Evaluating machine learning classifiers and explainability for monitoring cow behaviour with wearable nose rings

Abstract

Wearable technologies are revolutionizing precision livestock monitoring by allowing continuous real-time monitoring of animal behaviour. This study investigates and evaluates the use of machine learning techniques to classify dairy cow behaviours using tri-axial accelerometer data collected from novel wearable nose ring sensor. The raw dataset initially included five distinct behaviours: Feeding, Ruminating, Standing, Lying and Walking. However due to data imbalance and data limitations we refined the classification to three core categories: Feeding, Rumination and Walking. While previous studies on this dataset focused solely on Long Short-Term Memory(LSTM) network, the comparative potential of other models remained unexplored. To address this gap, we performed a comparative study on multiple classifiers, including Random Forest (RF), Artificial Neural Network (ANN), Gated Recurrent Unit (GRU) and a hybrid Convolutional Neural Network with LSTM (CNN-LSTM). The obtained results showed that GRU model performed well with an accuracy of 97.78 %, followed by CNN-LSTM, ANN and RF which scored 97.78 %, 68.27 % and 67.6 % respectively. To enhance model transparency, Explainable AI techniques were utilized. SHAP and LIME were utilized to showcase feature importance and interpretability of these models. These findings showcase the effectiveness of deep learning models (GRU, CNN-LSTM) and emphasizes the importance of model explainability in precision livestock management.