一种基于深度学习的台风强度估测技术  

A Deep Learning Technique of Typhoon Intensity Estimation

作　　者：钱奇峰 王川[2] 徐雅静[2] 周冠博 刘达 聂高臻 QIAN Qifeng;WANG Chuan;XU Yajing;ZHOU Guanbo;LIU Da;NIE Gaozhen(National Meteorological Centre,Beijing 100081;Beijing University of Posts and Telecommunications,Beijing 100876)

机构地区：[1]国家气象中心,北京100081 [2]北京邮电大学,北京100876

出　　处：《气象》2021年第5期601-608,共8页Meteorological Monthly

基　　金：广东省重点领域研发计划重点专项项目(2019B111101002);国家重点研发计划(2017YFC1501604)共同资助。

摘　　要：台风客观定强是提高台风业务现代化水平的重要支撑技术,深度学习通过机器对大量样本的分析和学习,能够隐式提取图像中深层抽象的复杂特征,越来越多地被应用到气象领域中。本文利用ResNet深度学习模型,采用预训练后迁移学习的方式,以2005—2018年西北太平洋及南海台风的卫星云图为样本,构建了一种自动、客观的台风强度估测技术。通过对2019年全年的业务台风云图的检验分析,结果表明利用该技术能够实现对不同强度、不同发展阶段的台风客观强度估测,且对2019年全年独立样本估测的平均绝对误差和均方根误差分别为4.3 m·s^(-1)和5.5 m·s^(-1),精度优于传统客观定强方法,具有一定的业务应用价值。Typhoon objective strength determination is an important supporting technique to improve the modernization level of typhoon forecasting operation.Deep learning can implicitly extract the deep complex features in the images through the learning of a large number of samples,and it has been increasingly applied to the meteorological field nowadays.In this paper,a ResNet deep learning model is used to study the satellite cloud images as samples by pre-training and transfer-learning.After studying the 2005-2018 typhoon images of the Northwest Pacific and South China Sea,we consturct an automatic and objective typhoon intensity estimation technique.By using the deep learning technique to analyze the typhoon satellite images in 2019,we find that this technique can be used to estimate the objective intensity of typhoon in different intensity and different developing stages,and the mean absolute error(MAE)and root mean square error(RMSE)of independent samples in 2019 are 4.3 m·s^(-1) and 5.5 m·s^(-1) respectively.The accuracy is better than that of the traditional objective intensity estimation method,so it has certain application values.

关 键 词：深度学习 卫星云图 台风定强 ResNet网格 

分 类 号：P456[天文地球—大气科学及气象学] P457