Acoustic detection of bees in the field using CASA with focal templates

This paper describes a method of detecting buzzes of bees in field audio. Detecting the buzzing of bees from environmental recordings is an instance of sound scene analysis. In this work, we build upon prior work in computational auditory scene analysis (CASA), employing spectral clustering techniques to mitigate the weakness of the target signal, coupled with a newly-introduced concept of “focal templates”. This system yields promising results on a previously acquired collection of environmental recordings, yielding results consistent with human performance, and, in some cases, improving upon human performance. Our success in this task suggests that the combination of focal templates and spectral clustering may prove valuable in other sound scene analysis tasks, especially when the target may be well-defined but may suffer from low signal-to-noise ratio (SNR). Survey recordings with manual (visual and acoustic) annotations were processed, and the algorithm yielded very favorable results. The potential for deploying this approach into a low-cost pollinator monitoring system is discussed.