A general hidden Markov model framework for capturing changes in species behaviour under disturbance in acoustic time series

Continuous biodiversity monitoring is crucial for understanding ecosystem dynamics in an era of global environmental change. Advances in bioacoustic hardware facilitate autonomous monitoring of vocalising animals in terrestrial and aquatic ecosystems. Time series of processed audio data can provide insights into multi-species behavioural responses to imminent disturbances. Here, we present a general Gaussian hidden Markov model (HMM) framework for analysing processed species detection data from audio recordings to identify changes in species behavioural dynamics under sudden and short-term (pulse) disturbances. Our framework transforms species detection data by calculating the logarithmic change in species detection counts between consecutive time points, focusing on shifts in temporal variability rather than counts per se. The framework includes a suite of HMMs with varying complexities in their number of states, constraints on the mean, and inclusion of covariates. We recommend an ensemble of in-sample and out-of-sample model selection methods that balance complexity, interpretability, and forecasting ability. We illustrate the framework using processed bird species detection data from an acoustic sensor array in Okinawa, Japan. To demonstrate the ability of our framework to detect changes in species vocalisation behaviour, we analysed 66 days of bird vocalisation data from before, during, and after two large typhoons struck Okinawa in 2018. A parsimonious three-state mean-constrained model and its non-homogeneous variant with precipitation were selected. The estimated HMM states represent ‘ambient’, ‘warning’ and ‘disturbed’ periods, respectively capturing low, medium, and high variability in vocal activity. A warning state consistently preceded a disturbed state, suggesting that our framework could help detect early behavioural responses to impending pulse disturbances. These findings demonstrate how species behavioural dynamics inferred from high-resolution monitoring can provide early warning signals of emerging ecological disturbances.