Application of hidden Markov models and their extensions to animal movement data

Hidden Markov Models (HMMs) have become increasingly popular in animal movement studies as they provide a flexible modelling approach and take the correlation between successive observations into account. They can segment the movement paths into latent states, which can be considered as rough proxies for the behaviours of the animals. This study comprises of two sections, both involving the application of HMMs to large terrestrial mammal movement data. Usually step lengths representing the displacement distances between successive observations, turning angles measuring the tortuosity, or a bivariate input of both variables are used as inputs in the models. It has been found in the literature that the turning angle is either included in the modelling process or it is excluded without much justification for doing so. The first part of this study investigates the nfluence of the turning angle on the model output and resultant interpretations of the HMMs when modelling the trajectories of large terrestrial mammals in southern Africa. Results revealed at different time scales, and for both predator and herbivore species in this study, that the turning angle does not influence the state allocation of the HMMs, which is the main output in terms of interpreting the behaviours of the animals. It is thought in most cases that the inclusion of the turning angle overcomplicates the models unnecessarily without contributing any additional information in terms of the behavioural interpretations or improving the overall fit of the models. This was found for the variety of movements of the species under observation in this study. The second part of this study attempts to validate the state allocation of the HMMs fitted to eland trajectories in the Greater Addo Elephant National Park in the Eastern Cape, with the use of camera trap data. This presented a unique opportunity as this type of data is mainly used for abundance or capture-recapture studies, and the HMMs are rarely validated as the true behaviours of the animals are seldom known. Results revealed that the same diel patterns were detected by the HMMs that were shown by the classified camera trap data. Direct comparisons of the observations where the dates and times matched for the telemetry and camera trap data could be done in several rare instances, which revealed many similarities. Although it was not an ideal comparison, the camera trap data provided a rough validation of the state allocation of the HMMs used in the study