Explainable classification of goat vocalizations using convolutional neural networks.

Area of Science:

• Animal Science
• Machine Learning
• Bioacoustics

Background:

• Precision livestock farming requires accurate animal and environmental data.
• Goat vocalizations offer insights into animal welfare and health.
• Automated analysis of animal sounds is crucial for efficient farm management.

Purpose of the Study:

• To develop and validate a Convolutional Neural Network (CNN) for classifying goat vocalizations.
• To enhance model robustness using data augmentation techniques.
• To interpret the CNN's decision-making process using explainable AI (XAI) for identifying critical acoustic features.

Main Methods:

• A CNN architecture was designed for goat vocalization classification.
• Dataset augmentation involved pitch shifting and time stretching.
• An explainability analysis (XAI) was performed to interpret model decisions.
• Performance was compared against contrasting approaches.

Main Results:

• The CNN achieved an average classification rate of 95.8% for discriminating goat emotional states.
• Data augmentation significantly boosted model robustness and classification accuracy.
• XAI identified specific time-frequency content crucial for accurate classification.
• The proposed model demonstrated superiority over alternative methods.

Conclusions:

• CNNs, enhanced by data augmentation and XAI, are effective for classifying goat vocalizations.
• XAI provides transparency and identifies key acoustic markers for animal welfare monitoring.
• The developed interactive scheme offers valuable insights to animal scientists for precision farming.