Statistical Advancements to Modelling Animal Movement in Aquatic Environments

Abstract

Movement of animals provides information on ecological processes that influence individual survival and fitness like migration, foraging, and breeding. Over the past several decades, large-scale interest in the tracking of aquatic animals has sparked both technological and statistical innovation. Predicting accurate locations from noisy tracking data, as well as classifying different kinds of movement that can be used to infer animal behaviour, have emerged as prime objectives that can lead to better understanding of where animals go, and what processes might be driving their behaviour. These goals persist across taxa, such that ecologists wish to apply similar methods to data with varying structures. In tandem, coordinated efforts by international collaborative projects have generated massive datasets that require new solutions capable of answering increasingly complex ecological questions.

This thesis develops new methods that can concurrently account for measurement error in the tracking technology, and predict different kinds of movement using switching hierarchical models. Traditionally, switching hierarchical models have been fitted with techniques that sample from the likelihood or posterior distribution and which are computationally intensive. Instead, the workhorse of this research is the combination of the Laplace approximation and Automatic Differentiation, which is shown to enable the rapid fitting of hierarchical models with large numbers of random effects. Because switching hierarchical models have both continuous and discrete random effects, an iterative procedure is adopted to efficiently optimize the marginal negative log-likelihood.

The novel switching hierarchical model frameworks that are presented can accommodate different combinations of discrete and continuous space and time processes. Throughout, simulations are used to demonstrate a high level of accuracy under each of the contrasting scenarios. These methods are applied to animal tracking datasets collected using a variety of technology, including both satellite and acoustic telemetry. The general implementation developed here is versatile, and could be expanded to incorporate alternative mixed-scale random effects (e.g., individual or temporal effects). A groundwork is thereby laid for the statistical innovation that will be needed to accommodate the increasing size and complexity of contemporary animal telemetry research.