Forecasting models for yellowfin tuna fishing ground in the central and western Pacific based on machine learning

An accurate forecast can be greatly contributed to the yellowfin tuna fishing ground in the western and Central Pacific. However, a large amount of fishery data, and high feature dimension have posed a great over-fitting on the various classification in recent years. The random forest parallel integration can be expected to achieve the excellent performance of the extreme gradient boosting decision tree algorithm. In this study, a hybrid integration model was proposed to combine the Xgboost with Random Forest (XGBRF) with the random forest and extreme gradient lifting decision tree. The fishery production data was also collected from the operation data of 43 distant-water longline fishing vessels of China Aquatic Group in the western and Central Pacific (0°-30°S; 110°E-170°W) from 2008 to 2019, including catch information, such as amount, job date, as well as the job latitude and longitude. A comparison was performed on the fishery data, including the concentration of chlorophyll a, eddy kinetic energy, sea surface height anomalies, temperature and salinity of the 0-500 m mixed water layer. A total of 36 variable combinations were used as the original data set, including the Southern Oscillation Index (SOI), the Arctic Oscillation Index (AOI), the Pacific Decadal Oscillation Index (PDOI), and North Pacific Gyre Oscillation Index (NPGOI). The original data set was divided into the training set and test set after the screening and normalization of the variance expansion factor, accounting for 80% and 20%, respectively. The training set was used to train eight models, including classification and regression, logistic regression, k-nearest neighbor, adaptive boosting, gradient boosting decision tree, xgboost, random forest, and XGBRF. The five-fold cross-validation was used for each model to determine the optimal parameters. Finally, the model was verified to superimpose the actual fishing ground of the test set. The experimental results showed that: 1) There was a significant correlation between the catch per unit fishing effort and various variable factors. There was also a great decrease in the degree of collinearity between the variables that were filtered by variance inflation factor. 2) The XGBRF hybrid ensemble model also significantly improved the performance of XGBoost and RF models. Specifically, the highest accuracy rate and Area Under Curve (AUC) were 75.39%, and 79.48%, respectively. The Receiver Operator Characteristic (ROC) curve of the XGBRF model was closer to the upper left, indicating the best performance of the forecasting model than before. 3) The sea surface temperature was the most important factor to dominate the distribution of yellowfin tuna fishing ground, accounting for 7.573%. The temperature of the 300 m water layer was equally important for the yellowfin tuna, which was 7.369%. In addition, the greater impact was also found in the salinity of the 50-meter water layer, the SOI, the concentration of chlorophyll a, and the surface salinity. There was a relatively low influence of other large-scale climatic factors, except for the SOI. 4) There was only a small deviation between the fishing ground predicted by the XGBRF model and the actual fishing ground, indicating the high accuracy and reliability of the prediction. Overall, the XGBRF ensemble model performed the best on the fishing ground forecast of yellowfin tuna in the western and Central Pacific. The finding can also provide a strong reference for the fishing ground forecast.