Dr Jasmeet Kaler of the University of Nottingham School of Veterinary Medicine and Science, UK, discusses the development of a system for the automatic detection of sheep behaviour and lameness
The livestock sector around the world is facing huge challenges. Over the next 15 years, global demand for meat is expected to increase by 40% and future milk production will need to increase by more than 1.8% per annum. The challenge for the livestock farms is to increase productivity whilst maintaining the highest standards of welfare. Precision livestock farming has been a growing area over the past decade that aims to offer solutions by providing tools that can help to collect automatic data on animal behaviour, health, welfare, and production in real time and aid in on-farm management decisions. However, these solutions have to be economically viable, reliable, and practical to be implemented by the farmer.1
The challenge: sheep lameness
Lameness in sheep is one of the biggest health and welfare problems in sheep industries around the world.2 Footrot caused by the bacteria Dichelobacter nodosus leads to 90% of lameness on farms. There are approximately £3m-worth (~€3.4m) of lame sheep each year in the UK alone, and the cost of lameness is around £80m, attributable to loss of production and treatment costs.3 Therefore, lameness substantially reduces the efficiency of sheep farming and adversely affects the drive towards national and international food security.
Our work over the past ten years has identified that even mildly lame sheep have footrot and that the early treatment of lame sheep along with culling repeatedly lame sheep is the best approach to manage this condition.4,5 This work impacted UK policy, as published in the Farm Animal Welfare Committee report, and the current best practice.3 This best practice relies on the rapid treatment of lame sheep to reduce the spread of disease and lower levels of lameness; however, sheep – being prey animals – tend to mask signs of lameness.
The solution: precision livestock technology
Precision behaviour monitoring systems and sensor technologies using accelerometers and gyroscopes can be used to automatically classify sheep behaviour and detect key signs of lameness (altered gait and posture). Currently there are no such devices/solutions or validated algorithms for sheep lameness and behaviour detection and classification.
Two key challenges in the development of such technology are to have a) algorithms that can correctly classify sheep behaviour, including lameness, with high accuracy; and b) optimum ways to gather and process data generated from sensors and implement algorithms. To achieve this one needs to do a systematic evaluation of key elements such as sampling rates, position of sensors, and window sizes for processing on accuracy of algorithms and also choice of architecture for data processing. Given that these elements dictate the overall accuracy of algorithms and the feasibility of having a real-time monitoring solution, it’s quite surprising to see scant literature in precision livestock that systematically investigates this and goes beyond mere algorithm development for any behaviour/welfare/health classification.
The approach: Early lameness detection for lameness control
In the multidisciplinary project EL4L (Early lameness detection for lameness control) – funded by the Biotechnology and Biological Sciences Research Council (BBSRC), Grant No. BB/N014235/1, and Innovate UK (132164) – my group is working with industry partners Intel and Farm Wizard. The consortium comprises specialists in engineering technology, software development, veterinary epidemiology, and data analytics.
The aim of the project is to develop and validate a system for the automatic detection of behaviour and lameness in sheep. In this project we gathered raw data from sheep of different breeds and body condition score using Intel custom-based devices with a 16 bit tri-axial gyroscope and 16 bit tri-axial accelerometer. We employed various machine learning approaches to develop algorithms and then implemented those on a device using ‘edge’ analytics. Our research work specifically looked into the following elements over a series of studies:
- We started by exploring the effect of different sampling rates (8, 16 and 32Hz), position of sensors (ear or collar), and window sizes (3s, 5s and 7s) for signal processing on accuracy of algorithms for sheep behaviour and also effect of energy consumption. From the magnitude of accelerometer and gyroscope 11 feature characteristics were extracted, resulting in a total of 44 feature characteristics. More details of this study can be found in the journal Royal Society Open Science.6 We investigated classification of three activities: ‘lying’, ‘standing’, and ‘walking’. Our algorithms indicated high accuracy (>90%) for classifying all activities. More importantly, the study results also highlighted that the above-mentioned elements impact accuracy and that there are trade-offs (between these elements) one needs to consider for a real-time monitoring solution;
- The next series of studies looked at the development of novel machine learning algorithms to classify sheep lameness, specifically the choice of algorithms and number of features; and
- We then worked on the implementation of algorithms on the ‘edge’ and employing ‘edge’ analytics for a classification of sheep behaviour and analysis. This is novel and the first of its kind in precision livestock monitoring as all the ‘thinking’ is done on the device and it offers potential benefits from a battery life perspective.
The implemented algorithms are currently being validated in a large trial, and we recently showcased our research and device (January 2018) at the BBSRC Innovation Hub at the Oxford Farming Conference. We hope that the results of this study will be useful not only for farmers but also for both industry and academia in the domain of precision livestock monitoring and data analytics.
The team: Ruminant Population Health Research Group
As part of the Ruminant Population Health Research Group at the University of Nottingham School of Veterinary Medicine and Science, we are very interested in using our understanding of animal behaviour and disease biology coupled with advanced analytics to develop solutions that improve ruminant health and welfare and have an impact on the farm. We work with industry and a wide range of stakeholders (farmers, veterinarians, processors, retailers). My group is exploring precision livestock approaches to detect lameness in other species and researching the development of algorithms using a range of precision technologies for cattle and sheep health in various other projects. For more details on these projects see: https://www.kaler-researchgroup.co.uk/
- Lima E, Hopkins T, Gurney E, Shortall O, Lovatt F, Davies P, Williamson G, Kaler, J (2018). Drivers for precision livestock technology adoption: A study of factors associated with adoption of electronic identification technology by commercial sheep farmers in England and Wales. PLoS ONE 13(1): e0190489. https://doi.org/10.1371/journal.pone.0190489
2. Kaler J, et al. Factors associated with changes of state of foot conformation and lameness in a flock of sheep (2010). Preventive Veterinary Medicine, 97 (3-4), 237-244. DOI: 10.1016/j.prevetmed.2010.09.019
3. Farm Animal Welfare Committee: Opinion on lameness in sheep (2011). https://www.gov.uk/government/publications/fawc-opinion-on-sheep-lameness
4. Kaler J, et al. The inter- and intra-observer reliability of a locomotion scoring scale for sheep (2009) Veterinary Journal, 180 (2), 189-194. DOI: 10.1016/j.tvjl.2007.12.028
5. Kaler J, et al. Randomized clinical trial of long-acting oxytetracycline, foot trimming, and flunixine meglumine on time to recovery in sheep with footrot (2010). Journal of Veterinary Internal Medicine, 24 (2), 420-425. DOI: 10.1111/j.1939-1676.2009.0450.x
6. Walton E, Casey C, Mitsch J, Vázquez-Diosdado J A, Yan J, Dottorini T, Ellis K A, Winterlich A, and Kaler J (2018). Evaluation of sampling frequency, window size and sensor position for classification of sheep behaviour. R Soc Open Sci 5:171442