Pipeline for acoustic classification of ant (Formica rufa) behaviours: A suggested workflow for soil macrofauna studies

Abstract

Acoustic monitoring has been touted as an automatable monitoring technique to study the presence, activity or acoustic communication of soil animals. Yet current acoustic approaches have been criticized for lacking understanding of what soil fauna behaviours they are detecting. We investigate whether specific insect behaviours could be identified with sound by recording Red Wood Ants (Formica rufa) behaviours (running, struggling, biting and stridulation-like behaviour) with a contact microphone. We label recordings by hand, then extract two types of acoustic features, spectral and cepstral features, and show that most of the selected ant behaviours could be separated from each other based on their acoustic profiles. Based on our results, we propose an analytical workflow including principal component analysis as well as random forest classification that generates correct behavioural classifications in 67–90% of the sound snippets. Key features to separate behaviours include cepstral skewness, duration of sound events and mean frequency of the generated sound. Even if we encountered background noise, our pipeline could select relevant features from audio files and train a machine learning classifier. Our study demonstrates that acoustic features can be used to inform about specific insect behaviours and thus broaden the application of passive acoustic monitoring to species that do not communicate acoustically.