Classification of Cattle Behavior Leveraging Accelerometer Data and Machine Learning

In recent years, with the increasing world population, the utilization of developing technologies in order to meet the sustainability needs in the livestock sector and to increase the efficiency in animal products and to ensure animal welfare has become an important field of study. In this study, in order to contribute to the improvement of livestock operational processes, animal behaviors were classified by using accelerometer sensor data obtained from cattle movements and machine learning methods. The accelerometer data collected from six cows (black Japanese cattle) in thirteen different behavior categories were analyzed. Different classification algorithms unlike the previous studies conducted with this data set were tried, and model performances were improved with the use of fewer features compared to previous ones. Statistical and entropy-based features were extracted from the accelerometer data and multi-class classification was performed to detect thirteen different behaviors using Extreme Gradient Boosting (XGBoost), Extra Trees (ET), Light Gradient Boosting Machine (LightGBM), Multilayer Perceptron (MLP) and Support Vector Machine (SVM) classifiers. When the classification performances were evaluated, it was seen that the ET Classifier model provided the best performance with an accuracy rate of 97.48%. The obtained result is above the highest accuracy rate obtained in the studies used this dataset in the literature and obtained using fewer features. This performance is also at a high level that can be compared with other studies in this fields. In conclusion, this study offers great potential for monitoring and management of animal behavior, and especially the determination of drinking and feeding behaviors with high accuracy will provide significant benefits in terms of monitoring and management of cattle health.