Thesis
Sheep grazing and lambing behaviour analysis from accelerometer data using deep learning
Masters by Research, Murdoch University
2022
Abstract
Classification of sheep behaviour from a sequence of tri-axial accelerometer data has the potential to enhance sheep management and improve lamb survival. Machine learning (ML) can be employed for classification, learning the patterns of the signals to identify sheep behaviour. However, sheep behaviour is inherently imbalanced (e.g., more ruminating than walking) resulting in underperforming classification for important minority activities. In this thesis, two studies were conducted: (i) Classification of grazing behaviour and (ii) Classification of lambing behaviour. For the grazing behaviours study, traditional ML, e.g., Random Forest (RF) and deep learning (DL) techniques were compared. Existing research is limited to statistical and traditional ML methods. We investigated two DL models, Long Short-Term Memory (LSTM) and Bidirectional LSTM (BLSTM), appropriate for sequential data. We designed several multi-class classification studies, with imbalance being addressed using synthetic data, and incorporated the number of steps a sheep performed in the observed 10s-timewindow. DL models achieved superior performance (e.g., 88%) to traditional ML models, especially with augmented data, and showed superior generalisability. LSTM, BLSTM, and RF achieved sub-millisecond average inference time, suitable for real-time applications.
The lambing behaviours study applied LSTM to a lambing behaviours task, focusing on identifying labour and post-birth licking behaviour. Previous research has not utilised DL for the task, and has focused on activity levels, or licking only, for lambing prediction. We executed studies with varying number of classes for the lambing data, including and excluding grazing behaviours, and performing fine-tuning on individual sheep to address generalisability. A combined grazing and lambing behaviours model achieved the best performance (e.g, 81%) with classification of the labour and licking behaviours strongest. Fine-tuning on individual sheep resulted in increased performance (e.g., 86% average).
The study presents a strong foundation for development of such models for real-time animal activity profiling, labour duration monitoring and detection of lambing events.
Details
- Title
- Sheep grazing and lambing behaviour analysis from accelerometer data using deep learning
- Authors/Creators
- Kirk E Turner
- Contributors
- Ferdous Sohel (Supervisor) - Murdoch University, Centre for Crop and Food InnovationHamid Laga (Supervisor) - Murdoch University, Centre for Biosecurity and One HealthDean Diepeveen (Supervisor)Andrew N Thompson (Supervisor) - Murdoch University, Centre for Animal Production and Health
- Awarding Institution
- Murdoch University; Masters by Research
- Identifiers
- 991005567366607891
- Murdoch Affiliation
- School of Information Technology; College of Science, Technology, Engineering and Mathematics
- Resource Type
- Thesis
UN Sustainable Development Goals (SDGs)
This output has contributed to the advancement of the following goals:
Metrics
128 File views/ downloads
251 Record Views