Behaviour classification of extensively grazed sheep using machine learning

Highlights

• Ear-tag accelerometers can detect multiple behaviours with reasonable accuracy.

• Epochs of 10 s and 30 s proved superior for behaviour detection.

• ML performance is dependent on the method of behaviour classification.

Abstract

The application of accelerometer sensors for automated animal behaviour monitoring is becoming increasingly common. Despite the rapid growth of research in this area, there is little consensus on the most appropriate method of data summation and analysis. The objective of this current study was to explore feature creation and machine learning (ML) algorithm options to provide the most accurate behavioural classification from an ear-borne accelerometer in extensively grazed sheep. Nineteen derived movement features, three epochs (5, 10 and 30 s) and four ML-algorithms (Classification and Regression Trees (CART), Linear Kernel Support-Vector Machines (SVM), Linear Discriminant Analysis (LDA) and Quadratic Discriminant Analysis (QDA)) were assessed. Behaviour classification was also evaluated using three different ethograms, including detection of (i) grazing, lying, standing, walking; (ii) active and inactive behaviour; and (iii) body posture. Detection of the four mutually-exclusive behaviours (grazing, lying, standing and walking) was most accurately performed using a 10 s epoch by an SVM (76.9%). Activity was most accurately detected using a 30 s epoch by a CART (98.1%). LDA and a 30 s epoch was superior for detecting posture (90.6%). Differentiation relied on identification of disparities between behaviours rather than pattern recognition within a behaviour. The choice of epoch and ML algorithm will be dependent on the application purpose, with different combinations of each more accurate across the different ethograms. This study provides a crucial foundation for development of algorithms which can identify multiple behaviours in pasture-based sheep. This knowledge could be applied across a number of contexts, particularly at times of change in physiological or mental state e.g. during parturition or stress-inducing husbandry procedures.

Introduction

Behaviour is often used by researchers to better understand an animal’s interaction with their environment and physiological state (Frost et al., 1997, Barwick et al., 2018a). However, behaviour is often difficult to consistently monitor, especially when animals are in large numbers or spread over vast distances (Dobos et al., 2015). The development of sensor technologies has improved our ability to remotely monitor livestock in a broad range of contexts and on scales not previously possible (Brown et al., 2013, Schmoelzl et al., 2016). Of these sensors, one type that has increased in popularity is the accelerometer (Fogarty et al., 2018). Accelerometers measure both gravitational and inertial acceleration associated with movement, usually on three different axes (called tri-axial) (Brown et al., 2013, Alvarenga et al., 2016, Barwick et al., 2018b). Recent advances in miniaturisation of technology has increased accelerometer uptake, with reduced size, mass and power consumption making attachment to the animal easier and less invasive (Watanabe et al., 2008, Alvarenga et al., 2016, Walton et al., 2018). Extensive beef and sheep industries are now exploring the potential for these systems to optimise production, reduce costs and enhance sustainability (Trotter, 2018).

In sheep, accelerometers have previously been used to detect basic behaviours (specifically, behavioural states rather than behavioural events) such as high- and low-level general activity (McLennan et al., 2015), gait and posture (Radeski and Ilieski, 2017) or some combination of grazing, lying, standing, ruminating, running and/or walking (Nadimi et al., 2012, Alvarenga et al., 2016, Giovanetti et al., 2017, Barwick et al., 2018b, Decandia et al., 2018, Mansbridge et al., 2018, Walton et al., 2018). More specific applications have included the detection of suckling (Kuźnicka and Gburzyński, 2017) and lameness (Barwick et al., 2018a). These studies vary in their approach, with differences in study purpose, design, sensor attachment and data sample rate. Some have also been conducted in controlled pen environments, either wholly (Alvarenga et al., 2016, Giovanetti et al., 2017, Barwick et al., 2018a) or in part (Radeski and Ilieski, 2017). As these sensors evolve into commercially affordable systems, further evaluation in sheep kept under normal grazing conditions is warranted. This is particularly important for behaviour signatures that may subtly differ between pen and pasture environments (e.g. differences in biting signatures of animals eating a total mixed ration compared to those actively tearing pasture (Martz and Belyea, 1986)) or behaviours that may arise due to extensive management conditions (e.g. increased insect-defence behaviours such as ear-flicking or head shaking (Dougherty et al., 1993, Mooring et al., 2003)).

