袁红春 张硕 陈冠奇

摘要： 为解决传统渔场渔情预测方法在处理高维复杂海洋数据时存在人工干预较多、拟合困难、精度不高的问题，提出了一种基于双模态深度学习的渔场渔情预测方法。首先，该方法将不同海洋环境因子在5°×5°渔业作业区域范围内按照空间相对位置映射为三维矩阵。然后，分别使用卷积神经网络模型（CNN）和深度神经网络模型（DNN）对海洋环境因子和时空因子2种异构数据进行特征提取。最后，将基于时空信息的深度神经网络模型与卷积结构进行特征融合，再将融合后的特征经过全连接层进行分类。试验结果表明，双模态深度学习模型对南太平洋长鳍金枪鱼中心渔场的渔场渔情预报率达到了89.8%，较其他渔场渔情预报模型精度提高10%～30%。同时由于该模型使用卷积神经网络，可以对任意空间分辨率的海洋环境因子进行特征提取，省去了手动匹配不同空间分辨率的过程，减少了人工干预，对南太平洋长鳍金枪鱼的渔业作业与渔场渔情预报有极高的指导意义。

关键词： 双模态深度学习模型;渔场渔情预报;长鳍金枪鱼

中图分类号： S934  文献标识码： A  文章编号： 1000-4440（2021）02-0435-08

Abstract： To solve the problems of much manual intervention， difficulty in fitting and low accuracy in processing ocean data of high-dimensional complex by traditional fishery forecasting methods in the fishing ground， a fishery forecasting method based on dual-modal deep learning was proposed. Firstly， the method mapped different marine environmental factors within a 5°×5° fishery operation area into a three-dimensional matrix according to their relative spatial positions. Secondly， features of two heterogeneous data such as marine environmental factors and spatiotemporal factors were extracted by convolutional neural network （CNN） model and deep neural network （DNN） model respectively. Finally， the deep neural network model based on spatiotemporal information and the convolution structure were fused by feature， and the fused features were classified through the fully connected layer. The results showed that， the forecast rate of fishery by dual-modal deep learning model in the central fishing ground of albacore in the South Pacific reached 89.8%， which improved the forecast accuracy by 10%-30% compared with forecast models in other fishing grounds. At the same time， because the model used a convolutional neural network， which could extract features of marine environmental factors with any spatial resolution， thus eliminated the process of matching different spatial resolutions by manual and reduced manual intervention， which showed extremely high guiding significance for the fishing operations and fishery forecast in the fishing ground of albacore in the South Pacific.

Key words： dual-modal deep learning model;fishery forecasting in the fishing ground;albacore tuna

長鳍金枪鱼（Thunnus alalunga）是高度洄游的大洋性鱼类，因其经济价值高，资源量相对丰富，故该鱼种具有较大开发潜力，已引起包括中国在内的很多渔业国家的关注与重视。如何提高南太平洋长鳍金枪鱼的渔场渔情预测水平已成为国内外学术界研究热点之一。通过调研区域渔业管理部门的渔获量和努力量数据，发现长鳍金枪鱼的种群主要分布于太平洋，并且在南太平洋海域近二十年来长鳍金枪鱼产量逐年增长。……