基于YOLOv5s模型的北极气旋目标检测方法  

Arctic cyclone detection method based on the YOLOv5s model

作　　者：王钰坤 谢涛 向儒萱毅 张雪红 白淑英 WANG Yukun;XIE Tao;XIANG Ruxuanyi;ZHANG Xuehong;BAI Shuying(School of Remote Sensing and Geomatics Engineering,Nanjing University of Information Science and Technology,Nanjing 210044,China;Laboratory for Regional Oceanography and Numerical Modeling,Pilot National Laboratory for Marine Science and Technology,Qingdao 266237,China;Technology Innovation Center for Integration Applications in Remote Sensing and Navigation,Ministry of Natural Resources,Nanjing 210044,China;Jiangsu Province Engineering Research Center of Collaborative Navigation/Positioning and Smart Application,Nanjing 210044,China)

机构地区：[1]南京信息工程大学遥感与测绘工程学院,江苏南京210044 [2]青岛海洋科学与技术试点国家实验室区域海洋动力学与数值模拟功能实验室,山东青岛266237 [3]自然资源部遥感导航一体化应用工程技术创新中心,江苏南京210044 [4]江苏省协同精密导航定位与智能应用工程研究中心,江苏南京210044

出　　处：《极地研究》2025年第1期1-10,共10页Chinese Journal of Polar Research

基　　金：国家重点研发计划(2021YFC2803302、2022YFC3004200/2022YFC3004202、2022YFC3104900/2022YFC3104905);国家自然科学基金(42176180);江苏省自然资源发展专项资金(海洋科技创新)项目(JSZRHYKJ202114);江苏省研究生科研创新计划(KYCX23_1345)资助。

摘　　要：北极气旋是影响北极环境的主要天气系统之一,对其准确识别对北极航海保障具有至关重要的意义。本文基于欧洲中期天气预报中心推出的第五代大气再分析产品(ERA5)平均海平面气压数据,构建北极气旋目标检测的数据集,训练北极气旋识别的深度学习目标检测You Only Look Once version5(YOLOv5s)模型,并与其他深度学习目标检测模型Single Shot Multibox Detector(SSD)、Faster R-CNN、YOLOv4进行性能的验证与比较。试验结果表明,YOLOv5s的检测精确率、平均准确率和刷新频率分别为95.26%、98.09%和64.85 s–1,刷新频率较其余4种模型中效果最好的SSD模型提高了15.65 s–1。YOLOv5s模型检测速度快、精度高且具有更好的识别能力,能够有效识别北极气旋目标,对北极气旋的识别具有较好的应用前景。因此,本文选取2021年冬季、夏季的北极气旋进行个例分析。结果显示,冬季3月8日12时产生的北极气旋生命周期为48 h,平均强度为54.62,夏季9月21日6时的气旋生命周期为84 h,平均强度为42.82,符合北极气旋生命周期夏长冬短,气旋强度冬强夏弱的特点。YOLOv5s模型为北极气旋的检测识别提供了新方法和新思路。Arctic cyclones are one of the main weather systems to affect the Arctic environment.Accurate identification of Arctic cyclones is of crucial importance to Arctic navigational support.Based on mean sea level pressure data from the new generation European Center for Medium-range Weather Forecasts ERA5 reanalysis product,this study constructed a dataset for Arctic cyclone target detection and trained the deep learning target detection You Only Look Once version 5(YOLOv5s)model for Arctic cyclone recognition,the performance of which was verified through comparison with other deep learning object detection models(i.e.,the Single Shot Multibox Detector,Faster-RCNN,and You Only Look Once version 4 models).Experimental results showed that compared with other deep learning models,the detection accuracy,average accuracy,and average detection speed of YOLOv5s were 95.26%,98.09%,and 64.85 s-1,respectively,and that detection speed increased by 15.65 s–1.This improved performance means that YOLOv5s can effectively identify Arctic cyclone targets.Based on the detection and recognition results of the YOLOv5s model,Arctic cyclones that occurred in winter and summer 2021 were selected for case analysis.The life cycle and average intensity of an Arctic cyclone generated at 12:00 on 8 March(winter)were 48 hours and 54.62,respectively,and those of an Arctic cyclone generated at 06:00 on 21 September(summer)were 84 h and 42.82,respectively.These findings are consistent with observed Arctic cyclones,i.e.,long and weak in summer,short and strong in winter.The YOLOv5s model provides new approaches and ideas for detection and identification of Arctic cyclones.

关 键 词：北极 北极气旋识别 ERA5 深度学习 YOLOv5s 模型 

分 类 号：P731.2[天文地球—海洋科学]