Accelerometers on sharks?

The overall aim of this project is to use motion sensing sensors, observational ethograms and machine learning algorithms to develop best practice guidelines for the classification of shark behavior using Machine Learning algorithms.

Sharks, like most wildlife, are facing increasing anthropogenic pressures, ranging from direct fishing, by-catch, habitat destruction to climate change. In addition to these pressures, Sharks are intrinsically biologically vulnerable due to their low fecundity and late sexual maturation.

In many cases, effective science based management relies on accurate data on the impacts of these many pressures on the behavior and ecology of individual animals, in order to devise effective policy. In many cases, these have been solved by the advent of telemetry able to resolve the distribution of individuals in space and time. However, in many cases, responses of individuals may not be distributional in nature and may be related to changes in the behavior or physiology of the individual.

Novel tags, such as accelerometers promise to solve these problems, but the potential of these techniques are yet to be fully realized.

The study of the behaviour of free-ranging animals in their natural environment is often problematic due to observational difficulties and inadvertently altering natural behaviour. The advent of telemetry has transformed our ability to make “observations” while animals are out of sight and undisturbed in their natural environment. For sharks, acoustic and satellite telemetry has yielded a wealth of knowledge on spatio-temporal distribution of individuals , in turn elucidating many aspects of their life-history, such as home ranges (Heupel et al. 2004), seasonal migrations (Weng and Block 2004), reproductive migrations (Campana et al. 2010) and ontogenetic changes in habitat-use (Whitty et al. 2009). Despite the great advances these technologies have made, relating these data to ecologically meaningful activities (e.g. foraging, refuging and mating) has relied on many assumptions (Whitney et al. 2012). Yet identifying spatio-temporal patterns in such activities may be crucial for some aspects of management and conservation, such as identifying critical habitat for mating or foraging.

Advances in bio-logging techniques now permit the recording of high-resolution motion data, which has the potential to allow researchers to classify behavior at sub-second resolution, as pioneered on largely terrestrial animals (Wilson et al. 2008). These types of tags are now increasingly applied in shark research (Gleiss et al. 2016; Gleiss et al. 2011a; Payne et al. 2016; Whitney et al. 2010; Whitney et al. 2007). Although relatively new, this technology has already resulted into a number of insights into fundamental aspects of shark biology, such as swimming behavior (Gleiss et al. 2011a; Nakamura et al. 2015; Nakamura et al. 2011), activity rhythms (Bouyoucos et al. 2017a; Gleiss et al. 2013) and energy demands (Gleiss et al. 2010; Lear et al. 2017), as well as increasingly studies of applied value, such as recovery fishing capture (Bouyoucos et al. 2017b; Whitney et al. 2016).

Despite the biological insights gathered by these studies to date, there remains significant scope for improving these methodologies. It is currently assumed that acceleration has to be recorded at high frequency to discern biologically relevant signals. However, this is largely based on anecdotal evidence of researchers classifying these data by eye, without the use of sophisticated classification algorithms. The reason for this paucity of data is that few researchers have access to large sharks under controlled conditions.

This project seeks to explore how the classification of shark behaviour using motion sensors (gyroscope, magnetometer and accelerometer) can be optimized through appropriate choices of sampling frequency and the sensors utilized. This will result in a best practice guide for the programming of such tags, in turn increasing the efficacy of future studies, by either requiring fewer animals to be tagged since more data is collected per individual or by increasing the duration per individual offering important insights into longer-term behavioural changes.

• Identify the most important sensors for the classification of behaviour
• Investigate the influence of tag sampling frequency on the classification of behaviour
• Evaluate the trade-off between recording duration (based on the number of sensors utilised and their sampling frequency) and the accuracy of classification