This document discusses automated monitoring and identification of Orthoptera (grasshoppers and crickets) based on their acoustic signals. It describes analyzing recordings of different species to identify patterns and periodicities in their calls at the pulse, syllable and echeme levels. Machine learning methods like random forests and dynamic time warping are used to classify and cluster recordings. A new project aims to evolve an automated identification system by mapping acoustic traits onto a supertree of Orthoptera and combining time and frequency domain features to identify known species from their recordings.

1. Towards automated monitoring of
Orthoptera
(and some other noisy stuff)
Ed Baker, Hannah O’Sullivan, Quentin Geissmann

2. A detour on frogs
Tachycnemis seychellensis
• Found on a number of
islands in the Seychelles
• Phylogeny of populations
shows no lineages specific
to individual islands
• Analysis of calls (16
parameters in frequency
domain from 3 islands)

3. A detour on frogs
Random forest method
(machine learning
classification)
Centre mixed as phylogeny
would suggest, but
differences between islands
Studying acoustics can reveal
interesting observations

4. Gryllotalpa vineae

5. Gryllotalpa vineae

6. Having fun in the time domain
• For automated studies lots of work has been done on pulse,
and inter-pulse features
• The pulse is the basic element of a song, but there is also
higher-level structures (syllables, echemes,… )
• The terminology of higher level structures generally (ideally)
reflects biological meaning

7. Chorthippus dorsatus
Can see high level structure
• Periodicity of approximately 1.5s (echeme)

8. Chorthippus dorsatus
Zoomed in, another level of structure is apparent
• Periodicity of approximately 80ms (syllable)

9. Chorthippus dorsatus
Repeat…
• Periodicity of approximately 4ms (pulse)

10. Chorthippus dorsatus
Towards computer identification
• The computer does not care about how individual authors
define syllables
• Just look for ‘rhythmicity’
In this example the rhythmicity spans over three orders of
magnitude
• Continuous wavelet transform

11. Chorthippus dorsatus

12. It’s good to have lots of
recordings

13. Dynamic Time Warping
Used to compare temporal sequences that may vary in speed
Can create clusters based on this

14. Very short
recording
Atypical

15. Time domain analysis
• Beat spectrum has potential for automated identification
• Particularly in combination with frequency domain
Requires:
• Annotated libraries of recorded songs (with multiple
recordings of each species)

16. A new project
Automated Acoustic Observatories
• Three year Leverhulme Trust funded project
• Two main themes
• Evolution of song in the Orthoptera
• Automated identification

17. Evolution
Supertree of Orthoptera
• 163 source trees
• Mapping of trait data, including acoustics to tree
• Methods from communications theory
As well as understanding the evolution of song, used to guide
identification algorithm in categorising unknown songs

18. Automated Identification
Combine time and frequency domain methods to identify
known orthoptera species.
• Acoustic features
• Weighted by location, time of year, habitat (niche models)
Provide as much information as possible about unknown songs
• Automatically record vouchers
• Best-effort attempt to categorise (e.g. family X, call adapted to
dense grass)