基于深度学习的南海东北部亚中尺度过程识别与分析研究  

IDENTIFICATION AND ANALYSIS OF SUBMESOSCALE PROCESSES IN THE NORTHEASTERN SOUTH CHINA SEA BASED ON DEEP LEARNING

作　　者：金宇涛 缪明芳 张志伟[1,2,3] JIN Yu-Tao;MIAO Ming-Fang;ZHANG Zhi-Wei(Key Laboratory of Ocean Observation and Information of Hainan Province,Sanya Oceanographic Institution,Ocean University of China,Sanya 572024,China;Sanya Oceanographic Laboratory,Sanya 572024,China;Key Laboratory of Physical Oceanography,MOE,China/Frontiers Science Center for Deep Ocean Multipheres and Earth System,Ocean University of China,Qingdao 266100,China)

机构地区：[1]中国海洋大学三亚海洋研究院海南省海洋立体观测与信息重点实验室,海南三亚572024 [2]三亚海洋实验室,海南三亚572024 [3]中国海洋大学物理海洋教育部重点实验室/深海圈层与地球系统前沿科学中心,山东青岛266100

出　　处：《海洋与湖沼》2025年第1期77-89,共13页Oceanologia Et Limnologia Sinica

基　　金：国家重点研发计划,2022YFC3105003号;国家自然科学基金,42222601号,92258301号。

摘　　要：亚中尺度过程是海洋学研究的前沿热点领域,从高分辨率资料中实现亚中尺度信号的快速提取对开展亚中尺度动力学研究具有重要意义。为此,根据亚中尺度过程的物理特性,提出一种基于深度学习的自动识别方法,构建了基于U-Net网络的海洋亚中尺度过程识别网络(submesoscale processes automatic identification network, SM-Net),该网络采用视觉几何组网络作为主干特征提取网络并引入改进的混合注意力模块以提升识别能力。基于高分辨率MITgcm (Massachusetts Institute of Technology general circulation model)模式数据,通过SM-Net准确识别出南海东北部全年的亚中尺度过程,并分类为冷涡、暖涡和锋面。南海东北部亚中尺度冷涡、暖涡和锋面均多发生于冬季,夏季的发生频率较低,但吕宋海峡的亚中尺度过程全年均较为活跃。除吕宋海峡外,亚中尺度冷涡夏季多发生于台湾岛西南海域、吕宋岛西南海域和吕宋岛沿岸,冬季多发生于南海北部陆坡陆架区;亚中尺度暖涡夏季多发生于吕宋岛沿岸,冬季在南海北部陆坡陆架区较为活跃;亚中尺度锋面的时空特征与冷涡相似,但黑潮流经区域的发生频率更高。亚中尺度过程罗斯贝数和动能的时空特征与发生频率具有较好的一致性,暖涡的动能、罗斯贝数和直径均弱于冷涡。上述识别方法在南海的成功运用,为应用SWOT (surface water and ocean topography)卫星数据研究亚中尺度过程提供了一定参考。Submesoscale processes are pivotal in advancing oceanographic research,and the rapid extraction of submesoscale signals from high-resolution data is crucial for advancing the study of submesoscale dynamics.Based on the physical characteristics of submesoscale processes,we proposed an automatic identification method utilizing deep learning,and established a U-Net-based submesoscale processes automatic identification network(SM-Net),in which the VGG(Visual Geometry Group)network is adopted as its backbone for feature extraction and an improved convolutional block attention module is integrated to enhance its recognition capabilities.By employing high-resolution data from MITgcm simulation with the SM-Net,submesoscale processes in the northeastern South China Sea throughout the year was accurately identified and categorized into submesoscale cyclonic eddies,anticyclonic eddies,and fronts.Further analysis revealed their seasonal patterns.The submesoscale cyclonic eddies,anticyclonic eddies,and fronts occur more frequently in winter than in summer,except for the Luzon Strait,where exhibits active submesoscale processes year-round.Outside the Luzon Strait,submesoscale cyclonic eddies are frequently found in the southwestern waters off Taiwan and Luzon and along the Luzon coastline in summer,while in winter,they are more active in the northern continental slope and shelf regions of the South China Sea.Submesoscale anticyclonic eddies are frequently found along the Luzon coast in summer and become more active in the northern continental slope and shelf regions of the South China Sea in winter.The spatiotemporal characteristics of submesoscale fronts are similar to those of cyclonic eddies,with a higher occurrence rate in the regions of the Kuroshio path.The spatiotemporal characteristics of the Rossby number and kinetic energy of submesoscale processes are consistent with their frequency of occurrence,with anticyclonic eddies exhibiting lower kinetic energy,less Rossby numbers,and smaller diameters than those of cyclonic

关 键 词：亚中尺度过程 自动识别 深度学习 语义分割 南海东北部 

分 类 号：P731[天文地球—海洋科学]