P.PSH.1205 - Developing and piloting a real-time monitoring system for sheep

Summary

Utilising sensors this pilot project focused on the development of artificial intelligence systems that can accurately determine normal sheep behaviours – and deviations from - ultimately resulting in the development of predictive and autonomous monitoring systems. The project also involved a review of other applications of sensors in sheep systems including for uses such as pasture monitoring and remote predation alerts.

Objectives

1. Developed and verified normal behaviour algorithms on 50 ewes, on a minimum of five properties utilising different breeds of sheep under different nutritional conditions
2. Utilised behavioural algorithms developed in objective one to predict key behaviours over lambing for synchronised ewes on collaborating properties
3. Utilised outcomes from objectives one and two to support remote monitoring of a small number of ewes using real time sensors that will test the ability of the algorithms and processes developed to autonomously monitor lambing ewes and to provide remote alerts to the producer
4. Conducted a detailed market analysis, analysing the further development that would be required to enable the development and deployment of a system across the wider sheep industry. This will include:
a. A detailed cost benefit analysis of the value of autonomous alerts at lambing supported by the experimental data outputs
b. A review of the unit cost and proportion of the flock that would require sensors to derive the benefits of the technology
c. A review of:
i. Engagement with a commercial partner to investigate the development and release of a commercial version of the developments achieved
ii. A competitor analysis and value proposition-to a commercial partner-analysis
5. Engaged a minimum of 25 sheep producers in the project and surveyed their perceptions of the value of autonomous monitoring systems for their enterprises
6. Attended a minimum of two industry events to communicate the objective and outcomes of the project
7. Published a minimum of two producer and industry facing articles
8. Submitted a minimum of one peer-reviewed journal paper for review

Key findings

This project clearly identified that producers were very interested in the concept of autonomous monitoring of livestock and very willing to invest time and energy into seeing progress in this area of work. The project has developed the world’s largest data set of sheep behaviour linked to accelerometer output. The work conducted here will be fundamental in guiding future endeavours in automating aspects of livestock management. The algorithms that have been developed and refined throughout the process have demonstrated the power that machine learning will hold in the future of the grazing livestock industries. The autonomous monitoring of lambing ewes has provided several challenges. A range of aspects that the project team considered to be unique to lambing events turned out to be specific to individual sheep and not generic to all sheep. This made it difficult to define an accelerometer pattern that described when a ewe was due to lamb or had recently lambed. This prompted the project team to move toward a vision-based approach to lambing identification. This approach holds promise and has been a useful development of the project. While commercial application of these technologies to farmers still requires additional development, the project has demonstrated the significant potential that already exists to automate some aspects of livestock research and bring a new level of precision to livestock research.

Benefits to industry

This project has thoroughly explored the sensor landscape as applicable to sheep and has built foundational data sets and techniques that can inform both future research efforts as well as commercial interests exploring this space. This work has been made publicly available so that technology developments in the future can start from a competitive advantage compared with where this project started. This project has developed a data set that is several fold larger than any database previously established. It has investigated the different algorithms that can be deployed to these types of data sets and found those that are most likely to deliver a successful outcome. The project has paved the way for more efficient and more accurate research in the future where sensors can be incorporated into research programs and animal activity can be predicted without the need to monitor them using research staff.

Future research

This project has set the foundation for a range of work using cameras and sensors in combination to automate several processes on Australian sheep farms. The work has demonstrated that while the technology holds significant promise and will eventually be a gamechanger for the industry, the foundational work is both difficult and risky. Future research efforts should be directed toward developing techniques that utilise camera-based approaches to ‘measure’ on-farm outcomes. The livestock industry sits at the edge of a technology revolution that will fundamentally change the way farms are operated. The transition to this new era will be considerably smoother if the likes of MLA continue to invest in technological change. The project has demonstrated that livestock research can significantly benefit from automation of data collection, and it is recommended that autonomous monitoring approaches become the norm in funded research projects.

Specifically, the project has demonstrated that:

• When utilising a neural network trained on a different data set, some labelling of data is required to ensure that predictions are relevant to a new data set
• Sensor systems based on cameras need to use video footage and animal tracking rather than still images
• The combination of metric-based behavioural models with time series transformers is the best method of prediction of time of lambing
• It is best to use both unsupervised learning and self-supervised learning to boost the effective size of the datasets
• Near real-time training of datasets on edge devices needs to be incorporated into future work to allow unsupervised learning and supervised learning to be merged in near real-time.