In addition to differing study design, there has also been a number of ways in which behaviour states have been interpreted and classified. For example, Alvarenga et al. (2016) used accelerometers to classify five mutually-exclusive sheep behaviours (grazing, lying, running, standing and walking). Similarly, Barwick et al., 2018b, Walton et al., 2018 classified lying, standing and walking behaviours, with Barwick et al. (2018b) including an additional classification of grazing. Other approaches include Mansbridge et al. (2018), where classification was focused on grazing, ruminating and other non-eating behaviours and Rurak et al., 2008, Umstätter et al., 2008 where behaviours were classified as either ‘active’ or ‘inactive’. While a higher degree of resolution in behavioural observation might be desirable from a research perspective, it is entirely feasible that simple bivariate classifications (e.g. active/inactive) may prove reliable enough to be applied in commercial contexts and should not be immediately dismissed.

The first step in most accelerometer-based data analysis involves the process of feature extraction. There have been a large number of features proposed in the literature, ranging from simple averages of a single accelerometer axis (Hokkanen et al., 2011) to more complex metrics designed to capture the variability of signal magnitude across all three axes (Alvarenga et al., 2016, Barwick et al., 2018b, Walton et al., 2018). Most features are created using a fixed time segment commonly referred to as the ‘epoch’ (Decandia et al., 2018). Epochs help to reduce both the amount and complexity of data and reduce noise in the dataset (Chen and Bassett, 2005, Barwick, 2016). However, choosing an optimal epoch length can be challenging, as epochs are required to be both short enough to maximise the likelihood of capturing a single behaviour yet sufficient in duration to allow adequate differentiation between behaviours (Chen and Bassett, 2005). While this method of data summation is common, there is not yet a consensus on the most appropriate epoch length for behaviour detection.

Once the features have been created, these are then analysed using any one of a number of machine learning (ML) algorithms e.g. discriminant analysis (Giovanetti et al., 2017, Barwick et al., 2018a, Barwick et al., 2018b, Decandia et al., 2018); classification trees (Alvarenga et al., 2016); random forest (Barwick et al., 2018a, Barwick et al., 2018b, Mansbridge et al., 2018, Walton et al., 2018); support vector machine (Mansbridge et al., 2018). Whilst some studies have compared the value of different epochs and ML algorithms, none have comprehensively evaluated the many different combinations in parallel and the resulting ability to classify multiple behaviour states in extensively grazed sheep.

The objective of this current study was to explore how a range of ML algorithms (Classification and Regression Trees (CART), Linear Kernel Support Vector Machine (SVM), Linear Discriminant Analysis (LDA) and Quadratic Discriminant Analysis (QDA)) can be used to predict behavioural states in sheep. Algorithms were assessed using three epoch lengths (5, 10, 30 s). Assessment of calculated features was also conducted to determine the best features for behaviour prediction. Analysis was applied to three different ethograms: (i) detection of four mutually-exclusive behavioural states (grazing, lying, standing, walking); (ii) detection of general activity (active – grazing, walking; inactive – standing, lying); and (iii) detection of body posture (upright – grazing, standing, walking; prostrate – lying). Though similar aspects of accelerometer application have been studied in sheep (Umstätter et al., 2008, Alvarenga et al., 2016, Giovanetti et al., 2017, Barwick et al., 2018b, Decandia et al., 2018, Mansbridge et al., 2018, Walton et al., 2018) and cattle (Martiskainen et al., 2009, Robert et al., 2009, Smith et al., 2016, Abell et al., 2017), the focus of this study was to assess multiple combinations of analysis protocols, with the objective of identifying algorithms appropriate for commercial application in extensively grazed sheep. Further to this, while comparable applications are already present in the dairy industry (Trotter, 2013), there is still a requirement to study these aspects in sheep given the differences between the two species e.g. bio-mechanical differences related to common behaviours and the resulting difference in acceleration signatures (Chambers et al., 1981, Barwick et al., 2018b).