Multi-Sensor Integration and Machine Learning for High-Resolution Classification of Herbivore Foraging Behavior.

Area of Science:

• Animal behavior science
• Machine learning applications in agriculture
• Livestock monitoring technologies

Background:

• Accurate classification of cattle behavior is crucial for welfare and productivity.
• Machine learning offers potential for automated analysis of complex animal movement data.
• Distinguishing subtle behaviors like grazing, resting, and ruminating requires sophisticated models.

Purpose of the Study:

• To evaluate and compare the performance of various machine learning models for classifying diverse cattle behaviors.
• To assess the effectiveness of random test split (RTS) and cross-validation (CV) data partitioning methods.
• To identify key movement-derived predictors influencing different behavioral states.

Main Methods:

• Applied six machine learning models: Perceptron, Logistic Regression, Support Vector Machine, K-Nearest Neighbors, Random Forest (RF), and XGBoost (XGB).
• Classified activity states, foraging behaviors (grazing, resting, walking, ruminating), posture states, and combined behaviors-posture states.
• Utilized random test split (RTS) and cross-validation (CV) for model evaluation.

Main Results:

• XGBoost showed highest accuracy for overall state and foraging behavior classification (74.5% RTS, 69.4% CV).
• Random Forest (RF) outperformed XGBoost in classifying specific behaviors like grazing, resting, ruminating, and posture states (e.g., 83.9% CV for posture).
• Movement data features like speed and Actindex were critical predictors, with specific sensor values (X, Z) influencing posture-related behaviors.

Conclusions:

• Cross-validation (CV) is a reliable method for evaluating machine learning models on complex cattle behavioral data, especially with Random Forest.
• Continuous recording devices and detailed movement data are essential for accurate cattle behavior monitoring.
• Machine learning, particularly RF, demonstrates significant potential for automated and precise analysis of livestock activities.