Multivariate Algorithmic Ensemble For Forecasting Fishing Grounds Of Different Grades Of Southern Pacific Albacore

To achieve high-accuracy forecasting of different grades of albacore fishing grounds in the South Pacific Ocean.This study used the albacore catch data from the Western and Central Pacific Fisheries Commission（WCPFC）and marine environmental factor data by the Copernicus Marine Environment Monitoring Service（CMEMS）from 2009 to 2019 as data sources.The Stacking ensemble model for albacore fishing grounds forecasting was constructed based on Random Forest（RF）,K-Nearest Neighbor（KNN）,Support Vector Machine（SVM）,Gaussian Process（GP）,and e Xtreme Gradient Boosting（XGBoost）algorithms.Comparing and analyzing the accuracy of different machine learning algorithms for fishing grounds forecasting;Exploring the forecasting capability of stacking ensemble learning for different grades of fishing grounds;Quantitatively assessment the effects of marine environmental factors（sea surface temperature（SST）,sea surface height（SSH）,sea surface salinity（SSS）,and chlorophyll-a concentration（Chl-a））on the accuracy of fishing grounds forecasting;Verifying the influence of marine environmental factors SST,SSS,Chl-a,and SSH on the changes of spatial and temporal distributions of albacore catch and catch per unit effort（CPUE）,and determine the contributions of different marine environmental factors to the fishing grounds forecasts based on RF feature importance analysis.The conclusions of the study are as follows:（1）In 2009,2010,2017 and from May to August,the CPUE and catch of albacore were higher,and the change in the distribution of catch and CPUE showed a positive correlation.Albacore resources were abundant in the 5°×5°area with 172.5°E,172.5°W and 157.5°W as the center longitude,and the CPUE was relatively high in the 27.5°S～37.5°S area,while the CPUE was relatively low in the 2.5°S and 47.5°S area.For high values of CPUE of albacore,the corresponding environments were SST at 16～22°C,SSS at 35.5 psu～36 psu,Chl-a at 0～0.1mg/m³,0.25～0.5 mg/m³,and SSH at 0.2～0.4 m,respectively.（2）Among the different machine learning models,the XGBoost model had the highest forecast accuracy（ACC=82.53%）for albacore fishing ground,which was 1.58%～15.09%better than other machine learning algorithms.The results of different grades of albacore fishing grounds based on machine learning models showed that the recall R₁ and accuracy P₁ of the high-yield fishing area were much higher than the recall R₀ and accuracy P₀ of the low-yield fishing area.Most of the low-yield fishing grounds of albacore in the South Pacific Ocean are distributed near the equator,and the high-yield fishing grounds are mainly distributed in the waters from 12.5°S to 37.5°S,and the trend of seasonal migration is obvious.（3）The Stacking ensemble model achieved high accuracy forecasting of albacore fishing grounds（ACC=86.92%）,which improved 4.39%～19.48%overall than the machine learning model.Based on the RF feature-importance analysis,it was found that Latitude（Lat）had the greatest effect on the forecast accuracy of different grades of albacore fishing grounds in February-December（0.377）,Chl-a had the greatest effect on the forecast of albacore fishing grounds in January（0.295）,and Longitude（Lon）had the least effect on the forecast of different grades of fishing grounds（0.037）.