地球信息科学学报 ›› 2018, Vol. 20 ›› Issue (11): 1562-1570.doi: 10.12082/dqxxkx.2018.180159

• 地球信息科学理论与方法 • 上一篇    下一篇

基于全卷积神经网络的建筑物屋顶自动提取

刘文涛1(), 李世华1, 覃驭楚2,*()   

  1. 1. 电子科技大学资源与环境学院,成都 611731
    2. 中国科学院遥感与数字地球研究所 遥感科学国家重点实验室,北京 100101
  • 收稿日期:2018-04-02 修回日期:2018-09-14 出版日期:2018-11-20 发布日期:2018-11-28
  • 通讯作者: 覃驭楚 E-mail:liuwentaoboy@126.com;qinyc@radi.ac.cn
  • 作者简介:

    作者简介:刘文涛(1989- ),男,硕士,主要研究方向为计算机视觉。E-mail: liuwentaoboy@126.com

  • 基金资助:
    中国科学院百人计划项目(Y6YR0700QM);国家自然科学基金项目(41471294)

Automatic Building Roof Extraction with Fully Convolutional Neural Network

LIU Wentao1(), LI Shihua1, QIN Yuchu2,*()   

  1. 1. School of Resources and Environment, University of Electronic Science and Technology of China, Chengdu 611731, China
    2. State Key Laboratory of Remote Sensing Science, Institute of Remote Sensing and Digital Earth, Chinese Academy of Science, Beijing 100101, China
  • Received:2018-04-02 Revised:2018-09-14 Online:2018-11-20 Published:2018-11-28
  • Contact: QIN Yuchu E-mail:liuwentaoboy@126.com;qinyc@radi.ac.cn
  • Supported by:
    100 Talents Program of the Chinese Academy of Sciences, No.Y6YR0700QM;National Natural Science Foundation of China, No.41471294.

摘要:

高分辨率遥感影像在地面自动目标提取中得到了广泛应用,然而利用传统算法,很难高精度地进行实时的建筑物屋顶绘图。本文使用深度学习方法探讨建筑物屋顶分割,由于卷积运算对形变、旋转、光照条件的不敏感,设计了一种用于建筑物屋顶提取的深度卷积神经网络,提出的网络为级联式全卷积神经网络,在深度卷积神经网络的设计中使用了特征复用和特征增强,实现建筑物的自动精确提取。以美国马萨诸塞州建筑物数据集为基础的实验结果表明,本文提出的网络结构取得了92.3%的总体预测精度,和其他方法相比,本文提出的方法具有更高的精度

关键词: 遥感图像, 建筑物, 深度学习, 卷积神经网络, 自动提取

Abstract:

The very high resolution remotely sensed imagery has been widely applied in automatic extraction of ground objects, However, it is hard to conduct timely operational building roof mapping with high accuracy using conventional algorithms. This paper investigate building roof segmentation with deep learning method, since the convolution operations upon images is not sensitive to deformation,rotation and illumination condition, a Deep Convolutional Neural Network (DCNN) is designed for building roof extraction, the proposed network has a cascaded structure with fully convolutional layers, with strategies for feature reuse and enhancement in the design of DCNN, it is expected to accurately extract building roof. The experiment is carried out upon building sample data set acquired in Massachusetts, USA, the results show that the proposed network achieved overall accuracy of 92.3%, and the comparison with other methods suggest the proposed network is able to map building roof with high accuracy.

Key words: remote sensing image, building, deep learning, convolutional neural network, automatic extraction