Dr Jasmeet Kaler, University of Nottingham, UK discusses her research of developing machine learning algorithms for livestock health and welfare.
Precision livestock farming (PLF) has been defined as the management of livestock farming by the continuous automated real-time monitoring of the livestock’s health and welfare. There are lots of different sensor technologies that are available that could help in this direction. For example, data gathered from sensors on animal movement and behaviour could be really beneficial for the early prediction of disease.
Although PLF offers a lot of promise, the progress in this area has been vastly limited in terms of the available technologies used for the reliable prediction of diseases. In this article, I will discuss some of the key issues surrounding this, based on our research in this area led by my group at the University of Nottingham, UK. The research reported in this article was funded by BBSRC and by Innovate UK.
Technical challenges around sensor selection
There are a lot of different aspects to consider when it comes to the choice of a sensor (e.g accelerometer, accelerometer and gyroscope, GPS etc). This includes where the sensor will be positioned, what the sampling rate will be, and how the data will be transmitted. All of these considerations have an impact on the accuracy of the algorithms, as well as the scalability and practicality of the solution that could therefore be utilised on the farm.
Although previous studies have tried to ‘predict’ a behaviour based on the signals from a sensor, not many studies have compared and/or evaluated all of these previously mentioned considerations. In our previous work, we evaluated sensor position, sampling frequency and window size for the processing of data for sheep behaviour classification. From these results, it was suggested that with the sampling rate of 16Hz and window size seven seconds, with sensor position on ear, it is possible to successfully classify basic sheep behaviour with an accuracy of around 95% (this included behaviours such as standing, walking, lying and grazing).
Feature selection and which algorithms are the best to implement
The next question we are faced with is what, and how machine learning features are needed, and which algorithms are therefore best to tackle the problem of classification. Again, this is another area lacking in evidence. It is important to consider these different features because a number of features and algorithms tend to have different computational costs.
In our study for classifying sheep grazing behaviour, we demonstrated that from a set of 44 features, only five to seven features are needed to yield highly accurate results, therefore, this offers valuable insights for implementation. For real-time systems, large feature sets are problematic due to computational complexity and higher storage requirements.
‘Concept drift’ and biologically complex systems
In addition to energy considerations, one of the key technical challenges for real-time and long-term behaviour monitoring is ‘concept drifts’. Concept drift occurs when a system is required to adapt to a change in data distributions within the concept.
In supervised classification problems, it is generally assumed that the data in the design model is randomly selected from the same distribution as the points that will be classified in the future. This is an unrealistic assumption due to the dynamic nature of many different classification problems. For example, when a system is trained in one environment, behavioural classification in animals can also show discrepancies in performance given environment variance or heterogeneity. Such discrepancies can be due to differences in the animals (age, breed, etc.) and/or environmental characteristics (terrain elevation, type of soil, particular farm constrains, etc.).
Discrepancies in behaviour classification performance with IceTag devices, have been reported with high (35%) mismatch values between the proportion of steps. One possible solution is the use of methods that incorporate concept drift in the architecture. We recently published such an approach for the first time in the precision livestock farming.
Our proposed system (see Fig. 1) is a hybrid system that combines offline and online learning algorithms (the offline algorithm is completely static), and it is pre-trained with offline information, while the online algorithm continues learning through the deployment lifetime utilising online information. We demonstrate that the accuracy of the approach outperforms both the offline classifier and online algorithm.
One of the biggest advantages of this approach is that it is computationally efficient due to embedding the offline part of the algorithm on to the device i.e. ‘on the edge’ and then using fewer features for online learning. In conclusion, this approach has considerable potential for improving classification accuracy in real-time deployment scenarios and potential for improving animal welfare and livestock production.
Farmers and technology adoption
Another important piece of the puzzle is the technology adoption on farms. It’s important to understand how farmers interpret the value of technology in the context of their farm. We explored sheep farmers view on technology, its use and adoption on farms using social science methods. Results suggested farmers have two conflicting frames of reference relating to precision technology.
Firstly, that it could potentially be of value to their overall farming business and secondly, that it posed a risk to their animals and their role/identify as a good stockman by creating a physical distance between the two. Frames of reference are complex sets of assumptions and attitudes which are used to filter perceptions and create meaning including beliefs, schemas, preferences, values and culture and other ways in which we bias or undermine our judgement. These frames of reference would need to be understood and acted on by key stakeholders (manufacturers, farmers, vets, suppliers) to stand a greater chance of getting farmers to act and use the technology to its full potential.
Please note, this article will appear in issue 32 of SciTech Europa Quarterly, which is available to read now.