Original papers
Freedom to lie: How farrowing environment affects sow lying behaviour assessment using inertial sensors

https://doi.org/10.1016/j.compag.2019.01.035Get rights and content

Highlights

  • Inertial data is used to classify activity transitions in free-farrowing sows.

  • Additional sensors improve activity classification accuracy.

  • Free-farrowing sows are more prone to less controlled lying behaviour.

  • Farrowing is preceded by an increase in activity transition frequency.

Abstract

We address the use of accelerometery to automatically monitor lying behaviour in free-farrowing sows; due to their freedom of movement and the consequent increased variety of movements the sows are able to exhibit, the challenges in automating this are greater than in sows housed in movement restricting farrowing environments. The methodology developed was applied to two salient applications: that of farrowing prediction through detection of nest building activity, and comparison of maternal lying behaviour in conventional movement-restricting and free-farrowing systems. Two sensors were attached at both the front and hind end to each of eight periparturient sows. Movement behaviour was recorded for a period of five days around parturition. Activity transitions were classified by a Support Vector Machine classifier, using data from both sensors individually, and combined; classifier output was validated against ground truth annotations collected from video data. We draw conclusions about the benefits of using multiple sensors over a single sensor, as well as the suitability of different sensor locations on the sow. Activity classification was found to improve through the use of multiple sensors, with a mean F1 score (a measure of predictive performance between 0 and 1) of 0.84, compared to use of the front sensor alone (mean F1 = 0.49) and the hind sensor alone (mean F1 = 0.57). Activity transitions were classified using the dual sensor setup with a mean F1 score of 0.77. Using a threshold-based approach, taking transition frequency as an indicator of nesting behaviour, we were able to detect the onset of nest building with an average latency to farrowing of 11.1 (±4.65) hours, and an average of 1 premature detection per sow; however, the majority of these premature were in a particular sow. We draw comparisons between the lying behaviour of free-farrowing and restricted sows. Using a mixed-design ANOVA we found a main effect of farrowing environment on transition duration (p=0.003), peak acceleration (p=0.007) and rate of change in pitch (p=0.009). Improving the classification accuracy of sow activity transitions through the addition of multiple sensors allows for improved performance in applications such as farrowing prediction, which has the capacity to reduce piglet mortality through enabling farrowing supervision. Understanding how movement restriction affects the lying behaviour of farrowing sows has the potential to inform decisions regarding restriction of sows and development of free-farrowing environments.

Introduction

The use of accelerometers to quantify animal behaviour has become widespread over recent years (Jukan et al., 2016). Studies have been conducted to investigate a range of animal behaviours, from routine activities such as running and playing in dogs (Ladha et al., 2013), to more specific behaviours targeted at assessing welfare, such as lameness assessment in cows (Pastell et al., 2009). Due to their small size and the versatility of the data produced, accelerometers have been found to be particularly effective at monitoring animal behaviour in large-scale settings (Matheson et al., 2016).

Automatic quantification of posture and lying behaviour has potential to enhance the welfare and productivity of various domesticated species. Changes in posture and activity may provide indications of underlying health and welfare issues (Matthews et al., 2016, Szyszka and Kyriazakis, 2013, Weary et al., 2009). Changes in activity and lying behaviour of pregnant sows can be indicative of the onset of farrowing, allowing for intervention and supervision, or to identify sows that pose less risk to their piglets by their lying behaviour (Marchant et al., 2001, Špinka et al., 2000). Accelerometery that quantifies the behaviour of sows during the farrowing process has used single sensors on animals (Cornou and Lundbye-Christensen, 2008, Oczak et al., 2015). The use of multiple sensors to perform activity recognition in sows has been undertaken (Ringgenberg et al., 2010), in which one sensor was mounted to the back of the sow, as well as one secured to the rear leg of the animal. The leg worn sensor was targeted at assessing stepping behaviour, and postural assessments only utilised a single sensor at a time. Prior to the work described in this paper, we also conducted experiments into the use of accelerometry to quantify maternal lying behaviour of sows housed in farrowing crates, using a single accelerometer secured to the hind end of the pig (Thompson et al., 2016).

It is standard practice in pig systems to move a parturient sow to a farrowing crate several days prior to the expected date of farrowing. This practice improves the survival rates of the piglets (Cronin and Smith, 1992), however has been shown to increase stress in the sow (Lawrence et al., 1994) and supress natural maternal behaviour (Damm et al., 2000, Jarvis et al., 2004). Farrowing crates are designed to restrict the movement of the sow, and as such approaches to automatically classify and quantify sow behaviours may be relatively straightforward. This restricted repertoire of behaviours reduces opportunity for misclassification between behaviours and allows for a more simplistic approach to classification. Despite this simplification, this does not account for the additional subtlety of movement that may be express by crated sows, given that they are moving under more tightly constrained conditions.

