Fishing ground forecasting of bigeye tuna (Thunnus obesus) in the tropical waters of Atlantic Ocean based on ensemble learning

In order to improve the accuracy of bigeye tuna (Thunnus obesus) fishing ground forecast model in the tropical waters of Atlantic Ocean, a series of fishery forecast models were established based on the logbook data of 13 Chinese longliners from 2013 to 2019 and the corresponding marine environment data, e.g. sea surface wind speed, chlorophyll a concentration, eddy kinetic energy, upper boundary depth of thermocline, vertical temperature, salinity and dissolved oxygen in 0-500 m water layer. T. obesus CPUE was calculated based on the logbook data. The environmental factors related to T. obesus CPUE were screened out from 29 environmental factors by correlation analysis. The non-collinear environmental factors were selected by collinearity analysis based on the variance expansion factor (VIF) and used to build the bigeye tuna fishing ground prediction models. The Spearman correlation coefficients between non-collinear environmental factors and T. obesus CPUE were calculated and used to analyze the relative importance of the environmental factors to the T. obesus CPUE. These series of prediction models, e g. K-Nearest Neighbor (KNN), Logistic Regression (LR), Classification and Regression Tree (CART), Support Vector Machine (SVM), Artificial Neural Networks (ANN), Random Forest (RF), Gradient Boosting Decision Tree (GBDT), and Stacking ensemble model (developed by KNN, RF and GBDT, STK) were built by using 75% of data and verified by using 25% of data. The time resolution of T. obesus CPUE and marine environment data was one day, and the spatial resolution was 2° × 2°. The performance of 8 models were evaluated by the area under the receiver operating characteristic curve (AUC) and prediction accuracy. The maps of the actual fishing ground and the predicted fishing ground were overlapped by ArcGIS and used to evaluate the performance of the best model. The central bigeye tuna fishing ground was determined by the nuclear density analysis tool of ArcGIS. The results show that (1) compared with the single model (KNN, LR, CART, SVM, ANN, RF and GBDT), the forecasting performance of T. obesus fishing ground of STK model was better and relatively stable. The accuracy (AUC) of the STK model, KNN, LR, CART, SVM, ANN, RF and GBDT were 81.62% (0.781), 79.44% (0.778), 72.81% (0.685), 74.84% (0.717), 73.67% (0.702), 67.70% (0.500), 80.96% (0.780), and 78.13% (0.747), respectively; (2) the distribution of central fishing ground predicted by STK model was basically consistent with the actual distribution of central fishing ground, all of them were mainly distributed in the area of 5 °N-10 °N, 33 °W-43 °W; (3) the marine environmental factors that affect the distribution of T. obesus fishing grounds in the Atlantic Ocean mainly included dissolved oxygen of 300 m layer, salinity of 500 m layer, sea surface wind speed and upper boundary depth of thermocline, and the relative importance were 13.24%, 9.12%, 9.12% and 8.81%, respectively. The results suggest that the accuracy of the STK model for T. obesus fishing ground forecast in the Atlantic Ocean is high.