袁红春，王敏，刘慧，陈冠奇

摘要：长鳍金枪鱼是南太平洋渔业生产中主要的捕捞对象，准确预测其渔场分布对提高渔业捕捞效率具有重要意义。针对传统渔场预测方法预测精度低的问题，本研究提出一种基于特征交互与卷积网络的渔场预测模型——CNN-Cross。该模型引入Embedding层对数据进行处理，解决了One-Hot Encoding（独热编码）带来的特征稀疏性问题以及手动特征工程对结果的影响。同时，引入Cross网络提取特征之间的交互信息，消除了单特征对目标拟合不足的问题，并且结合CNN网络对Embedding层生成的二维特征图进行高阶隐藏信息提取，最后将两部分网络提取到的特征融合，输出分类结果。使用渔业数据对模型预测效果进行验证，结果表明，模型预测南太平洋渔场总召回率达到87.4%，中心渔场召回率达到89.4%。表明，将特征交互网络与卷积神经网络相结合可以明显提高渔场预报精度，且精度能够较好地满足现实渔业作业需求。

关键词：长鳍金枪鱼;Cross网络;卷积神经网络;特征交互

中图分类号：S931.41文献标识码：A文章编号：1000-4440（2021）06-1501-09

Fishing ground prediction model based on feature interaction and convolutional network

YUAN Hong-chun 1，2，WANG Min 1，LIU Hui 1，CHEN Guan-qi 1

（1.College of Information Technology， Shanghai Ocean University， Shanghai 201306，China;2.Key Laboratory of Fisheries Information，Ministry of Agriculture and Rural Affairs，Shanghai 201306，China）

Abstract：Thunnus alalunga is the main fishing target of fishery production in the South Pacific Ocean. It is of great significance to accurately predict the fishery distribution of T. alalunga for improving fishery efficiency. In view of lack of accuracy of traditional fishery prediction methods， this paper proposed a fishery prediction model based on feature interaction and convolutional network—CNN-Cross. In this model， the Embedding layer was introduced to process the data， which solved the problem of feature sparsity caused by One-Hot Encoding and the influence of manual feature engineering on the result. At the same time， the Cross network was introduced to extract interactive information between different features to eliminate the problem of insufficient target fitting by single feature， and the two-dimensional feature map generated by the Embedding layer was extracted with the CNN network for high-order hidden information extraction. Finally， the features extracted by two networks were fused and the classification results were output. The effect of the model was verified by fishery data. The results showed that the predicted total recall rate of the South Pacific fishery reached 87.4%， and that of the central fishing ground reached 89.4%. The research results show that the combination of feature interaction network and convolutional neural network can obviously improve the accuracy of fishery forecast， and the accuracy can better meet the needs of practical fishery operations.

Key words：Thunnus alalunga;Cross network; convolutional neural network;feature interaction

長鳍金枪鱼作为一种大洋性中上层高度洄游生物，具有较高的经济价值、营养价值[1]。长鳍金枪鱼广泛分布于各热带海域与温带海域[2]，尤以南太平洋海域分布较广。近年来，长鳍金枪鱼已成为南太平洋金枪鱼延绳钓的主要捕捞对象之一[3-5]。因此，准确预测南太平洋长鳍金枪鱼渔场，对于降低渔业作业成本、提高寻找渔场效率具有极大的意义。

目前国内外学者仍然使用传统的方法与模型对渔场进行预测，并没有对渔场预测与深度学习结合进行进一步探究。……