A spatiotemporal attention network-based analysis of golden pompano school feeding behavior in an aquaculture vessel

Abstract

Behavior analysis and recognition of fish school have substantial management value and optimization implications for many aquaculture activities, such as feeding, which is currently the most expensive and polluting element of the aquaculture process. However, fish school behaviors analysis for complex marine environments is an extremely challenging task. Recent developments in artificial intelligence have demonstrated that Deep-Learning techniques can be ideally adapted for behavior analysis. Seeing such prospects, a spatiotemporal attention network (STAN) was proposed in this study to analyse the feeding states and behaviors of golden pompano school. Specifically, spatial images and optical flow images were created from videos using image processing techniques and the Lucas–Kanade algorithm. STAN was then used to extract intuitive and perceptual features from the intuitive channel and perceptual channel, respectively. Finally, the states of Feeding or Non-Feeding of the golden pompano school were established by the fusing channel with LSTM and Fully–connected networks. To evaluate the performance of the proposed method, quantitative and qualitative experiments were conducted. Results showed that STAN outperformed other models with a test accuracy of 97.97%. Further validation on a genuine golden pompano farming vessel was implemented, showing that the STAN architecture delivers state-of-the-art accuracy in the task of analysing feeding behavior for golden pompano school in aquaculture.