Research on fishing grounds forecasting models of albacore tuna based on ensemble learning in South Pacific

According to the data of longline fishing of Thunnus alalunga in the South Pacific Ocean from 2008 to 2015, we investigated  11 environmental indicators (sea surface temperature, chlorophyll a concentration, sea surface temperature anomaly, chlorophyll anomaly, sea surface temperature gradient, chlorophyll gradient, sea level anomaly, and sea surface temperature and chlorophyll values one month before and after the corresponding fishing area grid) and three spatio-temporal indicators (month, longitude and latitude).  Besides, based on six integrated learning models, taking month as time resolution and 0.5°×0.5° as space resolution, we carried out the model construction and prediction of T. alalunga fishery in the South Pacific. The optimal parameters of the model are determined by 10 fold cross validation and grid search. The accurary rates of RF (Random forest), Treebag, C5.0 decision tree, GBDT (Gradient boosting decision tree), AdaBoost (Adaptive boosting) and Stacking integration model were 75.52%, 73.87%, 72.99%, 71.14%, 71.33% and 75.84 %, respectively. The Stacking integration model had the highest accuracy. We used 2015 environmental data to test the forecast accuracy, and find that the overall forecast accuracy was 63.86%−82.14%, with an average of 70.99%; the forecast accuracy of catch per unit effort (CPUE) fishing area was 62.71%−97.85%, with an average of 78.76%. The results show that the Stacking integration model has a good effect and feasibility on the prediction of T. alalunga fishery in the South Pacific.