The next BRAG meeting will be held next Thursday the 20th June at 1 pm via Zoom/GP-Y801. Next week we will have presentations by Scott Forest and Charlotte Patterson.
Zoom Link: https://qut.zoom.us/j/83410867532?pwd=anTpu44owkkqht6HHvX9dN6ZA8tl1O.1
Meeting ID: 834 1086 7532
Password: 819498
Scott’s Talk
Title: Predicting animal movement using a deep learning step selection framework
Abstract: Predicting animal movement is difficult due to the fine-scale and complex decisions made by animals as they move. A common approach to infer the relationships between animals and their external environment is a step selection analysis (SSA). SSA approaches have recently been used to generate animal trajectories for prediction, although the SSA model fitting framework is limited in several ways, namely that regression is performed on point locations that ignore the structure and composition of habitat features, and it is prohibitively difficult to fit and interpret models that include interacting components such as multi-scale temporal dynamics. Fortunately, the goal of SSA model fitting is simply to determine the next step probability given the surrounding environment, and deep learning is aptly suited to this task. This study will train and compare between neural network architectures that receive scalar values and multiple habitat layers as inputs, and output a single layer representing the next-step probability. This is equivalent to SSA model fitting, but benefits offered by deep learning include the ability to learn features that are present in the habitat covariate layers, such as linear features (rivers, forest edges), and the composition or size of certain habitat areas. It can also represent the complex and seemingly abstract interactions between habitat covariates, time of day or year, and memory and social dynamic processes. Additionally, there is promising potential for integrating non-spatial data sources such as accelerometers and physiological sensors. We expect that the deep learning approach will lead to generating more accurate animal movement trajectories. Our research is motivated by the need to accurately predict invasive water buffalo and cattle locations in Northern Australia, both of which cause significant environmental damage and represent economic opportunities for traditional landowners.
Charlotte’s Talk
Title: When can integrated SDMs improve spatial model transfers? Antarctic ‘islands’ as a case study
Abstract: The rapid expansion of biological data collection and immediate need for biodiversity assessments motivate the development of methods that can harness benefits from multiple data types. Integrated Species Distribution Models (ISDMs) have been shown, under certain conditions, to outperform SDMs derived from single datatypes in the precision and accuracy of parameter estimates and can improve spatial prediction within the sampling domain. However, the limits to benefits arising from data integration are not well explored, particularly when an ISDM is transferred to a spatially or environmentally distant site. Two components of an ISDM might improve spatial model transfer: the shared learning of coefficients from different observation datasets, and the incorporation of a shared spatial random effect, that can capture unobserved ecological processes. We present a case study across Antarctica’s ice-free ‘islands’, where spatially constrained survey design and an urgent need for improved prediction of biodiversity patterns motivate the use of ISDMs. With this case study we pose two questions: (1) Can ISDMs improve spatial transfer relative to single dataset approaches and under what conditions? And (2) How does adding or removing a spatial random effect change the success of a spatial transfer? Using recent data from East Antarctica, we test these questions with integrated presence-only and presence-absence models. We validate model predictions at a separate ice-free site with an independent presence-absence dataset.