基于CNN-typhoon模型的全球台风报文收集方法研究  

Global Typhoon Message Collection Method Based on CNN-typhoon Model

作　　者：韩瑞 顾春利 李哲[4,5] 伍康 高峰[1] 沈文海 HAN Rui;GU Chun-li;LI Zhe;WU Kang;GAO Feng;SHEN Wen-hai(National Meteorological Information Center,China Meteorological Administration,Beijing 100081,China;School of Environment,Tsinghua University,Beijing 100084,China;Beijing Institute of Applied Meteorology,Beijing 100029,China;Department of Precision Instruments,Tsinghua University,Beijing 100084,China;Beijing Xiaopeng Automobile Co.,Ltd.,Beijing 100084,China)

机构地区：[1]中国气象局国家气象信息中心,北京100081 [2]清华大学环境学院,北京100084 [3]北京应用气象研究所,北京100029 [4]清华大学精密仪器系,北京100084 [5]北京小鹏汽车有限公司,北京100084

出　　处：《计算机科学》2020年第S02期11-17,共7页Computer Science

基　　金：国家重点研发计划(2016YFA0600301);上海台风研究基金项目(重点学科领域)(TFJJ201904);国家自然科学基金项目(91744209)。

摘　　要：台风是一种高影响的强对流天气系统。台风报文资料作为提供台风初值的最初来源,对改进台风预报准确性有一定的帮助,因此做好全球台风的报文快速识别收集工作至关重要。针对现有全球台风报文实时性差、延迟高、报文被动接收的问题,本研究利用MSG,Meteosat-5,MTSAT,GOES-W,GOES-E卫星图像数据,通过训练2006年1月-2020年8月的1351次全球热带气旋过程,共计8983张红外卫星图像,基于深度学习算法,提出了一种CNN-typhoon模型,可以对无台风、台风生成、台风最强等3种图像进行识别分类。实验证明:CNN-typhoon模型训练集的识别精度可接近100%,验证集精度高于88.1%;同时将模型代入模拟业务,在一定时段内增加了接近31.0%的报文收集种类,报文收集时效提高了23.5倍。Typhoon is a highly convective weather system with high impact.As the source of typhoon initial values,typhoon message data is helpful to improve the accuracy of typhoon forecasts.Therefore,it is very important to do a good job in the rapid identification and collection of global typhoon messages.Aiming at the problems of poor real-time performance,high latency,and passive message reception of global typhoon messages,this study uses MSG,Meteosat5,MTSAT,GOES-W,and GOES-E satellite 8983 infrared satellite images of 1351 typhoon processes from January 2006 to August 2020.Based on the deep learning algorithm,a CNN-typhoon model is proposed,which can identify and classify three types of images:no typhoon,typhoon generation,and strongest typhoon.Experiments have proved that the recognition accuracy of the CNN-typhoon model training set can be close to 100%,and the verification set accuracy is higher than 88.1%.At the same time,the model is substituted into the simulation service,and within a certain period of time,nearly 31.0%of the message collection types are increased,saving the message collection timeliness is improved by 23.5 times.

关 键 词：卷积神经网络(CNN) 台风报文 识别 分类 收集 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]