Murdoch University Research Repository

Welcome to the Murdoch University Research Repository

The Murdoch University Research Repository is an open access digital collection of research
created by Murdoch University staff, researchers and postgraduate students.

Learn more

Establishing best practice for the classification of shark behaviour from bio-logging data

Hounslow, Jenna L. (2018) Establishing best practice for the classification of shark behaviour from bio-logging data. Honours thesis, Murdoch University.

[img]
Preview
PDF - Whole Thesis
Download (3MB) | Preview

Abstract

Understanding the behaviours of free-ranging animals over biologically meaningful time scales (e.g. diel, tidal, lunar, seasonal, annual) gives important insights into their ecology. Bio-logging tools allow the remote study of elusive or inaccessible animals by recording high resolution multi-channel movement data, however archival device recording duration is limited to relatively short temporal-scales by memory and battery capacity. Machine learning (ML) is becoming common for automatic classification of behaviours from large data sets. This thesis develops a framework for the programming of bio-loggers for the classification of shark behaviour through the optimisation of sampling frequency (Chapter 2) and the choice of movement sensor (Chapter 3).

The effects of sampling frequency on behavioural classification were assessed using data published in a previous study collected from accelerometer equipped juvenile lemon sharks (Negaprion brevirostris) during captive trials in Bimini, Bahamas. The impacts of different combinations of movement sensors (accelerometer, magnetometer and gyroscope) were assessed using data collected from sub adult sicklefin lemon sharks (Negaprion acutidens). Sharks were equipped with multi-sensor devices recording acceleration, angular rotation and angular velocity during captive trials at St Joseph Atoll, Seychelles. Catalogues of discrete classes of behaviours (ethograms) were developed by observing sharks during captive trials.

Behaviours (swim, rest, burst, chafe, headshake) were classified using a random forest ML algorithm with predictor variables extracted from the ground-truthed data. A range of sampling frequencies (30, 15, 10, 5, 3 and 1 Hz) and combinations of movement sensors were tested. For each dataset, a confusion matrix was determined from model predictions for calculation and comparison of evaluation metrics. Classifier performance was best described by the class or macro F- score, a measure of model performance, one indicating perfect classification and zero indicating no classification.

As sampling frequency decreased, classifier performance decreased. Best overall classification was achieved at 30 Hz (F- score >0.790), although 5 Hz was appropriate for classification of swim and rest (>0.964). Behaviours characterised by complex movements (headshake, burst, chafe) were best classified at 30 Hz (0.535- 0.846). Classification of behaviours was best with a tri-sensor combination (0.597), although incorporating an additional sensor (magnetometer or gyroscope) resulted in little increase in classifier performance compared to using an accelerometer alone (0.590 compared to 0.535 respectively).

These results demonstrate the ideal sampling frequencies and movement sensors for best-practice programming of bio-logging devices for classifying shark behaviour over extended durations. This thesis will inform future studies incorporating behaviour classification, enabling improved classifier performance and extending recording duration of bio-logging devices.

Item Type: Thesis (Honours)
Murdoch Affiliation: School of Veterinary and Life Sciences
Supervisor(s): Gleiss, Adrian and Daly, R.
URI: http://researchrepository.murdoch.edu.au/id/eprint/41604
Item Control Page Item Control Page

Downloads

Downloads per month over past year