Lightweight Fine-Tuning for Pig Cough Detection.

Area of Science:

• Agricultural Science
• Animal Health
• Machine Learning

Background:

• Respiratory diseases are a major concern in intensive pig farming, impacting animal welfare and productivity.
• Early detection of pig coughs is crucial for timely intervention but is hindered by limited labeled data and challenging farm acoustics.

Purpose of the Study:

• To develop a lightweight and accurate pig cough recognition method for early disease detection in resource-limited agricultural settings.
• To address the challenges of small sample sizes and complex acoustic environments in pig farming.

Main Methods:

• Utilized a pre-trained audio neural network, freezing its backbone and fine-tuning the classifier for knowledge transfer and domain adaptation.
• Incorporated a time-frequency dual-stream module to enhance the capture of cough-specific temporal-spectral features.
• Evaluated the method on a dataset of pig coughs and environmental noise clips.

Main Results:

• Achieved 94.59% accuracy and 92.86% F1-score on the test dataset, outperforming baseline models.
• Cross-validation demonstrated a mean accuracy of 96.99%, indicating robust generalization.
• The proposed lightweight fine-tuning approach proved effective for small-sample audio recognition in agricultural contexts.

Conclusions:

• The developed framework offers a reliable technical solution for early warning of respiratory diseases in pig farms through accurate cough recognition.
• Transfer learning presents a viable strategy for small-sample audio recognition in resource-constrained agricultural environments.
• The study highlights the potential of AI for improving animal health monitoring and management in intensive farming.