Machine Learning Techniques for Classification of Livestock Behavior

Animal activity recognition is in the interest of agricultural community, animal behaviorists, and conservationists since it acts as an indicator of the animal's health in addition to their nutrition intake when the observation is performed during the circadian circle. Machine learning techniques and tools are used to help identify the activities of livestock. These techniques are helpful to discriminate between complex patterns for classifying animal behaviors during the day; human observation alone is labor intensive and time consuming. This research proposes a robust machine learning method to classify five activities of livestock. To prove the concept, a dataset was utilized based on the observation of two sheep and four goats. A feature selection technique, namely Boruta, was tested with multilayer perceptron, random forests, extreme gradient boosting, and k-Nearest neighbors algorithms. The best results were obtained with random forests achieving accuracy of 96.47% and kappa value of 95.41%. The results showed that the method can classify grazing, lying, scratching or biting, standing, and walking with high sensitivity and specificity.