基于U-Net的海洋锋智能检测模型  

Oceanic Front Detection Model Based on U-Net Network

作　　者：任诗鹤 韩焱红[2] 李竞时 赵亚明 匡晓迪[1] 吴湘玉[1] 杨晓峰 REN Shihe;HAN Yanhong;LI Jingshi;ZHAO Yaming;KUANG Xiaodi;WU Xiangyu;YANG Xiaofeng(Key Laboratory of Marine Hazards Forecasting,National Marine Environmental Forecasting Center,Ministry of Natural Resources,Beijing 100081;Public Meteorological Service Center,China Meteorological Administration,Beijing 100081;P.O.Box 5111,Beijing 100094;State Key Laboratory of Remote Sensing Science,Aerospace Information Research Institute,Chinese Academy of Sciences,Beijing 100101)

机构地区：[1]国家海洋环境预报中心自然资源部海洋灾害预报技术重点实验室,北京100081 [2]中国气象局公共气象服务中心,北京100081 [3]北京市北京100094 [4]中国科学院空天信息创新研究院遥感科学国家重点实验室,北京100101

出　　处：《空间科学学报》2023年第6期1091-1099,共9页Chinese Journal of Space Science

基　　金：国家自然科学基金项目(41806003);遥感科学国家重点实验室开放基金项目(OFSLRSS202219);国家重大科技基础设施项目“地球系统数值模拟装置”共同资助。

摘　　要：海洋锋作为海洋中两种不同性质的水体之间的边界,对渔业和海洋环境保护等许多领域有重要影响,如何快速准确实现海洋锋的自动检测和识别对于海洋监测和预报具有重要的科学意义。将深度学习图像分割网络与提取锋面特征的方法相结合,利用基于U-Net架构的实例分割模型,分别建立海洋锋区和锋面中心线的智能检测模型,同时在编解码过程中采用残差学习单元对模型特征提取网络进行改进。研究结果表明,锋面智能检测模型能够准确提取先前锋面检测算法所识别的锋区和锋面中心线特征,Dice系数分别达到了0.92和0.97,达到了很好的检测效果。同时,利用不同锋面阈值得到的样本数据对模型进行训练,比较结果表明,降低样本集阈值之后模型精度有了显著的提升。As a boundary of two water masses with different properties,oceanic fronts have impor-tant influences on many fields such as fishery,marine military and environmental protection.How to quickly and accurately implement automatic detection and identification of ocean front is of great scien-tific significance for ocean monitoring and forecasting.In this paper,the deep learning image segmenta-tion network is combined with the method of extracting frontal features,and the detection models of frontal area and frontal line are established by using U-Net architecture.Meanwhile,the residual unit is used to improve the feature extraction network in the processes of encoding and decoding.The results show that the deep learning frontal detection model can accurately extract the features of frontal area and frontal line.The Dice coefficients reach 0.92 and 0.97 respectively,achieving a good detection perfor-mance.In this paper,the model is trained by the sample data of different frontal thresholds.The com-parison results show that the accuracy of model is significantly improved after the threshold of sample set is reduced.

关 键 词：海洋锋 海表温度 深度学习 U-Net 

分 类 号：P731.1[天文地球—海洋科学]