A Two-Stage Deep Learning Method for Non-Invasive Sow Body Temperature Prediction Fusing Thermal Imaging and...

Area of Science:

• Animal Science
• Artificial Intelligence
• Biomedical Engineering

Background:

• Traditional rectal temperature measurement in pigs is stressful, labor-intensive, and risks cross-infection.
• Accurate, non-invasive temperature monitoring is crucial for swine health management.

Purpose of the Study:

• To develop and validate a non-invasive deep learning approach for predicting porcine rectal temperature.
• To combine infrared thermal imaging of thermal windows with environmental parameters for enhanced accuracy.

Main Methods:

• Utilized Mask R-CNN and YOLOv8s/YOLOv11s for thermal window detection/segmentation.
• Developed a two-stage hybrid regression framework with CNN, FCNN, Transformer, and Random Forest.
• Collected multimodal data including thermal images, environmental parameters, and rectal temperatures.

Main Results:

• The "eye + vulva" dual-region scheme achieved the best performance (MAE: 0.1796 °C, R²: 0.8212).
• The proposed multimodal deep learning model reduced prediction error by 42.3% compared to unimodal models.
• Demonstrated the effectiveness of combining spatial features, surface temperatures, environmental data, and cross-modal dependencies.

Conclusions:

• The non-invasive deep learning method offers a fast, accurate, and stress-free solution for porcine temperature monitoring.
• This approach supports intelligent health management in livestock farming.
• The "eye + vulva" thermal window combination is optimal for non-invasive temperature prediction in pigs.