Fishing Ground Prediction Of Neon Flying Squid(Ommastrephes Bartramii)in The North Pacific Ocean Based On Deep Learning

The neon flying squid,Ommastrephes bartramii,is one of the essential economic cephalopods in the Northwest Pacific Ocean.It has become one of the leading fishing targets for Chinese squid jigging fleets.Improving the accuracy of fishing ground prediction for oceanic economic species has been one of the most concerned issues in fishery research.In the era of big data of marine remote sensing and fishery,it has been difficult to mine valuable information from the complex and large amount of data using traditional fishing prediction methods.And it has been difficult to make full use of spatiotemporal correlation to reduce model complexity and improve model robustness.Thus,the task can not be completed with high efficiency and accuracy.Contrary to traditional methods,deep learning,as a powerful technology emerging in the field of artificial intelligence in recent years,has achieved better results especially in the problem of image big data processing,which has development prospects and exploration value.Therefore,this study used the fishing data of O.bartramii from 1998 to 2020,combined with marine remote sensing environmental factors(such as sea surface temperature,sea surface height,El Ni?o-Southern Oscillation index,etc.),used the improved U-Net model in deep learning to study the following items:(1)designed the cases of combining different spatial and temporal scales with multiple environmental factors;(2)analyzed the temporal correlations of environmental factors of different lengths;(3)fishing ground prediction of real time and future time with single and multiple environmental factors;(4)explored the optimal spatial and temporal scale criteria.This study lays the scientific foundation for the development of AI-based fishing ground research and provide technical support for the scientific and sustainable development and scientific management of fishery production enterprises in the North Pacific Ocean.The results of this study could be summarized as follows:(1)Exploration of optimal temporal and spatial scales for real time fishing ground prediction.The deep learning-based fishing ground prediction model takes the sea surface temperature as the input and the fishing ground as the output.The model is trained with the data from July to November in 1998--2019 and tested with the data of2020.We consider and compare 20 cases with a combination of four spatial scales(0.05°,0.1°,0.25°,and 0.5°)and five temporal scales(3,6,10,15,and 30 days).By comparing different cases,we found that the optimal temporal and spatial scales of the deep learning-based fishing ground prediction model exist.The optimal temporal scale is 15 days,and the optimal spatial scale is 0.25°.The study concluded that the deep learning-based fishing ground prediction model requires both the quantity and quality of environmental data and fishery data.The reasons for the influence of different spatial and temporal scales on the model results may be the spatial distribution characteristics of fishery data and SST data,and the corresponding sensitivity of fishery data to SST data.The results provide criteria for future AI-based fisheries research and deepen the understanding of fisheries prediction from the perspective of artificial intelligence.(2)Prediction of fishing ground for future time.The deep learning-based fishing ground prediction model takes the sea surface temperature as the input and the fishing ground as the output.The model is trained with the data from July to November in1998--2019 and tested with the data of 2020.We constructed 28 cases with different temporal scales and lead periods.By comparing different cases,we found that the larger the temporal scale,the more stable the model performance and the higher the accuracy.The optimal temporal scale of the deep learning-based fishing ground prediction model for future time is 15 days,the optimal lead periods are 4,and the best lead fishing date is May 1.The fluctuation information between different lead periods of the environmental field SST and the coupling degree of the SST range changes may be important factors affecting the model performance differences.Therefore,it is necessary to combine the characteristic information of environmental data with the actual operational production needs to select the optimal lead periods.(3)Prediction of fishing ground with multiple environmental factors.Based on the modified U-Net model,the deep learning-based fishing ground prediction model takes the sea surface temperature,sea surface height,sea surface salinity and Chlorophyll a as the inputs and the center fishing ground as the output.The model is trained with the data from July to November in 2002--2019 and tested with the data of 2020.We consider and compare five temporal scales(3,6,10,15,and 30 days)and seven multiple environmental factor combinations.By comparing different cases,we found that the optimal temporal scale is 30 days,and the optimal multiple environmental factor combination contained SST and Chl a.We analyzed the application effect index of fishing ground,which greatly reduced the area of fishing ground and improved the concentration of fishing ground while ensuring the high yield coverage.The inclusion of multiple factors in the model greatly improved the concentration of center fishing ground.And it makes the fishery prediction results more accurate.This result shows that the importance of each factor varies in different time periods.The selection of a suitable combination of multiple environmental factors is beneficial to the precise spatial distribution of fishing ground.