Please wait a minute...
img

官方微信

遥感技术与应用  2020, Vol. 35 Issue (1): 202-210    DOI: 10.11873/j.issn.1004-0323.2020.1.0202
数据与图像处理     
基于迁移学习及气象卫星云图的台风等级分类研究
郑宗生(),胡晨雨(),黄冬梅,邹国良,刘兆荣,宋巍
上海海洋大学 信息学院,上海 201306
Research on Transfer Learning Methods for Classification of Typhoon Cloud Image
Zongsheng Zheng(),Chenyu Hu(),Dongmei Huang,Guoliang Zou,Zhaorong Liu,Wei Song
School of Information, Shanghai Ocean University, Shanghai 201306, China
 全文: PDF(2238 KB)   HTML
摘要:

针对传统卫星云图特征提取方法复杂且深度卷积神经网络(Deep Convolutional Neural Network, DCNN)模型开发困难的问题,提出一种基于参数迁移的台风等级分类方法。利用日本气象厅发布的近40 a 10 000多景台风云图数据,构建了适应于迁移学习的台风云图训练集和测试集。在大规模ImageNet源数据集上训练出3种源模型VGG16,InceptionV3和ResNet50,依据台风云图低层特征与高层语义特征的差异,适配网络最佳迁移层数并冻结低层权重,高层权重采用自适应微调策略,构建出了适用于台风小样本数据集的迁移预报模型T-typCNNs。实验结果表明:T-typCNNs模型在自建台风数据集上的训练精度为95.081%,验证精度可达91.134%,比利用浅层卷积神经网络训练出的精度高18.571%,相比于直接用源模型训练最多提高9.819%。

关键词: 台风等级迁移学习深度卷积神经网络迁移层数自适应微调    
Abstract:

Aiming at the complexity of traditional methods for feature extraction about satellite cloud images, and the difficulty of developing deep convolutional neural network from scratch, a parameter-based transfer learning method for classifying typhoon intensity is proposed. Take typhoon satellite cloud images published by Japan Meteorological Agency, which includes 10 000 scenes among nearly 40 years to construct training and test typhoon datasets. Three deep convolutional neural networks, VGG16, InceptionV3 and ResNet50 are trained as source models on the large-scale ImageNet datasets. Considering the discrepancy between low-level features and high-level semantic features of typhoon cloud images, adapt the optimal number of transferable layers in neural networks and freeze weights of low-level network. Meanwhile, fine-tune surplus weights on typhoon dataset adaptively. Finally, a transferred prediction model which is suitable for small sample typhoon datasets, called T-typCNNs is proposed. Experimental results show that the T-typCNNs can achieve training accuracy of 95.081% and testing accuracy of 91.134%, 18.571% higher than using shallow convolutional neural network, 9.819% higher than training with source models from scratch.

Key words: Typhoon grade    Transfer learning    Deep convolutional neural network    Number of transferable layers    Adaptive fine-tuning
收稿日期: 2018-09-25 出版日期: 2020-04-01
ZTFLH:  TP79  
基金资助: 国家自然科学基金项目(41671431);上海市科委地方院校能力建设项目(17050501900);国家海洋局数字海洋科学技术重点实验室开放基金项目共同资助
通讯作者: 胡晨雨     E-mail: zszheng@shou.edu.cn;1105814265@qq.com
作者简介: 郑宗生(1979-),男,河北唐山人,博士,副教授,主要从事海洋信息化及深度学习、迁移学习应用方面的研究。E?mail:zszheng@shou.edu.cn
服务  
把本文推荐给朋友
加入引用管理器
E-mail Alert
RSS
作者相关文章  
郑宗生
胡晨雨
黄冬梅
邹国良
刘兆荣
宋巍

引用本文:

郑宗生,胡晨雨,黄冬梅,邹国良,刘兆荣,宋巍. 基于迁移学习及气象卫星云图的台风等级分类研究[J]. 遥感技术与应用, 2020, 35(1): 202-210.

Zongsheng Zheng,Chenyu Hu,Dongmei Huang,Guoliang Zou,Zhaorong Liu,Wei Song. Research on Transfer Learning Methods for Classification of Typhoon Cloud Image. Remote Sensing Technology and Application, 2020, 35(1): 202-210.

