测绘通报 ›› 2019, Vol. 0 ›› Issue (6): 29-33,40.doi: 10.13474/j.cnki.11-2246.2019.0179

• 学术研究 • 上一篇    下一篇

深度学习GoogleNet模型支持下的中分辨率遥感影像自动分类

陈斌1, 王宏志1, 徐新良2, 王首泰1, 张亚庆2   

  1. 1. 华中师范大学城市与环境科学学院地理过程分析与模拟湖北省重点实验室, 湖北 武汉 430079;
    2. 中国科学院地理科学与资源研究所资源与环境信息系统国家重点实验室, 北京 100101
  • 收稿日期:2018-09-03 修回日期:2018-11-07 出版日期:2019-06-25 发布日期:2019-07-01
  • 通讯作者: 徐新良。E-mail:xuxl@lreis.ac.cn E-mail:xuxl@lreis.ac.cn
  • 作者简介:陈斌(1992-),男,硕士生,研究方向为深度学习算法及GIS应用研究。E-mail:giscb@foxmail.com
  • 基金资助:

    中国科学院A类战略性先导科技专项(XDA20010302);华中师范大学研究生教育创新资助项目(2018CXZZ001)

The classification by medium resolution remote sensing images based on deep learning algorithm of GoogleNet model

CHEN Bin1, WANG Hongzhi1, XU Xinliang2, WANG Shoutai1, ZHANG Yaqing2   

  1. 1. Hubei Province Key Laboratory for Analysis and Simulation of Geographical Process, College of Urban and Environmental Sciences, Central China Normal University, Wuhan 430079, China;
    2. State Key Laboratory of Resources and Environmental Information System, Institute of Geographic Sciences and Natural Resources Research, CAS, Beijing 100101, China
  • Received:2018-09-03 Revised:2018-11-07 Online:2019-06-25 Published:2019-07-01

摘要:

提出了一种基于深度学习技术的遥感分类方法,它能有效解决中分辨率影像在分类过程中出现的像元混分问题。研究选用2016年5月12日武汉市Landsat 7 ETM+遥感影像,基于GoogleNet模型中的Inception V3网络结构,借助迁移学习方法,构建出遥感分类模型,实现了对武汉市主城区4类典型地物(不透水层、植被、水体和其他用地)的自动分类提取,并将分类结果与传统最大似然分类(ML)结果进行了对比分析。研究表明:基于深度学习方法的遥感影像总体分类精度高达88.33%,Kappa系数为0.834 2,明显优于传统ML方法总体分类精度83%和Kappa系数0.755 0,而且有效抑制了地物在分类过程中出现的像元混分现象。

关键词: 空间分辨率, 深度学习, 遥感分类, GoogleNet

Abstract:

We proposed a remote sensing classification method based on deep learning technology, which can effectively solve the problem of pixel mixing in the medium resolution images classification. The research selected the Landsat 7 ETM+ remote sensing image of Wuhan City on May 12, 2016. Based on the Inception V3 network structure in the GoogleNet model, the remote sensing image classification model was constructed by means of migration learning method, and four typical features of the main urban area of Wuhan were realized. Automatic classification of permeable layers, vegetation, water bodies and other land uses, and the classification results were compared with traditional maximum likelihood classification (ML) results. The research shows that the overall classification accuracy of remote sensing image based on deep learning method is as high as 88.33%, and the Kappa coefficient is 0.834 2, which is obviously better than the traditional classification accuracy of 83% and Kappa coefficient of 0.755 0, and it effectively suppresses the phenomenon of pixel mis-or leakage in the classification process.

Key words: spatial resolution, deep learning, remote sensing classification, GoogleNet

中图分类号: