# Real‐Time Behaviour Recognition on Bio‐Loggers Enables Autonomous Audio Playback Experiments in Free‐Ranging Seabirds

## TL;DR

A new system uses bio-loggers with real-time behavior recognition to play audio in the wild, enabling experiments on free-ranging seabirds.

## Contribution

An autonomous audio playback system integrated into bio-loggers that enables location-free behavioral experiments in wild animals.

## Key findings

- The system achieved high accuracy (macro F1-score = 0.91) in real-time behavior recognition using acceleration data.
- Playback experiments successfully triggered responses in black-tailed gulls, detected through acceleration, GPS, and video data.
- The system demonstrated a proof of concept for autonomous, location-free audio playback experiments in natural environments.

## Abstract

Audio playback experiments in the natural environment have been a powerful tool in animal behaviour and ecology, revealing causal relationships between animal movements/behaviours and audio stimuli. However, traditional audio playback experiments could only be performed in limited locations and/or situations where direct observation and/or video recording by human observers or installation of automated devices, such as camera traps, were possible. To overcome the limitation, we designed an autonomous audio playback system on bio‐loggers in the natural environment. In this system, an on‐board machine learning model estimates animals' behavioural state (e.g., flying or not) in real time using data from a low‐power accelerometer. If the target behaviour (e.g., flying) is detected and other predefined criteria are met, the logger activates high‐cost sensors, including a video camera, and plays audio from a built‐in speaker. The logger can record fine‐scale behavioural data before, during, and after the playback using multiple modalities (e.g., acceleration, GPS, and video). To examine the validity of the system, we performed field experiments targeting freely ranging black‐tailed gulls (
Larus crassirostris
) in Japan. The real‐time behaviour recognition using acceleration data demonstrated high accuracy in the field experiments (macro F1‐score = 0.91). The playback experiments were performed almost perfectly as we intended when birds were flying outside the colony (46 playback events were collected from eight birds), except for several failures due to hardware malfunctions. Using three response indicators (based on acceleration, GPS, and video data), Bayesian statistical modelling and causal inference analysis showed that several birds clearly responded to the audio stimuli, but to both predator call and noise sound. Despite some remaining practical challenges, the results demonstrated a successful proof of concept for the proposed audio playback system on bio‐loggers. By removing the location constraints of traditional playback experiments, the system allows a variety of playback experiments to be tested in various situations. In the future, the system can be extended to stimulate other sensor modalities (e.g., magnetic sensors), expanding the possibilities for intervention methods in the wild environment.

We developed an autonomous audio playback system integrated into bio‐loggers to conduct behavioural experiments in wild animals without location constraints. Field tests with free‐ranging black‐tailed gulls demonstrated accurate real‐time behaviour recognition and successful audio playback events, revealing behavioural responses to audio stimuli. This system could expand the potential for intervention‐based studies in natural environments.