Decoding Sheep Actions: Deep Learning Algorithms for In-Depth Analysis of Behavioral Reactivity

Abstract

This study presents a motion-based framework for classifying sheep behavior using deep learning algorithms and dense optical flow. To evaluate behavioral reactivity related to fear in sheep, a standard method known as a scale test was applied to 116 sheep. Behavioral data were clustered using K-means and hierarchical clustering algorithms on three datasets containing manually scored behavioral features, a five-point scoring system, an index score, and a new motion index derived from dense optical flow. The accuracy and reliability of the obtained clusters were evaluated using Support Vector Machines (SVM). The labeled video sequences were preprocessed using dense optical flow and used to train three different deep learning architectures: CNN, ConvLSTM, and VGG19. Among these, the ConvLSTM model trained on the motion index-based dataset demonstrated the highest performance, correctly classifying active sheeps with 92% accuracy and achieving 83% overall classification accuracy. The CNN model showed similar results, while VGG19 performed lower with 71% accuracy. These findings suggest that the integration of motion indices with deep learning architectures offers a promising alternative to manual behavior scoring. The proposed framework enhances classification reliability and highlights the potential of preprocessing techniques such as dense optical flow to improve model performance in livestock behavior analysis.