摘要
台风预测可为台风预警预报提供先验信息,辅助相关部门进行科学决策,以减少灾害损失。利用时间序列台风卫星云图,提出一种新的台风等级预测模型SeqTyphoon,将注意力机制和序列到序列引入模型预测未来时刻台风图像,然后利用卷积神经网络对预测的台风图像进行台风等级预测。通过日本气象厅发布的1981—2017年3万多张时序台风卫星云图,构建了训练集、验证集和测试集,分别对应29519、3804、1995张台风图像。针对SeqTyphoon模型,分别进行了台风云图的不同时间间隔、不同预测时长及不同空间分辨率对台风图像预测精度影响的对比实验。实验结果表明,台风云图均为32像素×32像素,时间间隔为6 h比时间间隔为12 h的训练集和验证集的均方根误差分别降低5.41%、5.72%,前者训练集的均方根误差达到0.0922,验证集为0.0954,前者台风等级预测准确率为后者的2倍;台风云图为32像素×32像素,时间间隔为6 h时,预测未来6~48 h的台风图像,训练集和验证集的均方根误差均递增,台风等级预测准确率递减;时间间隔为6 h,图像为64像素×64像素的训练集的均方根误差为0.0896,验证集为0.0911,台风等级预测总体准确率为83.2%。综上,影响台风图像的最主要因素是相邻台风云图的时间间隔,其次是预测时长与空间分辨率大小。
Typhoon prediction seeks to provide prior information for relevant departments to make scientific decisions and to reduce disaster losses.In this paper,a new typhoon grade prediction model,SeqTyphoon,is proposed based on time series typhoon satellite cloud images.Firstly,attention mechanism is introduced into the sequence to sequence model to predict future typhoon images.Then,to measure the quality of predicted images,further prediction of its typhoon grade with convolutional neural network is done.More than 30000 typhoon satellite cloud images released by the Japan Meteorological Agency from 1981 to 2017 are divided into training set,verification set and test set,corresponding to the typhoon images of 29519,3804 and 1995.In SeqTyphoon model,comparative experiments are carried out in different time intervals,different prediction durations and different spatial resolutions,which reflect on the prediction accuracy of typhoon images.The experimental results show that,when pixels of typhoon images are 32×32,RMSE of training and verification set of the 6 h time interval are lower than 12 h time interval,which is 5.41%,5.72%,respectively.RMSE of the former training set and the verification set reaches 0.0922,0.0954,respectively,and typhoon grade prediction accuracy of the former is twice than that of the latter.When typhoon image pixels are 32×32 and time interval is 6 h,typhoon images of future 6~48 h are predicted.RMSE of the training and verification set increases progressively and grade prediction accuracy decreases relatively.The time interval is 6 h,RMSE of the training set of the image pixel 64×64 is 0.0896,RMSE of the verification set is 0.0911,and the overall accuracy of the typhoon grade prediction is 83.2%.In summary,the most significant factor affecting typhoon image prediction is time interval of the adjacent typhoon picture,followed by the prediction duration and the spatial resolution.
作者
郑宗生
刘敏
胡晨雨
傅泽平
卢鹏
姜晓轶
ZHENG Zongsheng;LIU Min;HU Chenyu;FU Zeping;LU Peng;JIANG Xiaoyi(College of Information,Shanghai Ocean University,Shanghai 201306,China;National Marine Information Center,Tianjin 300171,China)
出处
《遥感信息》
CSCD
北大核心
2020年第4期16-22,共7页
Remote Sensing Information
基金
国家自然科学基金项目(41671431)
上海市科委地方高校能力建设项目(15590501900)
国家海洋局数字海洋科学技术重点实验室开放基金项目(B201801034)。
关键词
时间序列
台风卫星云图
注意力机制
序列到序列
图像预测
time series
typhoon satellite image
attention mechanism
sequence to sequence
image prediction
