Fish feeding behavior recognition model based on the fusion of visual and water quality features

Objective To improve the accuracy of fish feeding behavior recognition in industrial aquaculture environment.

Method A fish feeding behavior recognition model was proposed based on the fusion of visual and water quality features, namely MC-ConvNeXtV2. To better capture the global features of different aggregation levels and the detailed features of feeding behavior, a context-aware local attention mechanism (Cloatt) was introduced in each convolution stage of ConvNeXtV2-T. To improve the behavior recognition performance of the model in high-density aquaculture, a multimodal feature fusion module (MFFM) was designed to achieve adaptive fusion of visual features and dissolve oxygen, temperature, and pH of water quality features. The model test was conducted in a Micropterus salmoides culture factory with a culture density of 160 fish/m³.

Result The test results showed that for the task of four feeding behaviors classification of fish school, the recognition accuracy, precision and recall of MC-ConvNeXtV2 model were 96.89%, 96.34%, and 96.59%, respectively. Compared with ConvNeXtV2-T, these indicators increased by 3.11, 2.42, and 2.72 percentage points, respectively.

Conclusion The proposed fish feeding behavior recognition model offers a new approach for intelligent aquaculture management.