Measuring Comfort Behaviours in Laying Hens Using Deep-Learning Tools.

Area of Science:

• Animal Welfare Science
• Computer Vision
• Machine Learning

Background:

• Automated monitoring of animal behavior is crucial for assessing welfare.
• Image analysis with machine learning (ML) offers potential for objective behavioral measurement.
• Precision Livestock Farming (PLF) techniques can enhance animal husbandry.

Purpose of the Study:

• To evaluate a machine learning tool for counting hens on the ground.
• To assess the capability of ML in identifying dust-bathing behavior in hens.
• To compare the performance of two YOLO (You Only Look Once) models for these tasks.

Main Methods:

• Utilized image analysis with a PLF technique in an experimental aviary setting.
• Trained and validated two YOLO models (YOLOv4-tiny and YOLOv4) for hen behavior classification.
• Evaluated model performance using precision, recall, harmonic mean, mean average precision (mAP), and frames per second (FPS).

Main Results:

• Both YOLO models achieved high accuracy (94% mAP) in classifying hens on the floor.
• The YOLOv4-tiny model trained significantly faster (4.26 h) than YOLOv4 (23.2 h).
• Classification performance for dust-bathing hens was poor for both models (28.2%–31.6% mAP), indicating limitations.

Conclusions:

• Machine learning effectively identifies laying hens on the floor, aiding in welfare assessment.
• Current ML models struggle to accurately classify dust-bathing behavior.
• Further research and development of PLF tools are needed for comprehensive behavioral monitoring.