Abstract:
Efforts to better understand patterns of animal behaviour have often been restricted by
several environmental, human and experimental limitations associated with the collection
of animal behavioural data. The introduction of new bio-logging technology has offered
an alternative means of recording animal behaviour continuously and is being used
in an increasing number of studies. Accurately calibrating these bio-loggers, however,
still remains a challenge in many cases. Using lions as an example species, we test
how audio recordings from animal-borne acoustic sensors can improve calibration and
behaviour classification. Through a collaborative effort between computer scientists,
engineers, and zoologists, custom designed acoustic bio-loggers were fitted to eight
lions and recorded audio simultaneously with accelerometer and magnetometer data.
Audio recordings were then used as the source of ground truth to train random forest
classification models as well as to provide additional predictor variables for behaviour
classification. We demonstrated near-perfect classification performance for five lion
behaviour classes when all component variables were combined, with an average per-
class precision of 98.5%. Using accelerometer features only, the audio-trained classifier
predicted behaviours with an average per-class precision of 94.3%. On-animal audio
recordings are therefore able to provide a valuable source of ground-truth for calibrating
bio-loggers while also offering additional predictive features for increasing the accuracy of
behaviour classification. This technological innovation has wide ranging application and
provides a useful tool for behavioural ecologists wishing to collect fine scale behavioural
data for animal research and conservation.
