Spatiotemporal Modeling and Intelligent Recognition of Sow Estrus Behavior for Precision Livestock Farming

Area of Science:

• Animal Science
• Computer Science
• Machine Learning

Background:

• Accurate estrus behavior recognition in sows is crucial for efficient pig farm breeding management.
• Traditional methods are inefficient and prone to errors due to the complex nature of estrus behaviors.

Purpose of the Study:

• To develop and compare deep learning models for high-precision recognition of four key sow estrus behaviors (SOB, SOC, SOS, SOW).
• To address the limitations of traditional methods in estrus behavior identification.

Main Methods:

• Video-based behavior recognition using three deep learning models: CNN + LSTM, 3D-CNN, and CNN + TCN.
• A sliding window strategy was used for data processing, creating uniform-length image sequence samples.
• Systematic comparative analysis of model performance using metrics like accuracy, F1-score, and AUC curves.

Main Results:

• The CNN + TCN model demonstrated superior performance with validation accuracy >0.98 and F1-score near 1.0.
• 3D-CNN excelled at recognizing short-term behaviors (e.g., SOC) with an F1-score of 0.91.
• CNN + LSTM showed robustness in identifying long-duration behaviors (e.g., SOB, SOS) with accuracy of 0.99.

Conclusions:

• Deep learning models, particularly CNN + TCN, offer a highly accurate solution for sow estrus behavior recognition.
• The developed intelligent recognition system supports digitalization and intelligent management in large-scale pig farms.
• Different models are suitable for specific behavior recognition tasks, enabling tailored applications in swine reproduction.