A multi-stage ensemble framework for classifying pig vocalizations under noisy animal farm environments

Area of Science:

• Animal Science
• Agricultural Engineering
• Bioacoustics

Background:

• Pig vocalizations are key indicators of animal health and emotional states.
• Precision livestock farming requires accurate analysis of these vocalizations.
• Existing methods struggle with noise and classifying diverse vocalizations.

Purpose of the Study:

• To develop a comprehensive framework for detecting and classifying a wide range of pig vocalizations.
• To assess pig health and emotional stress in real-world farm conditions.
• To improve the practical applicability of pig vocalization analysis.

Main Methods:

• Introduced the Pig Vocalization Multi-stage Classification (PVMC) model.
• Integrated cough/scream detection with emotional state classification.
• Employed an ensemble learning strategy combining Wav2Vec2 and Audio Spectrogram Transformer (AST) models.

Main Results:

• Achieved up to 4.9dB SNR improvement and 95.80% accuracy in vocalization segmentation.
• Reached 98.88% accuracy in key vocalization classification and 92.15% in emotional state detection.
• Ensemble learning significantly enhanced precision, recall, and F1-score.

Conclusions:

• The PVMC model demonstrates robustness and practical utility for real-time pig vocalization monitoring.
• PVMC contributes to intelligent, welfare-oriented livestock management systems.
• The model's multi-stage approach and ensemble learning enhance performance in diverse farm conditions.