Automated identification of chicken distress vocalizations using deep learning models

Area of Science:

• Animal Welfare Science
• Machine Learning Applications
• Bioacoustics

Background:

• Global chicken production involves billions of birds housed in large groups, raising welfare concerns.
• Chicken distress calls are key welfare indicators, but manual identification is labor-intensive.
• Automated systems are needed to efficiently monitor welfare in intensive farming.

Purpose of the Study:

• To develop and evaluate a novel AI model for automatic detection of chicken distress calls.
• To assess the performance of the light-VGG11 model compared to existing methods.
• To explore data augmentation techniques to further enhance detection accuracy.

Main Methods:

• A convolutional neural network (CNN) model, light-VGG11, was developed using farm audio recordings.
• The model was trained on 3363 distress calls and 1973 natural barn sounds.
• Data augmentation techniques including time masking, frequency masking, and Gaussian noise were applied.

Main Results:

• The light-VGG11 model achieved high performance: 94.58% precision, 94.89% recall, 94.73% F1-score, and 95.07% accuracy.
• It demonstrated a 55.88% faster detection speed with significantly fewer parameters than VGG11.
• Data augmentation improved distress call detection by up to 1.52%.

Conclusions:

• The light-VGG11 model offers an efficient and accurate method for automated chicken distress call detection.
• This technology has significant potential for real-time welfare monitoring in commercial chicken flocks.
• Automated bioacoustic analysis can enhance animal welfare assessment in large-scale agriculture.