3 Towards Automated Behavior Analysis for Individually Housed Piglets–Sampling Considerations

Abstract

Abstract Behavior sampling is commonly applied to represent a complete picture of animal behavior. Traditional sampling methods often fail to capture complete sequences of events, diurnal behavior, or behavior patterns. As digital technologies allow more frequent collection of behavior, there is a need to understand if the potential increase in sampling frequency yields an increase in the knowledge we gain from additional observations. In this study, 24-h behavior was labelled for three individually housed research piglets, creating a ground truth dataset. An ethogram with five categories (resting, consumption, maintenance, exploratory, and social interactions) including 23 behavior labels was applied for behavior observation for every frame in the video data set. Using Python to automatically sample behavior data from the complete dataset, the results of instantaneous and continuous sampling strategies at various sampling intervals and times of day were compared with the ground truth continuous data set using a two-sample t-test for: (a) total percentage of observations, (b) average bout length, and (c) normalized number of occurrences. Results were considered significant for alpha< 0.05, which indicated a sampling method was different from ground truth. Continuously sampled data were collected at the start of each hour for 2, 5, 10, 15, and 20-min periods. Instantaneously sampled data were collected at the intervals: 1, 5, 10, 20, and 40-s; 1, 2, 5, 10, 20, and 60-min. Every sampling strategy resulted in misrepresentation of behavior. For example, results for the 1-s instantaneous sampling method (every 15th frame versus every frame) revealed that seven of the 23 behaviors observed were misrepresented in at least one metric: (a) total percentage of observations [P=0.027 (drinking behavior); P=0.027 (sham chewing), P=0.011 (head tossing)]; (b) average bout length [P=0.033 (feeding behavior), P=0.032 (nosing), P=0.040 (exploring pen), P=0.039 (head tossing)]; and (c) normalized number of occurrences [P=0.005 (exploring pen), P=0.018 (head tossing)].