摘要
深度学习在计算机视觉领域取得了显著的成果,如图像分类、人脸识别、图像检索等。对于遥感领域而言,获取用于训练CNN的有标签数据集通常是一个重大挑战。本文研究了如何将CNN用于高分辨率遥感影像的场景分类,为了克服缺乏大量有标签遥感影像数据集的问题,结合CNN采用了两种技术:数据增广和迁移学习。在UC Merced Land Use数据集上,验证了VGG16、VGG19、Res Net50、InceptionV3、Dense Net121等5种网络的性能,分别达到了98.10%、96.19%、99.05%、97.62%、99.52%的分类准确率。
Deep learning has achieved remarkable results in the field of computer vision,such as image classification,face recognition,image retrieval and so on.For remote sensing,obtaining a labeled dataset for training DCNN is often a major challenge.In this paper,the use of DCNN for scene classification in high-resolution remote sensing imagery is investigated.In order to overcome the lack of a large number of labeled remote sensing image datasets,two technologieswere combined with DCNN:data augmentation and transfer learning.On the UC Merced Land Use dataset,the performances of 5 networks including VGG16,VGG19,Res Net50,Inception V3,and Dense Net121 were verified,which achieved classification accuracy of 98.10%,96.19%,99.05%,97.62%,and99.52%,respectively.
作者
乔婷婷
李鲁群
QIAO Tingting;LI Luqun(School of Information and Mechanical Engineering,Shanghai Normal University,Shanghai 201400,China)
出处
《测绘通报》
CSCD
北大核心
2020年第2期37-42,共6页
Bulletin of Surveying and Mapping
基金
上海教委重点项目(304-AC9103-19-368405029)
教育部产学合作协同育人项目(309-C-6105-18-060).
关键词
高分辨率遥感影像
场景分类
卷积神经网络
数据增广
迁移学习
high-resolution remote sensing imagery
scene classification
convolutional neural network
data augmentation
transfer learning
