Characterising sheep vocals using a machine learning algorithm : A thesis submitted in partial fulfilment of the requirements for the Degree of Master of Applied Science at Lincoln University

Abstract

New Zealand’s economy is mainly dependent on the farming sector and the sheep sector is one of the most important farming sectors, playing a backbone role to the agricultural industry and placing New Zealand among the top five sheep exporter countries in the world. International consumer trends show concerns over the well-being of animals before slaughter and research also indicates potential negative effects on meat quality of stressed animals. Indicators for sheep well-being have largely been limited to physical weight gain and visually observable behaviour and appearance. There has been recent interest but little substantive research on sheep vocalisation as a means of monitoring sheep well-being. This assumes that sheep vocalisation can be classified as representing different states of well-being. Therefore, this thesis investigated the potential to be able to classify sheep vocalisations in a way that would enable automated assessment of the well-being of New Zealand sheep using recorded vocalisations.

A supervised machine learning approach was used to classify the sheep vocals into happy and unhappy classes. Sheep sounds were collected from a New Zealand Ryeland sheep stud farm and online databases. After collection, these sounds were labelled by an expert, pre-processed to make them clean from unwanted background sound noises and features were extracted and selected for classification. Models were built and trained and tested.

Model use in this research shows that sheep sounds were classified into happy and unhappy classes with an accuracy of 87.5%, for the sheep vocals used in this research. Through demonstrating the ability for automated classification of sheep vocalisations this research opens the door for further study on the well-being of sheep through their vocalisations. Future researchers could also collect larger vocal data sets across different breeds to test for breed-related variance in vocalisations.. This may enable future sheep well-being certification systems to be established to assure consumers of the well-being of pre-slaughter sheep life.