Marine mesoscale eddies are a common marine ocean phenomenon. Identifying mesoscale eddies can play an important role in shipping, military, and ocean resource development. Compared to traditional recognition algorithms, artificial intelligence methods can identify mesoscale eddies more accurately and efficiently. An intelligent detection algorithm with loss function gradient optimization was used to realize identification of mesoscale eddies. Firstly, the mesoscale eddy dataset was built from the satellite altimeter remote sensing data and the global ocean physics reanalysis data provided by the Copernicus Marine Environment Monitoring Service (CMEMS), and the PET method and Shapely tools were applied to generate labels for the dataset which would be used in intelligent models. Then, the U-Net AI model was realized to identify the mesoscale eddies. In this model, an attention mechanism for multi sources and fusion of features on each layer were added to optimize model performance. Finally, a loss function gradient optimization was introduced to make the model more consistent with the mesoscale eddies. By introducing 60% loss function gradient constraint weights in the model, the accuracy can be improved from 93.57% to 94.46%. The result show that multi source data, fusion of features and loss function gradient optimization constraints significantly improve the ability of artificial intelligence methods to recognize mesoscale eddies.
LI Yong, ZHANG Jie, YANG Jun-gang, CUI Wei, ZENG Wei, JIAO Fang-tao.
Multi Source Mesoscale Eddies Intelligent Detection Algorithm Based on Loss Function Gradient Optimization[J].
Science Technology and Engineering, 2026, 26(8): 3213-3222 DOI:10.12404/j.issn.1671-1815.2502716
近年来,新兴的人工智能方法也为中尺度涡的检测提供了新的思路[10-21],如基于U-Net模型的EddyNet算法[12]、用于涡旋多目标检测的EddyYolo算法[13]和用于语义和细节融合的金字塔场景解析网络(pyramid scene parsing network,PSPNet)算法[14],可以有效地识别海面涡旋。基于SSH数据的中尺度涡检测是常见方法。Duo等[15]设计并优化了一种OEDNet目标检测模型,该模型以深度残差网络和特征金字塔网络为主要结构,可以自动识别和定位涡流。董子意[16]基于卫星反演得到的海平面异常数据,对U-Net网络进行了一系列改进,嵌入一种卷积注意力机制,以增强特征提取阶段对具有高类判别的区域的注意力。然而,由于人工智能是由纯数据驱动的,如果数据本身具有缺陷,则这些缺陷将会影响中尺度涡的检测结果。如果采用来源不同的数据,例如遥感数据和再分析数据相结合,多源数据无疑会弥补单源数据的不足。此外,深度学习是采用非线性进行特征提取的算法,其特征提取是为了将模型的损失函数降到最低以尽可能拟合真实数值,但模型的特征提取能力存在极限,如果将损失函数的梯度差异也加入到损失函数本身,进一步增加模型的特征差别,势必会提高人工智能模型的检测准确度。
Cui WeiWangWei, MaYi, et al. Identification and characteristics analysis of mesoscale eddies in the northwest Pacific based on altimeter data from 1993 to 2014[J]. Acta Oceanography, 2017, 39(2): 16-28.
LiZhijun, NingLiang, ChenRenlin. Exploration and development of international seabed area resources in China[J]. Science Technology and Engineering, 2024, 24(28): 11558-11564.
[7]
CheltonD B, GaubeP, SchlaxM G, et al. The influence of nonlinear mesoscale eddies on near-surface oceanic chlorophyll[J]. Science, 2011, 334(6054): 328-332.
[8]
张宇伯. 习近平关于海洋生态环境治理重要论述研究[D]. 大连: 大连海事大学, 2023.
[9]
ZhangYubo. Research on important discussions on marine ecological environment governance from Xi Jinping[D]. Dalian: Dalian Maritime University, 2023.
[10]
OkuboA. Horizontal dispersion of floatable particles in vicinity of velocity singularities such as convergences[J]. Deep Sea Research and Oceanographic Abstracts, 1970, 17(3): 445-454.
[11]
SadarjoenI A, PostF H. Detection, quantification, and tracking of vortices using streamline geometry[J]. Computers & Graphics, 2000, 24(3): 333-341.
[12]
NencioliF, DongC, DickeyT, et al. A vector geometry-based eddy detection algorithm and its application to a high-resolution numerical modelproduct and high-frequency radar surface velocities in the southern california bight[J]. Journal of Atmospheric & Oceanic Technology, 2010, 27(3): 564.
[13]
ChaigneauA, EldinG, DewitteB. Eddy activity in the four major upwelling systems from satellite altimetry (1992—2007)[J]. Progress in Oceanography, 2009, 83(1-4): 117-123.
ZhangJiahao. Research on the identification method of mesoscale eddies in the ocean based on high-resolution neural networks[D]. Changsha: National University of Defense Technology, 2023.
[18]
LguensatR, SunM, FabletR, et al. EddyNet: a deep neural network for pixel-wise classification of oceanic eddies[C]// IGARSS 2018—2018 IEEE International Geoscience and Remote Sensing Symposium. Valencia: IEEE, 2018: 8518411.
XiePengfei, SongTao, XuDanya, et al. Mid scale vortex detection technology based on deep learning and its application in acoustic field[J]. Ocean Information, 2020, 35(1): 18-26.
[21]
XuG, ChengC, YangW, et al. Oceanic eddy identification using an AI scheme[J]. Remote Sensing, 2019, 11: 1349.
[22]
DuoZ J, WangW K, WangH Z. Oceanic mesoscale eddy detection method based on deep learning[J]. Remote Sensing, 2019, 11: 1921.
[23]
董子意. 基于深度学习的海洋中尺度涡自动检测模型[D]. 杭州: 浙江大学, 2022.
[24]
DongZiyi. A deep learning based automatic detection model for mesoscale eddies in the ocean[D]. Hangzhou: Zhejiang University, 2022.
[25]
MasonA E, PascualJ, McWilliamsC. A new sea surface height-based code for oceanic mesoscale eddy tracking[J]. Journal of Atmospheric and Oceanic Technology, 2014, 31: 1181-1188.
[26]
RonnebergerO, FischerP, BroxT. U-Net: convolutional networks for biomedical image segmentation[M]. Berlin: Springer International Publishing, 2015: 234-241.
[27]
吴琦. 基于深度学习的海洋表面温度预测方法研究[D]. 哈尔滨: 哈尔滨工业大学, 2020.
[28]
WuQi. Research on deep learning based ocean surface temperature prediction method[D]. Harbin: Harbin Institute of Technology, 2020.
[29]
BadrinarayananV, HandaA, CipollaR. SegNet: a deep convolutional encoder-decoder architecture for robust semantic pixelwise labelling[J]. arXiv Preprint, arXiv.1505.07293, 2015.
LuXuyi, ShanGuihua, LiGuan. Identification and visualization of ocean mesoscale eddies based on deep learning[J]. Computer Systems Applications, 2020(4): 65-75.