改进残差网络及时序气象卫星云图的台风等级分类研究  

Typhoon classification research based on improved residual network and temporal meteorological satellite cloud images

作　　者：郑宗生[1] 傅泽平 刘敏 胡晨雨 卢鹏[1] 姜晓轶[2] ZHENG Zongsheng;FU Zeping;LIU Min;HU Chenyu;LU Peng;JIANG Xiaoyi(School of Information,Shanghai Ocean University,Shanghai 201306,China;National Marine Information Center,Tianjin 300171,China)

机构地区：[1]上海海洋大学信息学院,上海201306 [2]国家海洋信息中心,天津300171

出　　处：《测绘工程》2023年第6期10-16,共7页Engineering of Surveying and Mapping

基　　金：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)。

摘　　要：结合了传统残差网络在数据样本小环境下分支卷积层特征的浪费问题,且考虑台风云图时空关联性强、特征复杂因素,参考日本国立情报学研究所在西北太平洋上空通过数个气象卫星拍摄的8000多景高分辨率台风云图,编制了适应于残差神经网络的时序台风云图分类训练集和测试集。为满足数据集及台风特征,有效优化了原始残差网络的残差块,并得到了恒等映射残差块。经由增加卷积输出来促进分支通路更好的被利用,保留台风图像时序特性,提高网络性能。实验显示,W-ResNets模型在自建台风数据集上的训练精度达到99.60%,测试精度达到76.19%,相较于浅层卷积神经网络的测试精度高出8.48%,相比于使用传统的残差神经网络提高了2.87%,为进一步验证模型的泛化性能,采用MNIST通用数据集进行网络对比实验,宽残差网络得到98.786%的测试精度,优于传统残差网络。文中的W-ResNets预报模型可在小样本台风数据集推广使用。Combining with the waste problem of the features of the branch convolution layer of the traditional residual network in the small environment of data samples,and considering the factors of strong spatial and temporal correlation and complex characteristics of typhoon cloud images,the high-resolution typhoon cloud images of more than 8,000 views taken by several meteorological satellites over the northwest Pacific Ocean by National Institute of Informatics(NII)in Japan were referred to.A training set and a test set for classification of temporal typhoon nephogram adapted to residual neural network are developed.In order to satisfy the data set and typhoon characteristics,the residual block of the original residual network is optimized effectively,and the identity mapping residual block is obtained.By increasing the convolutional output,the branch channel is promoted to be better utilized,and the timing characteristics of typhoon image are preserved to improve the network performance.Experiments show that the training accuracy of W-ResNets model on the self-built typhoon data set is up to 99.60%,and the test accuracy is up to 76.19%,which is 8.48%higher than that of the shallow convolutional neural network and 2.87%higher than that of the traditional residual neural network.In order to further verify the generalization performance of the model,MNIST general data set was used for network comparison experiments,and the wide residual network obtained a test accuracy of 98.786%,which was better than the traditional residual network.The W-ResNets prediction model can be generalized to small sample typhoon data sets.

关 键 词：卷积神经网络 残差神经网络 宽残差网络 小样本台风数据集 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] TP751[自动化与计算机技术—控制科学与工程]