链接本文:

http://www.rsta.ac.cn/CN/10.11873/j.issn.1004-0323.2020.1.0202        http://www.rsta.ac.cn/CN/Y2020/V35/I1/202

台风等级分类 最大风速(10 min平均值)
kt m/s km/h
热带低压(TD) <33 <17 <62
热带风暴(TS) ≥33~<48 ≥17~<25 ≥62~<89
强热带风暴(STS) ≥48~<64 ≥25~<33 ≥89~<118
台风(TY) ≥64~<85 ≥33~<42 ≥118~<150
强台风(STY) ≥85 ≥42 ≥150
表1  台风等级标准
图1  台风云图样本注:(a)为一级台风部分数据样本,(b)为二级台风部分数据样本,(c)为三级台风部分数据样本,(d)为四级台风部分数据样本,(e)为五级台风部分数据样本
图2  深度卷积神经网络
图3  研究方法流程图
模型 原始图像 参数 网络深度 大小
VGG16 224 × 224 × 3 138 357 544 23 528 MB
InceptionV3 299 × 299 × 3 23 851 784 159 92 MB
ResNet50 224 × 224 × 3 25 636 712 168 99 MB
表2  3种源模型参数配置
图4  T-typCNNs迁移模型结构
网络结构

代码

层数

网络参数 训练精度/% 测试精度/%

时间

/epoch

CNN_8 6 35 474 75.314 72.563 71 s
VGG16 20 138 357 544 81.647 74.031 153 s
InceptionV3 311 23 851 784 85.033 78.472 304 s
ResNet50 175 25 636 712 86.951 81.315 212 s
表3  不同深度模型仅迁移网络结构的分类性能
网络结构 训练精度/% 测试精度/% 训练时间/s
VGG16 83.131 81.092 137s/epoch
InceptionV3 87.772 84.726 253s/epoch
ResNet50 92.649 88.611 176s/epoch
表4  深度模型迁移整体网络结构和参数的分类性能
图6  迁移InceptionV3模型各层的分类精度
图5  迁移VGG16模型各层的分类精度
图7  迁移ResNet50模型各层的分类精度
网络结构 最佳冻结层数 训练精度/% 测试精度/% 训练时间/s
VGG16 12 89.323 82.712 146s/epoch
InceptionV3 223 92.744 87.398 294s/epoch
ResNet50 110 95.081 91.134 182s/epoch
表5  自适应微调后的各模型性能
图8  T-typCNNs模型准确率曲线
图9  T-typCNNs模型损失值曲线
图10  不同样本数下的模型性能
1 Liao Fei , Li Wenting , Zhang Ziran ,et al .Statistical Characteristics of Rapid Changes in the Intensity and Track of Tropical Cyclones in the South China Sea from 1949 to 2017[J].Journal of Ocean,2019,41(9):126-135.廖菲,李文婷,张子然, 等 .1949~2017年南海海域热带气旋强度和路径快速变化统计特征[J].海洋学报,2019,41(9):126-135.
2 Yu Jinhua , Sheng Siwei .Comparison of Tropical Cyclone Intensity Data in the Northwest Pacific[J].Meteorological Science,2010,30(6):778-784.余锦华,盛思伟.西北太平洋热带气旋强度资料的对比[J].气象科学,2010,30(6):778-784.
3 Yin Chengtuan , Zhang Jinshan , Xiong Mengjie ,et al .Analysis on the Disaster trend of typhoon and tide in south China sea [J].Journal of Tropical Oceanograpy,2019,38(1):38-45.殷成团,张金善,熊梦婕, 等 .我国南海沿海台风及暴潮灾害趋势分析[J].气象科学,2019,38(1):38-45.
4 Tang Tingting ,LiQingxi, Li Guangxin ,et al .Research on Typhoon Intensity Prediction Statistical Model based on Meteorological Big Data[J].Integration Technology,2016,5(2):73-84.汤婷婷,李晴岚,李广鑫, 等 .基于气象大数据的台风强度预测统计模型研究[J].集成技术,2016,5(2):73-84.
5 Zhao Xiaoli .Research on Cloud Image Cloud Detection and Classification of Meteorological Satellites[D].Guangzhou:South China University of Technology,2014.赵晓利. 气象卫星云图云检测及分类的研究[D].广州:华南理工大学,2014.
6 Tian Wenzhe , Fu Randi , Jin Wei ,et al .Adaptive Fuzzy Support Vector Machine for Classification of Clouds in Satellite Imagery[J].Journal of Wuhan University,2017,42(4):488-495.陈田文哲,符冉迪,金炜, 等 .面向卫星云图云分类的自适应模糊支持向量机[J].武汉大学学报,2017,42(4):488-495.
7 Cao Minjie , Liu Zenghong , Wang Zhenfeng ,et al .A Review of Real Time Marine Monitoring and Its Application in Typhoon Area[J].Advances in Meteorological Technology,2017(4):47-52.曹敏杰,刘增宏,王振峰, 等 .台风海域实时海洋监测及其应用研究综述[J].气象科技进展,2017(4):47-52.
8 Shi Lanhong , Cui Linli , Zhao Bingke ,et al .Study on the Vertical Distribution of Cloud Properties in the Eye Sall and Surrounding Spiral Cloud Belt of Typhoon[J].Journal of Tropical Meteorological,2015(1):53-64.史兰红,崔林丽,赵兵科, 等 .台风眼壁及周围螺旋云带云属性垂直分布研究[J].热带气象学报,2015(1):53-64.
