Predicting animal movement with deepSSF : A deep learning step selection framework

Predictions of animal movement are vital for understanding and managing wild populations. However, the fine-scale, complex decision-making of animals can pose challenges for the accurate prediction of trajectories. Integrated step selection functions (iSSFs), a common tool for inferring relationships between animal movement and the environment, are also increasingly used to simulate animal trajectories for prediction. Although admitting a lot of flexibility, the iSSF framework is limited to its reliance on pre-defined functional forms for fitting to data, and iSSFs that involve complex functional forms to model detailed processes can be prohibitively difficult to fit and interpret.

Here, we present deepSSF, an approach to fit and predict animal movement data using deep learning. The deepSSF approach replaces the log-linear model of an iSSF with a neural network architecture that receives multiple environmental layers and scalar values as inputs and outputs a single layer representing the next-step probability. We demonstrate an example deepSSF model, built in PyTorch, consisting of distinct but interacting habitat selection and movement subnetworks. This allows for explicit representation of both selection and movement processes, thus giving interpretable intermediate outputs. We apply our model to GPS data of introduced water buffalo (Bubalus bubalis) in the tropical savannas of Northern Australia.

Our deepSSF model was able to learn features that are present in the habitat covariate layers, such as linear features (rivers, forest edges) and the composition of certain habitat areas, without having to specify them pre-emptively within the model framework. It was able to capture complex interactions between the habitat covariates as well as temporal dynamics across time of day and year. Finally, our deepSSF model generally had better in- and out-of-sample predictive accuracy than the analogous iSSF model.

We expect that the deepSSF approach will generate accurate and informative predictions about animal movement, which can be used for deepening our understanding of animal–environment systems and for the practical management of species. We discuss how the wide range of existing deep learning tools could enable the deepSSF approach to be extended to represent memory and social dynamic processes, with the potential for integrating non-spatial data sources such as accelerometers and physiological sensors.