Classifying Calf Behavior from Data of the Nedap SmartTag Ear with Self-Supervised Learning

Thumbnail Image
Issue Date
Journal Title
Journal ISSN
Volume Title
Calves are the future of a farm. They have to grow up to become dairy cows to make a profit for the farmer. Unfortunately, health issues in calves can have a long-term effect on the calf’s welfare, development, and future performance. To detect health issues in adult cattle, artificial intelligence has been widely used, however, only a few sensors are available to be used for calves. Therefore, this research will focus on behavior classification in calves, based on data from the Nedap SmartTag Ear. Three different classes are recognized; Lying to predict diseases in calves, Ruminating to predict the ideal weaning period, and Other for the remaining behaviors. A convolutional neural network has been developed with fully-supervised learning to recognize the three classes, reaching an overall F1 score of 0.826 ± 0.073. Because fully-supervised learning needs a lot of labeled data and only limited labeled data is available, self-supervised learning (SSL) is used to incorporate unlabeled data in the training of a model. The pretext task used for SSL is transformation recognition with six transformations. Four SSL approaches have been compared and the approach with a single-task learning classifier to recognize the transformations and two-step training with a lambda value of 0.5 reaches the highest performance, with an overall F1 score of 0.817 ± 0.057, which is slightly worse than the fully-supervised model. Both model types reach the highest F1 score for the Lying class and the lowest F1 score for the Ruminating class. Further experimentation shows that with only 1% of the labeled data, the SSL model was still able to reach an overall F1 score of around 0.8, while the performance of the fully-supervised model only reached approximately 0.6. This shows that there is potential in SSL, however, in the current situation SSL is not able to outperform fully-supervised learning. Additionally, the models have been developed based on 2 Hz data, while previous experiments are with 8 Hz data. This results in a performance drop of 0.036 for the fully-supervised model and 0.107 for the SSL model. Overall, the results seem promising, but further research on handling the data diversity between calves and increasing the performance of the SSL model is required before the model can be used on farms.
Faculteit der Sociale Wetenschappen