9 Dvorak V F .Tropical Cyclone Intensity Analysis Usingsatellite Data[R].Noaa Technical Report, 1984, NESDUS,11.
10 Yang Wei , Ouyang Sida , Fan Kuikui ,et al .Classification of Remote Sensing Image based on BP Neural Network based on Double Tree Complex Wavelet Decomposition[J].Remote Sensing Technology and Application,2018,33(2):313-320.杨朦朦,欧阳斯达,范奎奎, 等 .基于双树复小波分解的BP神经网络遥感影像分类[J].遥感技术与应用,2018,33(2):313-320.
11 Zhao Liangliang .Research and Application of Cloud Image Recognition and Ultra-short-term Direct Solar Radiation Prediction based on Neural Network[D].Nanjing:Southeast University,2017.赵亮亮.基于神经网络的云图识别与超短期直接太阳光辐射预测研究及应用[D].南京:东南大学,2017.
12 Cui Xianliang , Chen Lifu , Xing Xuemin ,et al .Remote Sensing Image Scene Classification based on Frequency Band Feature Fusion and GL-CNN. Remote Sensing Technology and Application, 2019, 34(4): 712-719. [崔先亮,,陈立福,邢学敏,等. 基于频带特征融合的GL-CNN遥感图像场景分类[J].遥感技术与应用,2019,34(4):712-719.]
13 Zhuang Fuzhen , Luo Ping , He Qing ,et al .Progress in Migration Learning Research[J].Journal of Software,2015,26(1):26-39.庄福振,罗平,何清, 等 .迁移学习研究进展[J].软件学报,2015,26(1):26-39.
14 Wang Wenpeng , Mao Wentao , He Jianliang ,et al .Smoke Recognition Method based on Deep Migration Learning[J]Journal of Computer Applications,2017,37(11):3176-3181.王文朋,毛文涛,何建樑, 等 .基于深度迁移学习的烟雾识别方法[J].计算机应用,2017,37(11):3176-3181.
15 Huang Jie , Jiang Zhiguo , Zhang Haopeng ,et al .Ship Target Detection based on Convolutional Neural Network for Remote Sensing Image[J].Journal of Beijing Univeristy of Aeronautics,2017,43(9):1841-1848.黄洁,姜志国,张浩鹏, 等 .基于卷积神经网络的遥感图像舰船目标检测[J].北京航空航天大学学报,2017,43(9):1841-1848.
16 Sun Chao , Lu Junwei , Liu Feng ,et al .Research on Infraredimage Super-resolution Method based on Migration Learning[J].Laser & Infrared,2017,47(12):1559-1564.孙超,吕俊伟,刘峰, 等 .基于迁移学习的红外图像超分辨率方法研究[J].激光与红外,2017,47(12):1559-1564.
17 Long M S , Cao Y , Wang J M ,et.al .Learning Transferable Features with Deep Adaptation Networks[C]∥In International Conference on Machine Learning,2015:97-105.
18 Afridi M J , Ross A , Shapiro E M .On Automated Source Selection for Transfer Learning in Convolutional Neural Networks[J].Pattern Recognition,2017,73:65-75.
19 Gao Song , Zhao Peng , Pan Bin ,et al .The Typhoon Path Nowcasting Model based on LSTM Neural Network[J].Acta Oceanologica Sinica,2016,37(5):8-12.高松,赵鹏,潘斌, 等 .基于长短时记忆神经网络的台风路径临近预报模型[J].海洋学报,2016,37(5):8-12.
20 Simonyan K , Zisserman A .Very Deep Convolutional Networks for Large-scale Image Recognition[J].Computer Science,2014,arxiv:1409.1556.
21 Szegedy C , Vanhoucke V , Ioffe S ,et.al .Rethinking the Inception Architecture for Computer Vision[C]∥In International Conference on Learning Representation,Computer Science,2015:2818-2826.
22 He K , Zhang X , Ren S ,et al .Deep Residual Learning for Image Recognition[C]∥Computer Vision and Pattern Recognition,2015:770-778.
23 Yosinski J , Clune J , Bengio Y ,et al .How Transferable are Features in Deep Neural Networks?[C]∥Advances in Neural Information Processing Systems,2014:3320-3328.
[1] 崔刚,吴金胜,于镇,周玲. 深度学习对不同分辨率影像冬小麦识别的适用性研究[J]. 遥感技术与应用, 2019, 34(4): 748-755.