Study on the influence of pre-partum activities pattern of sows on piglet survival using deep learning and computer vision

Abstract

Over the years genetic selection for ovulation rate substantially increased litter size. However, there was a concomitant increase in young piglet mortality. Pre-weaning mortality can range from 5-20%, with 50-80% of these deaths occurring during the first week of life, with the first 72 hours after farrowing being the most critical period for piglet survival. There is evidence that the sow activity patterns in the pre-partum period (postures, postural changes and nest building behavior) can predict poor maternal behaviors during farrowing and early lactation. Traditionally, behavior analysis is performed through human observation, in a short period of time and with a limited number of animals, making it laborious, subjective and susceptible to human error. The use of computer vision presents as an alternative to overcome these challenges and more robustly analyze the impact of maternal behaviors on piglet survival and development. Based on the foregoing, the objective of this project is to develop an automated computer vision monitoring system using deep learning algorithms to predict sow activity patterns in the peripartum period, namely those associated with nest building behavior, and estimate how these influence maternal behaviors and piglet's performance during early lactation. A total of 56 sows will be used. The experiment will be carried out in two periods: summer (n = 28) and winter (n = 28). In each period, the sows will be moved to the farrowing unit seven days before the expected date of farrowing and assigned according to parity in two groups: sows with no provision of nesting substrates (n = 14) and sows provided with nesting substrates (n = 14). The sows will be housed in standard farrowing crates equipped with video recording cameras. The sows will be filmed 24h a day, from two days before farrowing until five days after farrowing. The total number of piglets born alive, stillborn, farrowing duration, inter-piglet birth interval, interval from birth to first suckle, piglet mortality rate and the cause of piglet deaths until day 5 after birth will be recorded. The behavioral measurements will be conducted using a pre-define ethogram. The video images will be manually labelled to compose an image database for training models based on convolutional neural network (CNN) for posture detection using the SLEAP software (Social LEAP Estimates Animal Poses). Different RNC architectures will be used to construct different models for posture classification. The models will be compared using the confusion matrix metrics (accuracy, precision and sensitivity) to obtain the best computational model for the vision system.