农业大数据学报 ›› 2019, Vol. 1 ›› Issue (1): 67-77.doi: 10.19788/j.issn.2096-6369.190107

• 应用研究 • 上一篇    下一篇

卷积神经网络在高分辨率影像分类中的应用

李贤江1,陈佑启2,*(),邹金秋2,石淑芹1,郭涛3,蔡为民1,陈浩4   

  1. 1. 天津工业大学经济与管理学院,天津 300387
    2. 中国农业科学院农业资源与农业区划研究所,北京 100081
    3. 四川智图信息技术有限公司,成都 610081
    4. 天津市环境保护科学研究院,天津 120000
  • 收稿日期:2018-11-08 出版日期:2019-03-26 发布日期:2019-04-04
  • 通讯作者: 陈佑启 E-mail:chenyouqi@caas.cn
  • 作者简介:李贤江,男,硕士研究生,研究方向:土地利用与GIS工程;Email: <email>lixianjiang100@163.com</email>
  • 基金资助:
    教育部人文社科青年基金项目(16YJCZH082,16YJC630149);科技基础性工作专项项目(2013FY110900);中央级公益性科研院所专项资金(720-36);国土资源部公益性行业科研专项(201511010-8);天津市高等学校创新团队培养计划(TD13-5038);天津工业大学研究生创新计划(18138)

Application of Convolutional Neural Networks in High-Resolution Image Classification

Li Xianjiang1,Chen Youqi2,*(),Zou Jinqiu2,Shi Shuqin1,Guo Tao3,Cai Weimin1,Chen Hao4   

  1. 1. School of Economics and Management, Tianjin Polytechnic University, Tianjin 300387
    2. Institute of Agricultural Resources and Regional Planning of the Chinese Academy of Agricultural Sciences, Beijing 100081
    3. Sichuan Zhitu Information Technology Co., Ltd. Chendu 610081
    4. Tianjin Academy of Environmental Science, Tianjin 120000
  • Received:2018-11-08 Online:2019-03-26 Published:2019-04-04
  • Contact: Chen Youqi E-mail:chenyouqi@caas.cn

摘要:

目的 将CNN应用于高分辨率遥感影像的实际分类中,并与传统的分类方法进行对比分析,揭示出不同分类方法在高分辨率遥感影像中的分类精度和适用性问题。方法 采用最大似然、平行六面体、K-Means均值聚类和传统神经网络等四类常用的ENVI传统分类方法以及CNN分类法,并利用混淆矩阵和空间像元误差分析对不同分类方法的分类结果进行精度评价。结果 根据分类精度对比分析发现在传统的四种ENVI分类方法中,传统神经网络和最大似然法的分类精度相对较好,K-Means均值聚类和平行六面体的分类精度相对较差,CNN的分类精度整体上要高于ENVI传统分类方法的精度。结论 CNN在高分辨率遥感影像分类中能够较好地提取地物信息和地物的轮廓特征,在高分辨率遥感影像分类中具有良好的适用性。

关键词: 高分一号, 卷积神经网络, 遥感, 深度学习

Abstract:

[Objective] Convolutional neural networks were applied to the classification of high-resolution remote sensing images and compared with traditional classification methods to assess the classification accuracy and applicability of different classification methods for high-resolution remote image sensing. [Methods] The convolutional neural network classification method was compared with four commonly used traditional ENVI classification methods: maximum likelihood, parallelepiped, K-means mean clustering, and traditional neural network approaches. The confusion results were examined to evaluate the effectiveness of the different classification methods. [Results] Among the four traditional ENVI classification methods, the classification accuracy of the traditional neural network and maximum likelihood methods was relatively good, the classification accuracy of k-means mean clustering and parallelepiped methods was relatively poor, and the convolutional neural networks method achieved higher classification accuracy. [Conclusion] Convolutional neural networks can extract feature information in high-resolution remote sensing image classification, improving the classification accuracy of remote image sensing.

Key words: GF-1, Convolutional Neural Network, Remote Sensing, Deep Learning

中图分类号: 

  • TP399