A Deep Transformer Model-Based Analysis of Fish School Starvation Degree in Marine Farming Vessels

Studying the physiological behavior of fish school starvation and dynamically adjusting the feeding scheme based on the starvation features of fish school can help promote the unmanned and intelligent process of offshore aquaculture. The current starvation algorithms trained based on breeding data from terrestrial pools and ponds are largely limited by the differences in environmental conditions and are therefore difficult to be applied to real marine aquaculture environments. In this paper, a deep attention model based on a dual-stream network (DADSN) was used to analyze the starvation behaviors of fish school and to grade them into five starvation levels. First, we collected the golden pompano aquaculture videos on a farming vessel and extracted optical flow images from the original videos to create a dual-stream dataset for golden pompano starvation analysis. Next, considering the characteristics of camera imaging in the marine environment, we used DADSN to extract fish behavioral features in the spatial domain and the optical domain, respectively, and fused the heterogeneous features by LSTM to enrich behavioral information without additional equipment. Finally, we conducted detailed control tests and qualitative and quantitative analyses, and the model obtained an accuracy of 83.43%, which is significantly higher than other mainstream models. Further tests were conducted on farmed workboats to confirm that the model can be applied to the vessel aquaculture environment of golden pompano.