Identification of estrus cows through their vocalizations using Ensemble Machine Learning

Ensemble Machine Learning for estrus detection

Abstract

Identification of estrus phase is important for effective reproductive management and to improve productivity of dairy cattle. Conventional estrus detection methods, including visual observation and activity-based monitoring, may require substantial labor and can miss estrus events, motivating automated and non-invasive approaches. This study proposes an Ensemble Machine Learning model that investigates cattle vocalizations as an acoustic indicator of estrus. Audio preprocessing included pre-emphasis filtering and data augmentation, followed by extraction of 40-dimensional Mel-frequency cepstral coefficient (MFCC) features. Convolutional neural network architectures, LeNet-5, VGG-16, and ResNet-50, were evaluated individually and integrated using a weighted prediction-averaging strategy. The proposed ensemble achieved an overall classification accuracy of 96.03%, with precision, recall, and F1-score of 97.3%, 95.5%, and 96.5%, respectively, for the estrus class. The results indicate that cattle vocalizations contain discriminative acoustic features connected with reproductive phase and that combining complementary deep-learning architectures can improve classification performance. Hence, the proposed approach provides a non-invasive basis for automated estrus monitoring and may support precision livestock farming.