Research On Fishing Grounds Forecasting Models Of Albacore (Thunnus Alalunga) Based On Ensemble Learning In The South Pacific

Albacore tuna(Thunnus alalunga)is an oceanic pelagic fish,which is widely distributed in tropical and temperate waters.The South Pacific is rich in tuna resources,and China has become the most important fishing country for albacore tuna in the South Pacific since 2008.Albacore tuna is a highly migratory fish.The location of its fishing grounds is closely related to migratory paths.At the same time,the formation of fishing grounds is also affected by the change of environmental conditions,showing inter annual and inter monthly changes.Accurately forecasting the central fishing grounds can guide the ocean fishing enterprises to arrange the operation position reasonably and improve the production capacity effectively,which plays an important role in China’s albacore tuna longline fishing industry.Based on the historical data of albacore tuna longline fishing in the South Pacific from 2008 to 2015,this study analyzed the changes of CPUE in the interannual,monthly,longitude and latitudinal directions,and studied the migration characteristics of albacore tuna in the South Pacific by analyzing the changes of the center of gravity of longitude and latitude in the fishing grounds over time.The study selected several environmental factors including sea surface temperature,chlorophyll a concentration,sea surface temperature anomaly,chlorophyll anomaly,sea surface temperature gradient,chlorophyll gradient,sea level anomaly,the previous and the next month’s sea surface temperature and chlorophyll value,and three spatiotemporal factors of month,longitude and latitude.Based on six kinds of ensemble learning models of bagging,boosting and stacking,the albacore tuna fishing grounds forecasting model were established,with a monthly temporal resolution and a spatial resolution of 0.5°×0.5°.And the optimal parameters of the model were determined by combining the 10fold cross validation and grid search idea.Finally,the highest accuracy integrated model was used to predict the South Pacific albacore tuna fishing grounds by month in2015,and the high CPUE fishing grounds was compared with the actual fishing grounds to verify the prediction model with the overall accuracy of each month and the high CPUE fishing area prediction accuracy as the standard.The results are as follows:1)The results of the study on the temporal and spatial distribution and the changing trend of the center of gravity of the albacore tuna fishery in the South Pacific show that:the CPUE level is lower from January to April,and higher from May to August,which is the most abundant period in the South Pacific.The CPUE from September to December is lower than that in the previous period,and enters the end of the fishing season.The working fishing grounds are mainly concentrated in the sea areas of 8°S～22°S and 158°E～176°E.In the latitude direction,the center of gravity changes from south to north,then to south and then to north,and in the longitude direction,it changes from east to west,then to east and then to west.From May to August,the center of gravity of fishing grounds is more inclined to the southeast sea area.2)By comparing the performance of the forecasting model based on ensemble learning,the results show that:The prediction accuracy of random forest(RF),bagging decision tree(Treebag),C5.0,gradient boosting decision tree(GBDT),adaptive boosting(Ada Boost)and stacking integrated models is 75.52%,73.87%,72.99%,71.14%,71.33%and 75.84%respectively.The accuracy rate of the training set data of the albacore tuna fishing grounds in this paper is 70%.The bagging model is better than the boosting model,and RF is higher.The accuracy rate of the stacking model with KNN added is slightly higher than RF,so the stacking integrated model is selected for further accuracy test.3)It is found that the contribution of sea surface temperature to the classification of albacore tuna fishing grounds is relatively large,followed by the sea level height anomaly,and the chlorophyll is the weakest.It is difficult to divide high and low CPUE fishing grounds by a single environmental factor,so it is necessary to combine multiple factors to increase the attribute difference between samples.In addition,the temperature of high CPUE fishing grounds is concentrated in the range of 26.4～30°C,chlorophyll is concentrated in the range of 0.03～0.15 mg/m~3,and the sea surface height anomaly is concentrated in the range of-0.03～0.19 m.4)Further test on the prediction accuracy of stacking integrated model shows that:in 2015,the comprehensive accuracy of monthly fishery prediction is 63.86%～82.14%,with an average accuracy of 70.99%;the accuracy of high CPUE fishery prediction is 62.71%～97.85%,with an average accuracy of 78.76%.Stacking integrated model has a good effect and feasibility on the prediction of albacore tuna fishing grounds in South Pacific.The innovative results of this study are mainly reflected in:1)In view of the single type of forecasting model in the South Pacific albacore tuna fishing grounds,the integrated learning model is introduced,which improves the forecasting model of albacore tuna fishing grounds in the South Pacific Ocean,and provides a new idea for the selection of forecasting model of ocean fishing grounds in China;2)From the perspective of high and low CPUE fishing grounds defined by the first third quantile,this paper analyzes the impact of various environmental factors on fishing grounds classification,and considers that multiple factors should be combined to increase the attribute differences among samples,which provides a reference for the selection of inputing attributes of the integrated model;3)The study combines 10 fold cross validation to output the average performance of the model.At the same time,the new data set is used to test the accuracy of the model again.The dual evaluation verifies the prediction reliability of the integrated learning model.