基于改进的XGBoost模型预测南太平洋长鳍金枪鱼资源丰度

Prediction of albacore tuna abundance insouth Pacific based on improved XGBoost model

为提高渔场资源丰度预测准确性,以单位捕捞努力量渔获量(catch per unit effort,CPUE)为长鳍金枪鱼(Thunnus alalunga)资源丰度指标,利用海洋遥感、Argo等获取的海洋环境因子,在最优分布式决策梯度提升树(XGBoost)模型基础上,采用卷积神经网络(Convolutional Neutral Network,CNN)进行高维海洋环境数据特征提取,并利用模拟退火算法(Simulate Anneal,SA)对最优分布式决策梯度提升树(Extreme gradient boosting,XGBoost)模型进行超参数优化,提出了改进的XGBoost模型CNN-SA-XGBoost模型,实现对南太平洋长鳍金枪鱼资源丰度的回归预测。实验表明,在南太平洋长鳍金枪鱼资源丰度预测中,CNN-SA-XGBoost模型的均方根误差为0.486,较XGBoost减少12.4%,较多元线性回归(Multiple Linear Regression)、随机森林(Random Forest,RF)、BP神经网络等模型预测误差降低11.8%~28.4%。且改进的XGBoost模型在一定程度上改善了传统资源丰度预测模型面对高维环境数据和缺失值较多的渔业生产数据时预测误差较大的问题,为远洋渔场预报提供了新方法。

In order to improve the accuracy of fishery resource abundance prediction,Catch per unit effort(CPUE) used as an indicator of the abundance of albacore tuna(Thunnus alalunga) resourcesmarine environmental factors such as marine remote sensing and Argo.Based on the xtreme gradient boosting model(XGBoost),onvolutional neutral network(CNN) used for feature extraction of high-dimensional marine environment data,and imulated annealing algorithm(SA) used to optimize XGBoost model.An improved XGBoost model CNN-SA-XGBoost model was proposed to realize the regression prediction of the abundance of albacore tuna resources in the South Pacific.Experiments show that in the prediction of abundance of albacore tuna resources in the South Pacific,the root mean square error of the CNN-SA-XGBoost model is 0.486,12.4% lower than XGBoost.Compared with ultiple linear regression and andom forest(RF),BP neural network and other models reduce the prediction error by 11.8 28.4%.he improved XGBoost model improved the traditional resource abundance forecasting model to a large extent when it face high-dimensional environmental data and fishery production data with many missing values,which provides a new method for the prediction of pelagic fisheries.