Automated oestrous detection in sows using a robotic imaging system

Accurate oestrous detection is critical to optimise sows' reproductive performance. The conventional method of oestrous detection relies on the laborious back-pressure test.

Objective: This study presents an automated oestrous detection method for sows housed in individual stalls using a robotic imaging system and neural networks.

Methods: A robotic imaging system consisting of a LiDAR camera was used to monitor a group of stall-housed sows at a 10-min interval to capture their postures and vulva volume. Imagery data were analysed using a previously developed pipeline.

Results: Results showed that significant changes were observed in daily standing index, sternal lying index, lateral lying index, posture change frequency, and vulva volume before the onset of oestrous. A 1-D convolutional neural network model architecture for oestrous detection was developed using Days from Weaning, behaviour features, and vulva volume features as inputs. The oestrous detection models were evaluated using 10-fold cross validation. The training and testing accuracies of the oestrous detection model were 96.1 ± 2.0% and 92.3 ± 10.1% when using the Days from Weaning and behaviour features as input. The model's training and testing accuracies increased to 98.1 ± 2.4% and 98.0 ± 4.2% when vulva volume features were added to the input.

Conclusion: While it is difficult to trace the behaviour of sows housed in group conditions, combining vulva volume features with Days from Weaning could be a suitable method to detect the onset of oestrous in these sows. The training and testing accuracies of this method of oestrous detection were 97.9 ± 1.4% and 95.2 ± 4.8%. However, further validation under real group house conditions is needed.