Research On Fishery Forecast Model Of South Pacific Albacore Tuna Based On Feature Interaction And Deep Learning

Albacore tuna,as a highly migratory pelagic fish,is widely distributed in the South Pacific region.Because of its high yield and rich nutritional value,albacore tuna has become one of the main fishing objects of tuna longline in the South Pacific.How to make a reasonable and effective forecast of albacore tuna fishery in the South Pacific has become one of the research hotspots in the field of pelagic fishing.Due to the simple model structure and small number of learning parameters,the traditional fishing ground prediction method often has poor fitting effect in the face of large scale,complex and multidimensional ocean big data,so its prediction accuracy is low and its guidance for fishery operation is limited.However,deep learning can dig out important semantic features in complex dynamic scenes,fit high-dimensional complex data,and extract implicit high-order feature information.At the same time,feature interaction technology can learn interactive information between features in an explicit and controllable way,which can form a complementary relationship with deep learning.Therefore,it is of great significance to combine deep learning with feature interaction technology to put forward a model suitable for the field of fishery prediction and further improve the accuracy of fishery prediction.By studying a large number of literature related to deep learning and feature interaction and combining its successful application in various fields,this paper proposed two different fishery prediction models based on feature interaction and deep learning:(1)In the first model,CNN-CROSS,a fishery prediction model based on feature interaction and convolution network,was proposed based on the production data of alba tuna longline fishery in the South Pacific from 2000 to 2015,combined with the data of three environmental factors and three spatio-temporal factors.The model introduces Embedding layer to process data,which solves the problem of feature sparsity caused by one-hot coding and the influence of manual feature engineering on results.At the same time,Cross network is introduced to extract interactive information between features to eliminate the problem of insufficient target fitting of single feature.In addition,CNN network is used to extract high order hidden information of two-dimensional feature graph generated by Embedding layer.Finally,the features extracted from two parts of the network are integrated through the full connection layer to output classification results.The results showed that the total recall rate of the south Pacific fishery was 87.4% and that of the central fishery was 89.4%.Compared with the F-ACN model with better effect,the recall rate of the central fishery was increased by 5.4% and the total recall rate was increased by 9.1%.It is proved that the combination of feature interactive network and convolutional neural network can obviously improve the accuracy of fishery prediction,and the accuracy can better meet the needs of realistic fishing operations.(2)The second model takes into account that there may be errors in the artificial division of fishery grade,so the prediction of fishery grade is changed to the direct prediction of CPUE value of fishery,which eliminates the error of artificial division of fishery and avoids label division for data set,and the prediction results are more reference.In addition,in order to further improve the performance of the model,additional 16 environmental factors were collected in the environmental factors part of the experimental data set of the model,and the number of environmental factors reached 19,which can better simulate the Marine complex environment.Combined with the selfattention mechanism,this model uses CIN network based on directional feature interaction to extract explicit feature interaction information of spatio-temporal factor data,and uses DNN to extract implicit high-dimensional feature information of environmental factor data.Finally,the two features are fused and the CPUE value is output.The experimental results show that the RMSE and MAE of the model in this chapter are 0.131 and 0.060 respectively compared with similar models,which both reach the minimum value,proving the validity and rationality of the model design in this chapter,and providing a new method for fishery prediction.Two models of the experimental results compared with the same type model achieves the highest accuracy,the experimental results also proved that the deep learning combined with characteristics of interactive technology and applied to the fishery prediction field can effectively improve the predictive accuracy of fishery as well as the subsequent fishery prediction research provides a new research idea.