The effect of different time epoch settings on the classification of sheep behaviour using tri-axial accelerometry

Highlights

• Tri-axial accelerometer sensors are widespread to monitor behaviour of animals.

• Discriminant analysis classifies feeding behaviours in grazing sheep.

• Sampling the three axes of acceleration produces an enormous quantity of data.

• An epoch of aggregation window can reduce the quantity of data.

• The 30 s epoch length yields the most accurate classification in grazing sheep.

Abstract

Monitoring behaviour of grazing animals is important for the management of grazing systems. A study was run to discriminate between the main behaviours (grazing, ruminating and other activities) of sheep at pasture wearing a halter equipped with an accelerometer (BEHARUM device), and to identify the epoch setting (5, 10, 30, 60, 120, 180 and 300 s) with the best performance. The BEHARUM device includes a three-axial accelerometer sensor and a force sensor positioned under the lower jaw of the animal. The halter was fitted to eight Sarda dairy sheep that rotationally grazed either a spatial association (mixture) or a time association of berseem clover (Trifolium alexandrinum L.) and Italian ryegrass (Lolium multiflorum Lam.) for 6 h day⁻¹. The behaviour of the animals was also video-recorded. The raw acceleration and force data were processed for each epoch setting to create 15 variables: the mean, variance and inverse coefficient of variation (ICV; mean/standard deviation) per minute for the X-, Y-, Z-axis and force, and the resultant. Multivariate statistical techniques were used to discriminate between the three behavioural activities: canonical discriminant analysis (CDA), and discriminant analysis (DA). To validate the derived discriminant functions, a bootstrap procedure was run. To evaluate the performance of DA in discriminating between the three activities, the sensitivity, specificity, precision, accuracy and Coehn’s k coefficient were calculated, based on the error distribution in assignment. Results show that a discriminant analysis can accurately classify important behaviours such as grazing, ruminating and other activities in sheep at pasture. The prediction model has demonstrated a better performance in classifying grazing behaviour than ruminating and other activities for all epochs. The 30 s epoch length yielded the most accurate classification in terms of accuracy and Coehn’s k coefficient. Nevertheless, 60 and 120 s may increase the potential recording time without causing serious lack of accuracy.

Introduction

Monitoring the behaviour of grazing ruminants is important to understand how animals meet their requirements in pastoral systems and to achieve optimal plant production, animal forage intake and performances (Carvalho, 2013). Since observing animal behaviour is a labour-intensive and difficult task, whether it is performed with direct observations or through video recordings, most research has concentrated on recognizing feeding behaviour of ruminants from animal attached sensors. A type of sensor that has recently become widespread in research studies is the tri-axial accelerometer, since it is small, inexpensive, and easy to wear (Brown et al., 2013).

Accelerometers have been widely used to automatically detect and classify several behaviours in cattle, e.g. oestrus detection (Ueda et al., 2011), walking (Robert et al., 2009) feeding and standing activities in a free-stall barn (Arcidiacono et al., 2017), sleeping posture (Fukasawa et al., 2018) and time (Hokkanen et al., 2011), and eating, ruminating and resting activities (Watanabe et al., 2008).

Fewer research studies have been conducted to classify sheep behaviours than cattle behaviours. Umstätter et al. (2008) used integrated pitch and roll tilt sensors, and found that they could distinguish between two categories: active and inactive, with more than 90% accuracy. Other studies on sheep behaviour used the collar attached Actiwatch accelerometer system for classifying activity levels and detecting diurnal rhythms (Piccione et al., 2010, Piccione et al., 2011). Other authors (Nadimi et al., 2012, Nadimi and Søgaard, 2009) used the ADXL202 accelerometer to detect grazing, lying down, standing, walking, mating and drinking in sheep with a mean accuracy of 76.2%. Alvarenga et al. (2016) successfully identified grazing and non-grazing states, with accuracies higher than 83%, in grazing sheep wearing an accelerometer under the lower jaw. More recently Giovanetti et al. (2017a), positioning a device containing an ADXL335 accelerometer sensor in the same place, were able to classify grazing, ruminating and resting behaviour of sheep at pasture with an overall accuracy of 93%.

Tri-axial accelerometer based devices can acquire and store information internally, thus consuming very little battery power. However, the amount of data that can collected is limited by the size of the memory card within the device. On the other hand, data can be directly transmitted to a central receiver for subsequent processing. This practice, however, requires a high power consumption (Vázquez Diosdado et al., 2015).

The sampling frequency of such devices usually ranges from 8 to 100 Hz, thus producing an enormous quantity of data, proportional to the sampling frequency, which can lead to a rapid depletion of the memory device and to high costs in terms of battery consumption caused by sending and receiving large data sets. These restrictions could be overcome by undertaking some form of preliminary processing of the accelerometer data on the device itself settling and applying to the data stream, for a given sampling frequency, an optimal aggregation window (called epoch).

Optimizing the epoch setting, without compromising classification accuracy, could imply a number of advantages. Short epoch settings could increase the labour involved in processing data, deplete the memory device, decrease the battery duration and may also cause erroneous attribution activities during processing. Actually, if an epoch shorter than the period of time an activity occurs is used, the number of false positive classifications for dynamic activities could increase probably due to transitioning between different activities or body shifts during static activities (Robert et al., 2009). Conversely, optimized longer epoch settings might reduce the memory depletion and increase the battery duration without compromising the performance of the sensor. Nielsen (2013) distinguished grazing from non-grazing behaviour with a 3D activity sensor that correctly classified the behaviours of dairy cows with a relatively high accuracy when the epoch was set at 5 s, 5 or 10 min. Other authors, as Vázquez Diosdado et al. (2015), while classifying lying, standing and feeding behaviours in dairy cows, reported a small increase in the decision-tree classification algorithm performance at the largest window size of 10 min if compared with 1 and 5 min epoch settings. In the present research, a customized tri-axial accelerometer based sensor, able to either store data in a micro SD card or send them to a remote computer, was used. In the future perspective of data pre-processing in the device itself, determining the optimum device settings before field application is crucial, because they could impact on monitoring system accuracy as well as on the effective battery and memory life.

The objectives of this study were: (1) to develop an algorithm based on the multivariate statistical analysis to discriminate the main behaviours (grazing, ruminating and other activities) of sheep at pasture equipped with a customized tri-axial accelerometer based sensor named BEHARUM; (2) to determine the performance of the algorithm in terms of accuracy, sensitivity, specificity, precision and Coehn’s k coefficient, at different epoch settings (5, 10, 30, 60, 120, 180 and 300 s); and (3) to select the epoch that optimizes the system accuracy of the device.