摘要
为解决用传统方法进行渔场预测时存在的性能欠佳、特征转换困难、拟合程度不足等问题,提出了一种基于深度学习和典型相关分析的新型渔情预测方法——CNN-DNN-CCA(连接融合)-RBF模型,该方法首先在5°×5°渔业作业区域内将不同海洋环境因子按相对空间位置映射为三维矩阵,然后分别采用卷积神经网络(CNN)和深度神经网络(DNN)对海表温度(SST)、叶绿素a(Chl-a)浓度、海面高度(SSH)3种环境因子和渔场时空因子两种多源异构数据进行模态特征提取,得到两种不同模态的特征向量,并将两种特征向量通过典型相关分析(CCA)进行特征级融合,最后将融合后的特征输入到径向基函数网络(RBF)中进行分类。结果表明,通过试验验证,基于深度学习和典型相关分析的渔场预报模型CNN-DNN-CCA(连接融合)-RBF对南太平洋长鳍金枪鱼Thunnus alalonga中心渔场的召回率达到了90.3%,相较于随机森林(RF)、CNN和DNN模型提高了6.8%~21.8%。研究表明,CNN-DNN-CCA(连接融合)-RBF新型渔情预测模型通过深度学习和典型相关分析方法分别进行特征自动提取和特征融合,消除了冗余信息,简化了特征转换,提高了运算速度和预测精度。
In order to solve the problems of poor performance,difficulty in feature conversion,and insufficient fitting degree in traditional methods of fishing ground prediction,a new fishing situation prediction method—CNN-DNN-CCA(fusion with connection)-RBF model is established based on deep learning and canonical correlation analysis.First,in this method different marine environmental factors were maped into a three-dimensional matrix according to their relative spatial positions within a 5°×5°fishery operation area.Then,the convolutional neural network(CNN)and the deep neural networks(DNN)are used to extract the modal features of the three environmental factors including sea surface temperature,concentration of chlorophyll a,and the sea surface height,and the spatiotemporal factor of fishing grounds.The two feature vectors are fused at the feature level through the canonical correlation analysis(CCA)method.Finally,the fused features were inputted into the radial basis function network(RBF)for classification.The experimental results showed that the fishing ground prediction model based on deep learning and canonical correlation analysis had a recall rate of 90.3%for the South Pacific albacore fishing center,increased by 6.8%-21.8%compared with the random forest(RF),CNN model and DNN model.The new fishing situation prediction model proposed in this study is shown to extract and fuse features automatically through deep learning method and canonical correlation analysis method,and is featured by to elimination of redundant information,simplified feature transformation,and improvement of the operation speed and prediction accuracy.The findings provide a new idea for the fishing ground prediction of albacore tuna.
作者
袁红春
刘慧
张硕
陈冠奇
YUAN Hongchun;LIU Hui;ZHANG Shuo;CHEN Guanqi(College of Information Technology,Shanghai Ocean University,Shanghai 201306,China;Key Laboratory of Fisheries Information,Ministry of Agriculture and Rural Affairs,Shanghai 201306,China)
出处
《大连海洋大学学报》
CAS
CSCD
北大核心
2021年第4期670-678,共9页
Journal of Dalian Ocean University
基金
国家自然科学基金(41776142)
国家重点研究发展计划“蓝色粮仓科技创新”项目(2018YFD0701003)。
