers Donald Solick, Matthew Clement, Kevin Murray, Christopher Nations, and Jeffery

Full-Spectrum (FS)

Time and Frequency

Amplitude

Multiple frequency content

-harmonics, multiple bats, calls

against background noise

Zero-Crossing (ZC)

Time and Frequency

Dominant frequency content

-loudest sound gets recorded

Page 4: Comparison of Full-Spectrum and Zero-Crossing Automated Bat Call Classifiers Donald Solick, Matthew Clement, Kevin Murray, Christopher Nations, and Jeffery

Objective

To determine which type of classifier is

better at species discrimination given

the same set of known calls

Assumption

More information = better

species discrimination?

Page 5: Comparison of Full-Spectrum and Zero-Crossing Automated Bat Call Classifiers Donald Solick, Matthew Clement, Kevin Murray, Christopher Nations, and Jeffery

FS Reference CallsSpecies Abbreviation Number of files

Perimyotis subflavus PESU 97

Lasiurus borealis LABO 40

Eptesicus fuscus EPFU 115

Lasionycteris noctivagans LANO 23

Lasiurus cinereus LACI 18

Myotis leibii MYLE 23

Myotis septentrionalis MYSE 46

Myotis lucifugus MYLU 263

Myotis sodalis MYSO 26

Page 6: Comparison of Full-Spectrum and Zero-Crossing Automated Bat Call Classifiers Donald Solick, Matthew Clement, Kevin Murray, Christopher Nations, and Jeffery

SonoBat 3.04 Northeast

Mean Classification: average of all calls in file

By Vote: majority of calls in file

Consensus: when Class and Vote agree

Page 7: Comparison of Full-Spectrum and Zero-Crossing Automated Bat Call Classifiers Donald Solick, Matthew Clement, Kevin Murray, Christopher Nations, and Jeffery

SonoBat 3.04 Northeast

Mean Classification: average of all calls in file

By Vote: majority of calls in file

Consensus: when Class and Vote agree

x = Lano Lano 0.9991

Page 8: Comparison of Full-Spectrum and Zero-Crossing Automated Bat Call Classifiers Donald Solick, Matthew Clement, Kevin Murray, Christopher Nations, and Jeffery

SonoBat 3.04 Northeast

Mean Classification: average of all calls in file

By Vote: majority of calls in file

Consensus: when Class and Vote agree

Lano Lano LanoEpfu Lano

Lano 4 of 5

Page 9: Comparison of Full-Spectrum and Zero-Crossing Automated Bat Call Classifiers Donald Solick, Matthew Clement, Kevin Murray, Christopher Nations, and Jeffery

SonoBat 3.04 Northeast

Mean Classification: average of all calls in file

By Vote: majority of calls in file

Consensus: when Class and Vote agree

Lano 4 of 5

Lano 0.9991Lano

Page 10: Comparison of Full-Spectrum and Zero-Crossing Automated Bat Call Classifiers Donald Solick, Matthew Clement, Kevin Murray, Christopher Nations, and Jeffery

ZC Classifiers

Bat Classification and Identification (BCID) East v2.4mAC

www.batcallid.com

EchoClass 64 v1

www.fws.gov/midwest/Endangered/mammals/inba/

inbasummersurveyguidance.html

Discriminant Function Analysis for New York

Developed by Eric Britzke for use by NY Dept.of Environmental

Conservation

Page 11: Comparison of Full-Spectrum and Zero-Crossing Automated Bat Call Classifiers Donald Solick, Matthew Clement, Kevin Murray, Christopher Nations, and Jeffery

Anabat Converter 0.8 (http://bertrik.sikken.nl/anabat/)

AnalookW 3.8e

Applied filter and extracted parameters

Except EchoClass

Converting FS to ZC

Page 12: Comparison of Full-Spectrum and Zero-Crossing Automated Bat Call Classifiers Donald Solick, Matthew Clement, Kevin Murray, Christopher Nations, and Jeffery

SonoBat

BCIDEcho Class NY DFA Class Vote Consensus

% Correct 49 56 45 53 43 56

% Incorrect 6 7 4 45 38 36

% Unknown 43 14 49 2 15 0

Overall Classification Rates (%)

