DCNN for Pig Vocalization and Non-Vocalization Classification: Evaluate Model Robustness with New Data.

Area of Science:

• Agricultural Technology
• Machine Learning
• Animal Science

Background:

• Pig vocalization is key for monitoring livestock health and welfare.
• Collecting sufficient pig sound data for deep learning is challenging and time-consuming.

Purpose of the Study:

• To develop an effective deep learning model for pig vocalization and non-vocalization classification.
• To introduce a novel audio feature extraction method to enhance classification accuracy.

Main Methods:

• A deep convolutional neural network (DCNN) was employed for classification.
• Evaluated Mel-frequency cepstral coefficients (MFCC), Mel-spectrogram, Chroma, and Tonnetz features.
• Proposed and integrated a novel Mixed-MMCT feature extraction method.
• Utilized audio data augmentation techniques and k-fold cross-validation (k=5).

Main Results:

• The Mixed-MMCT method achieved superior classification accuracy, reaching up to 99.67% on farm datasets.
• Robustness experiments demonstrated an average performance of 95.67% accuracy, 96.25% precision, 95.68% recall, and 95.96% F1-score.
• The proposed method outperformed existing feature extraction techniques.

Conclusions:

• The Mixed-MMCT feature extraction method is highly effective for pig vocalization classification in real farming conditions.
• This approach offers a promising solution for improving pig welfare and farm management through advanced audio analysis.