Post-hoc validation and prediction of hunting behaviour using unsupervised accelerometry classification: a case study of Adélie penguins.

Abstract

1. Fine-scale monitoring of free-ranging animals remains logistically challenging, particularly for cryptic diving marine predators that spend much of their lives underwater in remote environments. Rapid miniaturisation and expansion of biologging technology have greatly improved the ability to uncover features of an animal’s behaviour, physiology, and environment, albeit with increasingly large and complex datasets. Unsupervised machine learning methods are therefore a popular approach for identifying latent behavioural states without requiring a priori labels. However, because these methods typically lack robust validation procedures, there is little objective basis for determining whether inferred behavioural states correspond to underlying biological processes.
2. To address this limitation, we develop a validation pipeline that benchmarks latent behavioural states against annotations of observable events. These annotations enable post-hoc assessment of model performance and improve interpretation of the inferred latent states. The validation outputs can also be integrated into GLMMs to generate probabilistic predictions of the annotated events based on associated behavioural parameters. Our pipeline therefore offers a resource-efficient way to incorporate variability-inclusive predictive tools into behavioural studies where a priori information is limited.
3. We tested this validation pipeline in R and Python, using accelerometry data from diving Adélie penguins (Pygoscelis adeliae) as a case study. By annotating prey capture events (PCEs) on a subset of accelerometry data using co-deployed video camera footage, we validated the performance of an unsupervised machine learning approach based on the Expectation-Maximization (EM) algorithm in defining a latent ‘hunt’ behaviour. We determined a pooled true positive rate of 98.7% for EM-derived hunting sequences across accelerometry datasets when validated against independently annotated prey capture events. Additionally, we demonstrate that the probability of a PCE in Adélie penguins can be determined through a combination of the sequence’s duration, maximum depth, and acceleration metrics.
4. Overall, our validation pipeline provides a flexible framework for validating, interpreting and generating predictive models from unsupervised machine learning approaches in behavioural and movement ecology.

KEYWORDS
Machine learning; validation; behavioural classification; accelerometry; Pygoscelis adeliae; prey capture event

Date of Award	19 Aug 2026
Original language	English
Awarding Institution	Bangor University
Sponsors	Santander Open Academy, Challenger Society for Marine Science & Taith
Supervisor	Marianna Chimienti (Supervisor), Akiko Kato (Supervisor) & Yan Ropert-Coudert (Supervisor)