基于注意力的多尺度残差U-Net的海洋中尺度涡检测  

A MULTI-SCALE ATTENTION RESIDUAL-BASED U-NET MODEL FOR DETECTION OF OCEAN MESOSCALE EDDIES

作　　者：王丽娜[1,2] 孙阳 张红春 王旭东 董昌明 WANG Li-Na;SUN Yang;ZHANG Hong-Chun;WANG Xu-Dong;DONG Chang-Ming(School of Artificial Intelligence(School of Future Technology),Nanjing University of Information Science and Technology,Nanjing 210044,China;Southern Marine Science and Engineering Guangdong Laboratory(Zhuhai),Zhuhai 519080,China;School of Computer Science,Nanjing University of Information Science and Technology,Nanjing 210044,China;School of Marine Sciences,Nanjing University of Information Science and Technology,Nanjing 210044,China)

机构地区：[1]南京信息工程大学人工智能学院(未来技术学院),江苏南京210044 [2]南方海洋科学与工程广东省实验室(珠海),广东珠海519080 [3]南京信息工程大学计算机学院,江苏南京210044 [4]南京信息工程大学海洋科学学院,江苏南京210044

出　　处：《海洋与湖沼》2025年第1期64-76,共13页Oceanologia Et Limnologia Sinica

基　　金：南方海洋科学与工程广东省实验室(珠海)资助项目,SML2020SP007号;国家重点研发项目,2023YFC3008204号。

摘　　要：海洋中尺度涡是一类重要的海洋现象,其特征是海洋中的螺旋运动,伴随着海水温度、营养物质以及能量的输送,对海洋生态系统和全球的气候变化起着重要影响。因此,海洋涡旋的智能识别成为海洋学的研究热点之一。由于海洋中尺度涡数量众多且大小不同,存在检测精度不高问题。为了提高海洋中尺度涡的检测精度,提出一种基于注意力的多尺度残差U-Net的海洋涡旋检测模型(dual cross-attention-pyramid spilt attention-Res U-Net, DCA-PRUNet)。该模型采用基于注意力的编解码器结构。编解码结构中,引入金字塔分割注意力(pyramid spilt attention,PSA)以提取多尺度特征,并捕获不同涡旋的特征信息;此外,为了解决网络过深导致模型无法训练的问题,引入残差学习模块。同时,为了使解码器更好地恢复涡旋细节信息,引入双交叉注意力模块(dual cross-attention, DCA)捕获编码器各个阶段的特征依赖。选取西北太平洋海域的海平面异常(sea level anomaly,SLA)与海面温度(sea surface temperature,SST)数据进行建模,实验结果表明DCA-PRUNet涡旋检测的准确率达到95.12%,F1分数达到91.21%,显著优于现有的模型,验证了该模型的有效性。The mesoscale eddies in the ocean is an important ocean phenomenon,which is characterized by the spiral movement in the ocean;with the transportation of seawater temperature,nutrients,and energy,it has an important impact on marine ecosystems and global climate change.Therefore,intelligent identification of ocean eddies has become one of the research hotspots in oceanography.Due to the large number of mesoscale eddies in the ocean and their different sizes,there is a problem of low detection accuracy.To improve the detection accuracy of mesoscale eddies in the ocean,a novel model,the dual cross-attention-pyramid spilt attention-Res U-Net(DCA-PRUNet),was proposed for ocean eddy detection.The model adopts an attention-based encoder-decoder structure,in which the pyramid spilt attention(PSA)is introduced to extract multi-scale features and capture the characteristic information of different eddies.In addition,to solve the problem that the model cannot be trained due to the network layers being too deep,a residual learning module is incorporated.At the same time,to enable the decoder to better restore the eddy details,a dual cross-attention module(DCA)is combined to capture the feature dependencies in each stage of the encoder.Sea level anomaly(SLA)and sea surface temperature(SST)data in the NW Pacific Ocean were selected for modeling.The experimental results show that the accuracy of DCA-PRUNet eddy detection reached 95.12%with F1 score of 91.21%,which is significantly better than the existing models,verifying the effectiveness of the novel model.

关 键 词：海洋涡旋 深度学习 金字塔分割注意力 残差学习 双交叉注意力 

分 类 号：P731.16[天文地球—海洋科学]