A vision-based framework for quantifying fish feeding behavior in industrial recirculating aquaculture systems.

Area of Science:

• Aquaculture technology
• Computer vision
• Machine learning

Background:

• Accurate fish feeding intensity quantification is vital for optimizing feed strategies and minimizing waste in industrial recirculating aquaculture systems (RAS).
• Real-world RAS face challenges like high fish density, water surface disturbances, and variable behaviors, complicating feeding analysis.
• Existing methods struggle with the dynamic and complex visual environment of industrial aquaculture.

Purpose of the Study:

• To develop a robust hybrid vision-based framework (HVIT) for accurate fish feeding intensity analysis in industrial RAS.
• To integrate Convolutional Neural Network (CNN) and Vision Transformer (ViT) for comprehensive feature extraction and context modeling.
• To incorporate Long Short-Term Memory (LSTM) for capturing temporal feeding dynamics and enabling continuous intensity characterization.

Main Methods:

• A hybrid vision-based framework (HVIT) combining CNN for local features and ViT for global context in parallel.
• Integration of a Long Short-Term Memory (LSTM) module to analyze temporal feeding activity.
• Development of a dedicated largemouth bass dataset with data augmentation for enhanced robustness.

Main Results:

• The proposed HVIT framework achieved over 98% accuracy in classifying four feeding intensity levels.
• The hybrid approach effectively captured complex group behaviors and temporal feeding dynamics.
• HVIT demonstrated superior performance compared to conventional CNN-based methods.

Conclusions:

• The HVIT framework offers a robust and accurate solution for quantifying fish feeding intensity in challenging RAS environments.
• This quantitative evaluation provides a practical foundation for real-time feeding decision support.
• The study paves the way for intelligent feeding system development in large-scale aquaculture.