Page 13: Comparison of Full-Spectrum and Zero-Crossing Automated Bat Call Classifiers Donald Solick, Matthew Clement, Kevin Murray, Christopher Nations, and Jeffery

SonoBat

BCIDEcho Class NY DFA Class Vote Consensus

% Correct 49 56 45 53 43 56

% Incorrect 6 7 4 45 38 36

% Unknown 43 14 49 2 15 0

Overall Classification Rates (%)

Page 14: Comparison of Full-Spectrum and Zero-Crossing Automated Bat Call Classifiers Donald Solick, Matthew Clement, Kevin Murray, Christopher Nations, and Jeffery

Species

SonoBat

BCID Echo Class NY DFA Class Vote Consensus

PESU 96 98 96 98 77 95

LABO 43 70 30 60 65 63

EPFU 88 93 86 51 57 78

LANO 74 83 74 52 26 65

LACI 53 68 53 53 58 63

MYLE 26 26 17 9 0 4

MYSE 33 33 26 57 46 46

MYLU 31 33 24 56 25 56

MYSO 0 0 0 38 31 35

Correct Classification Rates (% Correct)

Page 15: Comparison of Full-Spectrum and Zero-Crossing Automated Bat Call Classifiers Donald Solick, Matthew Clement, Kevin Murray, Christopher Nations, and Jeffery

Species

SonoBat

BCID Echo Class NY DFA Class Vote Consensus

PESU 96 98 96 98 77 95

LABO 43 70 30 60 65 63

EPFU 88 93 86 51 57 78

LANO 74 83 74 52 26 65

LACI 53 68 53 53 58 63

MYLE 26 26 17 9 0 4

MYSE 33 33 26 57 46 46

MYLU 31 33 24 56 25 56

MYSO 0 0 0 38 31 35

Correct Classification Rates (% Correct)

Page 16: Comparison of Full-Spectrum and Zero-Crossing Automated Bat Call Classifiers Donald Solick, Matthew Clement, Kevin Murray, Christopher Nations, and Jeffery

Correct Classification Rates (% Correct)

Species

SonoBat

BCID Echo Class NY DFA Class Vote Consensus

PESU 96 98 96 98 77 95

LABO 43 70 30 60 65 63

EPFU 88 93 86 51 57 78

LANO 74 83 74 52 26 65

LACI 53 68 53 53 58 63

MYLE 26 26 17 9 0 4

MYSE 33 33 26 57 46 46

MYLU 31 33 24 56 25 56

MYSO 0 0 0 38 31 35

Page 17: Comparison of Full-Spectrum and Zero-Crossing Automated Bat Call Classifiers Donald Solick, Matthew Clement, Kevin Murray, Christopher Nations, and Jeffery

Summary

• None of the classifiers performed well overall

• SonoBat better for non-Myotis, ZC better for

Myotis

• SonoBat more conservative

• Caution when using automated classification

for Indiana bat surveys

Page 18: Comparison of Full-Spectrum and Zero-Crossing Automated Bat Call Classifiers Donald Solick, Matthew Clement, Kevin Murray, Christopher Nations, and Jeffery

Caveats

• Low sample size for some species

• Lost in translation?

• Different Analook filters could improve or

worsen ZC classifier performance

• Focus on trends, not absolute comparisons

Page 19: Comparison of Full-Spectrum and Zero-Crossing Automated Bat Call Classifiers Donald Solick, Matthew Clement, Kevin Murray, Christopher Nations, and Jeffery

Conclusion

• Illustrates limitations of automated

classification of bat acoustic data

• Species presence/probable absence should be

based on multiple lines of evidence

Page 20: Comparison of Full-Spectrum and Zero-Crossing Automated Bat Call Classifiers Donald Solick, Matthew Clement, Kevin Murray, Christopher Nations, and Jeffery

Thank You!

Ryan Allen, Bat Call Identification, Inc.

Eric Britzke, US Army Engineer Research & Development Center

John Chenger, Bat Conservation and Management

Carl Herzog, New York Dept. of Environmental Conservation

Amie Shovlain, Montana Natural Heritage Program

Craig Stihler, West Virginia Dept. of Natural Resources

Joe Szewczak, Humboldt State University