Automated monitoring of swine behavior using accelerometers: Classification of behavior of nursery pigs from acceleration patterns

Abstract Text:

Behavioral monitoring is essential for animal welfare research. This study was to validate automated monitoring of swine behavior using accelerometers. Twelve pens of 8 pigs weaned at 4 wk were used. In each pen, 4 focal pigs were selected randomly for behavioral monitoring which occurred at 5 wk and 7 wk of age. Each focal pig was fitted with a digital accelerometer (Onset Pendant G. Data Logger) on the rear leg, which recorded instant acceleration in 3-dimensions at 10 sec intervals for 24 h. During the same period, behaviors of focal pigs were recorded continuously using digital cameras. Video-recordings were viewed to register four postural behaviors (lying on right side, lying on left side, lying sternally, and active behaviors consisting of standing, walking, eating, and drinking) that the pigs were performing continuously for 2-min. A total of 1,276 two-minute observation sets were collected from the video-recordings, including 618 sets for active behaviors, 330 for lying sternally, 182 for lying on left side, and 146 for lying on right side. Acceleration data series corresponding to each behavior were extracted, resulting in a total of 16,579 individual extracts taken from the accelerometers every 10 seconds. The Discriminant Analysis Procedure of SAS (Enterprise Miner, Version 12.3) was used to predict each behavior using the acceleration data. Data were randomly divided into training and evaluation datasets, each containing approximately half the individual extracts for each behavior. The variables used to classify behavior were the instantaneous accelerations in the x, y, and z directions, the magnitude of the acceleration vector, and the angle tilts for the x, y, and z components. On average, 83% individual extract and 86% series of 2-min were classified correctly for the four behaviors (Table 1). The correct classification was the highest for active behaviors and the lowest for lying on right. The classification results suggest excellent ability for the classification model to distinguish between four postural behaviors, but difficulty distinguishing between active behaviors like standing, walking, drinking, and eating in nursery pigs.

Table 1. Correct classification (%) for postural behaviors in nursery pigs.

Behavior	Series of 2-min	Individual Extract
Active Behaviors	97.1	91.2
Lying sternally	86.1	83.9
Lying on left side	89.1	85.5
Lying on right side	71.2	69.8
Average	85.8	82.6

Keywords: Accelerometer, behavior, pigs.