Dam behavior patterns in Japanese black beef cattle prior to calving: Automated detection using LSTM-RNN

Abstract

This study develops a recurrent neural network (RNN) with a long short-time memory (LSTM) model to detect and recognize calving related behaviors using inertial measurement unit (IMU). The models were trained using IMU data collected from three expectant cows during the last three days before calving. Classified behavior pattern classes included feeding, ruminating (lying), ruminating (standing), lying normal (collected during 72 h–24 h before calving), standing normal (same as lying normal), lying final and standing final, which were defined as the lying and standing behavior that occurred during the last 24 h before calving. The LSTM-RNN models were trained to classify cow behavior classes across window-size of 32, 64, 128 and 256 respectively (1.6 s, 3.2 s, 6.4 s and 12.8 s). The best overall performing LSTM-RNN model had a window-size of 32 (accuracy, precision, recall, f1-score were 79.7%, 81.1%, 79.7% and 79.8%, respectively). With a window-size of 32, the model classification accuracy for specific behaviors was 76.0% (feeding), 92.6% (ruminating (lying)), 88.3% (ruminating (standing)), 63.2% (lying normal), 78.0% (standing normal), 74.7% (lying final) and 70.1% (standing final). These results demonstrate the potential of a LSTM-RNN model to automatically recognize behaviors patterns prior to birth. In the future, more related indicators will be added to improve the accuracy and robustness of this recognition model. With further work, statistically significant changes in behavior could be streamed to farmers informing them of the progress of calving and alerting them to critical changes in the situation.