Modelling animal movement for behavioural inference

Abstract

Effective animal conservation is more important than ever in the face of biodiversity and climate crises. Understanding animal movement and behaviour can aid spatial planning and inform conservation management. However, it is difficult to directly observe and record behaviours in remote and hostile terrain such as the marine environment. Researchers often attempt to infer animal behaviour from telemetry data using hidden Markov models (HMMs), as animal movement patterns are influenced by their underlying behaviour. However, these inferred behaviours are not typically validated due to difficulty in obtaining ‘ground truth’ behavioural information. In practice, ecological researchers often use the inferred behaviours to suggest practical ways of improving conservation efforts. This includes identifying and designating important foraging areas as Marine Protected Areas (MPAs), spatial planning, and construction of offshore wind farms. Additionally, since some of these animals, particularly those in the marine environment, receive less protection at sea, it is crucial that the inferred behaviors are actually identified in the right areas. Furthermore, large-scale animal telemetry datasets are readily available due to advances in biologging technologies, and studies on animal movement and behaviour have been on the increase as a result. The use of these telemetry datasets for statistical modelling can be computationally costly due to the large volume and high-resolution of the datasets. To easily fit these models, the need to reduce the number of data points prior to analysis often arises. In the process of reducing the number of data points, some of the information embedded in the original dataset that may be useful for analysis can be lost.

In the first part of this thesis, we have a dataset obtained from the visual tracking method that provides a ground truth behavioural dataset of terns and the boat GPS track as a proxy for bird tracks. Leveraging on this unique data, we assess whether (i) the visual tracking information from the boat is a good proxy for true bird locations in relation to inferred behaviours of the fitted HMM, and (ii) the inferred behaviours from HMMs fitted to visual tracking data accurately represent true behaviours as identified by behavioural observations taken from the boat. We demonstrate that visual tracking data can be regarded as a good proxy for true movement data in terms of similarity in inferred behaviours. In the second validation, we assess the validity of HMMs-inferred behaviour by fitting HMMs to boat visual tracks, inferring behaviours from fitted models, and assessing inferred behaviour using ground-truth observed behavioural data. Our results suggest that HMMs fitted to tracking data can accurately identify important conservation-relevant behaviours in seabird species, as demonstrated using visual tracking data.

In the second part of this thesis, we examine a more efficient way of subsampling animal telemetry data in addition to the standard thinning approach. We adopt the Nyquist-Shannon sampling theorem (NSST) to inform the choice of subsampling frequency. While NSST can be used to inform the temporal resolutions for obtaining animal movement trajectories, it also describes how to sub-sample a discrete-time signal (animal movement data) in a way that there is little or no information loss about the underlying latent behavioural process. We explain how to adopt NSST in animal movement modelling to inform the choice of sampling frequency and subsequently reduce the number of data points while minimizing information loss. We initially show that NSST can be useful for determining whether a chosen sampling interval is sufficient for behavioural analysis. We also showed using two scenarios (i) cases where the standard thinning approach alone produces results similar to the NSST approach with respect to behavioural inference and close representation of original movement location, (ii) cases where the NSST approach provides better results than the thinning approach. Lastly, we examine how the accuracy of inferred behaviours is sensitive to the choice of subsampling frequencies informed by NSST. Results suggest that the NSST approach is useful for subsampling as it provides improved behavioural accuracy compared to the standard thinning approach.