On the other hand, alternative free-farrowing systems, such as PigSAFE (Edwards et al., 2012) aim to optimise welfare and economic performance by allowing increased freedom of movement and expression of natural behaviour, whilst providing enhanced safety features for new-born piglets. When allowed free movement, the problem of automated behaviour classification becomes considerably more complex. Additional behaviour states must be considered and can be expressed with fewer physical constraints, specifically behaviours that involve movement from one area of the pen to another. Understanding the differences in sow lying behaviour between movement restricting and free-farrowing environments has potential impact on the management of farrowing. Quantification of these effects, if possible, may also have implications for promoting uptake of higher welfare systems.

Throughout this paper, we will refer to the combination of posture state and moving behaviours as “activities”, and the period in which a sow moves from one activity to another as “activity transitions”. It is the aim of this work to explore the potential for increasing activity transition classification accuracy through the use of multiple accelerometers and quantify improvement in the two farrowing systems. The methodology was applied to two salient applications that of farrowing prediction and comparison of maternal lying behaviour in movement restricting and free-farrowing systems. This allows us to demonstrate the applicability of the approach in scenarios that have practical significance. We hypothesised that the improvement in accuracy through the use of two sensors will be mainly in the alternative farrowing system.

Section snippets

Study animals

Eight hybrid sows from the same batch at Newcastle University Cockle Park pig unit were used for assessment and moved to farrowing accommodation three days prior to the expected date of farrowing. The sows were between 2nd and 4th parity. They were housed in PigSAFE farrowing environments2, providing them with freedom of movement throughout the farrowing process. A floor plan for the PigSAFE systems is shown in Fig. 1.

Activity classification

Given 5 frames of data per second for 6 h per day, for all 8 pigs on each of 5 days there were a total of 4,320,000 frames of activity data to be classified. Using a leave-one-pig-out cross validation technique the F1 scores for the three sensor setups are described in Table 1. For all behaviours F1 scores were higher when two sensors were used as opposed to one. Mean F1 scores were taken from classifier results from all 8 subjects. Mean F1 score for the combined front and rear sensor setup was

Discussion

Whilst alternative farrowing accommodation is currently only employed for a relatively small proportion of farrowing sows within the pig industry, this is set to increase due to increasing concerns over sow welfare (Edwards et al., 2012). The advantages of being able to quantify sow lying behaviour have been explored in movement-restricted sows in a previous work (Thompson et al., 2016). There we showed that lying behaviour assessment has the potential to identify sows with a predisposition to

Conclusion

We present a novel approach to classification of activity and activity transitions in farrowing sows, and assess the efficacy of various sensor placements and numbers. We found that through the use of multiple sensors we were able to achieve higher levels of accuracy for activity classification, and that if only a single sensor is to be used, it is better placed at the rear of the sow, rather than at the head end. We drew comparisons between movement-restricted and free-housed sows,

Ethics

The experiments were performed at Newcastle University, Cockle Park farm. Attaching the accelerometer(s) on the sows was the only deviation from normal husbandry procedures; previous work suggests that this does not change the repertoire of sow behaviours. The study was approved by the Newcastle University Animal Welfare Ethics Review Board.

Acknowledgements

This project has received funding from the European Union’s Seventh Framework Programme for research, technological development and demonstration under grant agreement n° 613574 (PROHEALTH). This project has also received funding from the Biotechnology and Biological Sciences Research Council (BBSRC) in the form of a studentship to RT.

References (34)

Cited by (15)

  • Application of deep learning in sheep behaviors recognition and influence analysis of training data characteristics on the recognition effect

    2022, Computers and Electronics in Agriculture
    Citation Excerpt :

    An embedded Multi-Layer Perceptron-based feed-forward neural network was used to process sensor data. Thompson et al. (2019) used accelerometry to monitor the lying behavior of free-farrowing sows automatically. Sensors were placed in various locations around the sows, and a specific behavior was determined based on the sensor data threshold.

  • Assessment of lying down behaviour in temporarily crated lactating sows

    2021, Animal
    Citation Excerpt :

    However, King et al. (2019) covered only very limited periods (6 h before and 6 h 2 days after crate opening) without analysing piglet behaviour. In line with this finding, Thompson et al. (2019) showed that free-farrowing sows are more prone to less controlled lying behaviour than sows in crates. In contrast, Höbel et al. (2018) reported that sows lie down more carefully after crate opening in temporary crating systems than in farrowing crates.

  • Accelerometer systems as tools for health and welfare assessment in cattle and pigs – A review

    2020, Behavioural Processes
    Citation Excerpt :

    A majority of studies report percentages of correctly classified lateral lying exceeding 90 % and F1 scores ranging from 0.90 to 0.98. The best results were achieved when using two accelerometers: one mounted on the front and one on the hind end of the back of a sow (F1 score for lying on the left side: 0.98, F1 score for lying on the right side: 0.97, as reported by Thompson et al. (2019)). For sternal lying, model evaluation results are in general a little bit lower than for lying on the side.

View all citing articles on Scopus
1

Current address: School of Interactive Computing, Georgia Tech, USA.

View full text