The 'aukward' use of hidden Markov models in a pursuit-diving seabird: semi-supervised HMMs improve behavioural classification and energetic inference in auks.

Abstract

Global seabird declines underline the need for accurate behavioural inference at sea to guide conservation. For pursuit-diving species such as auks, differentiating between resting and foraging is difficult due to similar above-water movement patterns, biasing estimates of behaviour-specific habitat use and energy expenditure. We deployed combined GPS-time depth recorders on 28 puffins during the 2021, 2023, and 2024 breeding seasons and compared inferred behaviours using outputs from GPS-only, dive-informed, semi and fully supervised hidden Markov models (HMMs) using state forcing rules. GPS-only and dive-informed models routinely conflated resting/drifting and foraging behaviours. The semi-supervised model best distinguished behavioural states, identifying 89% of track segments containing at least one dive as foraging, providing a more conservative approach compared to the supervised model when classifying transit, while avoiding inflating the number of foraging for segments with large step lengths. Compared to the GPS-only model, used here as baseline, the dive-informed model provided greater estimates of energetic expenditure by 12–21% relative to the models using forcing rules. Integrating dive data into a semi-supervised HMM refines behavioural classification, enabling more realistic energy budgets and precise mapping of important behaviour-specific habitats for conservation planning.