Feeding intensity identification method for pond fish school using dual-label and MobileViT-SENet

Abstract

Accurately identifying the fish feeding intensity is crucial for timely understanding the feeding demand, dynamically adjusting the feeding strategy, and reducing costs. For the variability and uncontrollability of the pond aquaculture environment, the densities of feeding fish schools aggregating to the feeding point exhibit significant variations. Consequently, different densities of fish schools present inconsistent characteristics in the image, even under the same feeding intensity, making the precise identification of feeding intensity difficult. To tackle this issue, a method for identifying the feeding intensity of pond fish schools based on dual-label and MobileViT-SENet (DL-MobileViT-SENet) was proposed. The fish school images were marked with labels indicating density and feeding intensity to establish the dual-label dataset. Subsequently, a proposed MobileViT-SENet network is trained using the dataset to obtain the dual-label pre-training weight incorporating both fish density and feeding intensity features. Two models are trained to identify density and feeding intensity based on the obtained weight. Finally, a dynamic feeding strategy for fish that integrates biomass, density, and feeding intensity is presented. The proposed method combines the density and feeding intensity labels to enhance the accuracy of identifying the feeding intensity of pond fish schools across various densities, and lays the groundwork for designing a dynamic feeding strategy. It was tested on authentic pond fish school images and yielded an accuracy of 97.95%. This value is superior to these comparison methods, demonstrating that this method can accurately identify the feeding intensity of pond fish and provide support for formulating a dynamic feeding strategy.