Construction And Evaluation Of Model For The Spatial And Temporal Distribution Of Albacore Tuna (Thunnus Alalunga) Habitats In The Western And Central Pacific Ocean Based On Machine Learning

Climate change and human activities are having a significant impact on the spatial and temporal distribution of marine fishes.Marine fishes have important management and economic values,and it is necessary to conduct research on the changes in the distribution of marine fishes and their response to climate change.With the development and widespread application of technologies such as machine learning and big data,fish distribution research has entered a new phase.As an emerging model,machine learning models outperform traditional modelling methods in screening and predicting the abundance of important environmental factors affecting fish distribution.In the face of an increasing number of new models and complex fish distribution prediction scenarios,there is an urgent need to investigate the comparison and selection of prediction models for specific fish species.To make full use of the advantages of different models,the most suitable model can be selected quickly and accurately for more accurate fish distribution studies and forecasts for different types of fisheries data,which can be an important guide for fisheries management organizations and enterprises to improve production efficiency,rationalize operational areas and formulate fish conservation policies.The albacore tuna(Thunnus alalunga)is a pelagic species with highly migratory characteristics,and its spatial and temporal distribution is significantly influenced by the elements of the marine environment.Currently,research on the distribution of T.alalunga is moving from simple statistical models to advanced and complex machine learning models.The ability of marine remote sensing to capture more elements of the marine environment over large areas has made machine learning useful in areas such as fish distribution and fishing ground forecasting,and many models and methods have shown great potential.It is of great practical and theoretical importance to try to apply machine learning models to study the spatial and temporal distribution dynamics of T.alalunga.In this study,marine environmental data were used as input factors,including Sea Surface Temperature(SST),Sea Surface Salinity(SSS),Chlorophyll-a(Chl-a),deep water temperature and deep dissolved oxygen data,and T.alalunga production data obtained from the Western and Central Pacific Fisheries Commission(WCPFC)from2013 to 2018 were used as output factors to construct artificial neural network models(ANN),Support Vector Machine models(SVM)and Random Forest Models(RF).The optimal parameters of each model were found by the manual grid search method.The artificial neural network model with 9 hidden layer nodes,the support vector machine model with Linear kernel function and penalty coefficient of 4,and the random forest model with ntree of 100 and mtry of 3 were found to be the optimal models,and their prediction accuracy reached 61.3%,56.4% and 64.8% respectively.Combining the evaluation and interpretation results of all optimal models,it was concluded that the random forest model constructed in this study with low learning cost,simple hyperparameter optimization and good ease of use was the optimal machine learning model for T.alalunga in the Western and Central Pacific Ocean,with the following results.(1)The accuracy of the optimal model reached 64% by cross-validation test,with the highest Kappa value of 0.51.(2)From the perspective of the factor contribution output,the total factor contribution of the RF model was higher than the other two models,with the highest contribution of spatial factors(longitude and latitude),followed by temperature(SST and deep water temperature),Chl-a slightly lower than temperature,then dissolved oxygen data,and the lowest contribution of both SSS.(3)Single-factor sensitivity analysis based on the random forest model showed that the appropriate range for SST was 23°C to 31°C,with the highest grade predicted between 24.5°C and 31;the appropriate range for SSS was 33.5 to 36,with the highest grade achieved between 34.5 and 36;the appropriate range for Chl-a was 0.08 to 0.9,with the highest grade achieved between 0.08 and 0.0.65 highest grade;T50 showed the most suitable range at temperatures between 11.5 and 25°C;T100 showed the most suitable range between 7.03 and 16°C;T200 showed the most suitable range between4.15 and 13°C;D50 showed the most suitable range between 0.09 and 8.5;D100 showed the most suitable range between 0.01 and 0.5 and around 2.8;D200 showed the most suitable range between 0.01 and 0.018 or so achieved two suitable grades.(4)The predicted spatial distribution of T.alalunga fisheries for each month in 2018 was predicted and compared with the actual yield classes,and it was found that the RF model was significantly more accurate than the other two models,with an overall accuracy of 88.3%,and all classes could be predicted with an average accuracy of around90%,and the variation between months was small,expressing the high stability and interpretation of the model.This study provides new theoretical support for the study of the spatio-temporal distribution of T.alalunga in the Central and Western Pacific Ocean and the study of fishing ground forecasts,and provides new ideas for machine learning models in the distribution studies of other marine fishes.