测绘通报 ›› 2021, Vol. 0 ›› Issue (11): 65-69,75.doi: 10.13474/j.cnki.11-2246.2021.340

• 学术研究 • 上一篇    下一篇

深度学习PaddlePaddle框架支持下的遥感智能视觉平台研究与实现

孙玉梅1, 刘昱豪2, 边占新1, 孙亮1, 陈敬周1   

  1. 1. 石家庄铁路职业技术学院, 河北 石家庄 050018;
    2. 中科北纬(北京)科技有限公司, 北京 100043
  • 收稿日期:2021-05-19 修回日期:2021-09-20 出版日期:2021-11-25 发布日期:2021-12-02
  • 通讯作者: 边占新。E-mail:bian0080@sina.com
  • 作者简介:孙玉梅(1975-),女,博士,副教授,主要研究方向为空间大数据分析、遥感图像处理、人工智能。E-mail:sunyumei2008@126.com

Research and implementation of remote sensing intelligent vision platform based on deep learning PaddlePaddle framework

SUN Yumei1, LIU Yuhao2, BIAN Zhanxin1, SUN Liang1, CHEN Jingzhou1   

  1. 1. Department of Surveying and Mapping Engineering, Shijiazhuang Institute of Railway Technology, Shijiazhuang 050018, China;
    2. North Latitude(Beijing) Technology Co., Ltd., Beijing 100043, China
  • Received:2021-05-19 Revised:2021-09-20 Online:2021-11-25 Published:2021-12-02

摘要: 百度深度学习PaddlePaddle框架支持下的遥感智能视觉平台,能够运用深度学习技术实现遥感影像的智能建模、训练和解译。本文通过深入分析PaddlePaddle图像分割模型库PaddleSeg的图像处理深度学习算法模型DeepLabV3+、U2-Net及RetinaNet,开发设计了遥感智能视觉平台,实现了遥感影像的地块分割、变化检测和斜框检测等专业功能。研究表明:遥感智能视觉平台提取的图斑总面积是目视解译的80%、有效图斑比例为76%、错误图斑比例为18%,实现了快速有效的遥感图像智能处理。

关键词: 深度学习, PaddlePaddle, 算法模型, 遥感图像, 智能处理

Abstract: The remote sensing intelligent vision platform supported by Baidu deep learning PaddlePaddle framework is researched and implemented. It can use deep learning technology to realize intelligent modeling, training and interpretation of remote sensing images.Through the deep analysis of the deep learning algorithm model DeepLabV3+, U2-Net and RetionaNet of PaddleSeg image segmentation model library, the remote sensing intelligent vision platform is developed and designed, which realizes the professional functions of parcel segmentation, change detection and oblique frame detection of remote sensing image. The results show that:the total area of the spot extracted by the remote sensing intelligent vision platform is 80% of the visual interpretation, the proportion of effective spot is 76%, and the proportion of error spot is 18%, which realizes the fast and effective remote sensing image intelligent processing.

Key words: deep learning, PaddlePaddle, algorithm model, remote sensing image, intelligent processing

中图分类号: