Reasoning About Neural Network Activations: An Application in Spatial Animal Behaviour from Camera Trap Classifications

Abstract

Camera traps are a vital tool for ecologists to enable them to monitor wildlife over large areas in order to determine population changes, habitat, and behaviour. As a result, camera-trap datasets are rapidly growing in size. Recent advancements in Artificial Neural Networks (ANN) have emerged in image recognition and detection tasks which are now being applied to automate camera-trap labelling. An ANN designed for species detection will output a set of activations, representing the observation of a particular species (an individual class) at a particular location and time and are often used as a way to calculate population sizes in different regions. Here we go one step further and explore how we can combine ANNs with probabilistic graphical models to reason about animal behaviour using the ANN outputs over different geographical locations. By using the output activations from ANNs as data along with the trap’s associated spatial coordinates, we build spatial Bayesian networks to explore species behaviours (how they move and distribute themselves) and interactions (how they distribute in relation to other species). This combination of probabilistic reasoning and deep learning offers many advantages for large camera trap projects as well as potential for other remote sensing datasets that require automated labelling.