Classification of multiple cattle behavior patterns using a recurrent neural network with long short-term memory and inertial measurement units

Abstract

In this study, we aimed to develop a recurrent neural network (RNN) with a long short-time memory (LSTM) model to monitor and classify cattle behavior patterns using inertial measurement units (IMU). This model was trained using motion data obtained from 6 Japanese steers. Each steer was fitted with an IMU sensor inside a waterproof box attached to a collar on the neck. Classified behavior classes included feeding, lying, ruminating (while lying), ruminating (while standing), licking salt, moving, social licking and head butt. LSTM-RNN model was trained to classify and measure cattle’s behavior across three window-sizes including window-size 64, 128 and 256 (3.2 s, 6.4 s and 12.8 s). A convolution neural network (CNN) model was used for comparison. The results reveal the LSTM-RNN model classification performance was superior to the CNN model. The LSTM-RNN model was found to achieve the best performance when using a window-size of 64 (accuracy, precision, recall, f1-score all were 88.7%). With a window-size 64, classification accuracy of specific behaviors was 97.8% (feeding), 88.7% (lying), 88.4% (ruminating-lying), 92.9% (ruminating-standing), 94.4% (licking salt), 84.8% (moving), 80.3% (social licking), and 81.9% (head butt). A few physically similar behaviors were easily misclassified. In conclusion, the LSTM-RNN demonstrated reasonable classification of cattle behavior. In future, additional sensors, such as a microphone, could be added to the cattle behavior monitoring system and behavior classification extended to cattle welfare and growth behaviors, such as feeding, reproduction and disease prediction.