Refining Accelerometer-Based Animal Behaviour Classifications With Sequence-Informed Post-Processing.

Supervised machine learning has been used to detect fine-scale behaviours from animal-borne accelerometers by dividing the continuous sequences of behaviour into discrete segments and classifying each with a distinct behavioural category. This approach, while widely implemented, discards the sequential information available in the temporal ordering of the behavioural series. 'Post-processing' (smoothing and error correction made after the initial classifications) can be used to improve the accuracy of the original predictions by learning from the natural transitions and durations of behaviours. While broadly implemented across other classification domains, this technique has been underutilised for accelerometer-based animal behaviour classification. In this paper, we compare the performance of five different post-processors (modal, duration-based, transition-based, Hidden Markov Model, and a Naive Bayes smoother) against the original predictions from the base classifier across 15 animal accelerometer datasets across 13 species. Overall, while there was no single post-processing method that was optimal in every dataset, we find Bayesian smoothing to have the best overall performance improvement (average 6.4% increase in