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

Accurate quantification of fish feeding intensity is critical for optimizing feeding strategies and reducing feed waste in industrial recirculating aquaculture systems (RAS). However, real-world aquaculture environments present significant challenges, including high-density fish populations, water surface disturbances, and dynamic behavioral variations. To address these issues, this study proposes a hybrid vision-based framework (HVIT) for robust feeding intensity analysis. The proposed method integrates a Convolutional Neural Network (CNN) for local feature extraction and a Vision Transformer (ViT) for global context modeling within a parallel architecture, enabling effective representation of complex group behaviors. Furthermore, a Long Short-Term Memory (LSTM) module is incorporated to capture temporal dynamics of feeding activity, allowing continuous characterization of feeding intensity over time. A dedicated dataset of largemouth bass (Micropterus salmoides) under industrial RAS conditions was constructed, with data augmentation strategies applied to improve robustness against environmental noise and visual disturbances. Experimental results demonstrate that the proposed framework achieves over 98% accuracy across four feeding intensity levels and outperforms conventional CNN-based approaches. More importantly, the proposed method enables quantitative evaluation of feeding activity, providing a practical basis for real-time feeding decision support and intelligent feeding system development in large-scale aquaculture.