Efficient Statistical Inference for High Frequency Movement Data

Recent years have seen a proliferation of high-frequency animal movement data, often exceeding 1 Hz, enabling much greater insight into behaviour than was possible with lower-resolution data. In particular, it is now possible to detect the precise moments at which animals make navigational decisions, placing the concept of movement as a sequence of “steps and turns” on a more rigorous footing.

Summarising high-frequency trajectories in terms of the timing and angles of discrete turning points is crucial for many downstream analyses. This is especially important given that widely used techniques in movement ecology, such as hidden Markov models (HMMs) and step selection analysis (SSA), were originally developed with coarser data in mind.

In this thesis, I propose a statistically principled method for identifying turning points in high-frequency movement data, with minimal reliance on arbitrary parameter choices. The method is formulated as a statistical changepoint detection problem, in which the animal’s path is represented as a stream of heading data (typically derived from magnetometry) and denoised as a stepwise function over time.

Building on this segmentation, I develop a hidden Markov model framework for behavioural state inference based on steps defined between successive turning points. The model accommodates variable-duration steps and incorporates emission distributions for step angles, durations, and speeds. I further extend this framework by allowing transition probabilities to depend on environmental covariates, enabling state-switching dynamics to be driven by external factors.

Together, these contributions provide a coherent and computationally efficient framework for analysing high-frequency animal movement data, with strong potential to improve behavioural inference in ecological studies.