数据与计算发展前沿 ›› 2021, Vol. 3 ›› Issue (2): 133-141.

doi: 10.11871/jfdc.issn.2096-742X.2021.02.015

• 技术与应用 • 上一篇    

基于深度学习的遥感影像地块分割方法

赵伟昱1,2(),张宏海1,*(),仲波3()   

  1. 1.中国科学院计算机网络信息中心,北京 100190
    2.中国科学院大学,北京 100049
    3.中国科学院空天信息创新研究院,北京 100094
  • 收稿日期:2020-12-31 出版日期:2021-04-20 发布日期:2021-05-18
  • 通讯作者: 张宏海
  • 作者简介:赵伟昱,中国科学院计算机网络信息中心,中国科学院大学,硕士研究生,研究方向为云计算与分布式系统、深度学习。
    本文主要承担工作为：模型设计、实验数据分析、文章撰写。
    ZHAO Weiyu is a master student in Computer Network Infor-mation Center, Chinese Academy of Sciences (University of Chinese Academy of Sciences). His research interests include cloud computing and distributed system, and deep learning.
    In this paper, he is mainly responsible for model design, experi-mental data analysis and article writing.
    E-mail: zhaoweiyu@cnic.cn|张宏海,中国科学院计算机网络信息中心,副研究员,研究方向为云计算与分布式系统、嵌入式操作系统与物联网、高性能计算环境软件与技术。
    本文主要承担工作为：文章框架的整体设计。
    ZHANG Honghai is an associate research fellow in Computer Network Information Center, Chinese Academy of Sciences. His research interests include cloud computing and distributed system, embedded operating system and Internet of things, high performance computing environment software and technology.
    In this paper, he is mainly responsible for the organization of the paper.
    E-mail: zhh@cnic.cn|仲波,中国科学院空天信息创新研究院,副研究员,研究方向为遥感大数据与产品生产系统。
    本文主要承担工作为：遥感图像分割介绍与研究指导。
    ZHONG Bo is an associate research fellow in Aerospace Information Research Institute, Chinese Academy of Sciences. His research interests include remote sensing big data and product production system.
    In this paper, he is mainly responsible for the introduction of Remote sensing image segmentation and supervising the research.
    E-mail: zhongbo@radi.ac.cn

A Deep Learning Based Method for Remote Sensing Image Parcel Segmentation

ZHAO Weiyu1,2(),ZHANG Honghai1,*(),ZHONG Bo3()   

  1. 1. Computer Network Information Center, Chinese Academy of Sciences, Beijing 100190, China
    2. University of Chinese Academy of Sciences, Beijing 100049, China
    3. Aerospace Information Research Institute,Chinese Academy of Sciences, Beijing 100094, China
  • Received:2020-12-31 Online:2021-04-20 Published:2021-05-18
  • Contact: ZHANG Honghai

RichHTML

64

PDF

403

摘要：

【目的】遥感影像地块分割是遥感影像解译的一项具体任务。良好的遥感影像地块分割结果可以为环境保护、农业生产、城镇建设提供指导意见。【方法】本文使用Pytorch框架搭建了DeepLabV3+网络,编码器使用ResNet101和空洞空间金字塔池化模块进行特征提取,解码器使用双线性插值的方法进行特征图尺寸还原。训练过程中,针对遥感影像地块分割任务,专门设计了训练时的数据增强策略,从而增强模型的泛化能力。使用联合Lovasz loss和Softmax loss的损失函数克服样本类别分布不平衡的问题。【结果】实验结果选用平均交并比作为评价指标,最终模型的平均交并比可以达到70.3%,相比遥感图像分割常用的UNet方法提高了7.6%。【局限】部分分割图像的区域不够完整,还需要进一步提高分割图像的连通性。【结论】本文提出的遥感影像地块分割方法,可以实现对高分辨率遥感图像的精细分割,为遥感图像分割的研究提供借鉴。

关键词: 遥感图像, 图像分割, 深度学习

Abstract:

[Objective] Remote sensing image parcel segmentation is a specific task of remote sensing image interpretation. Good results of remote sensing image parcel segmentation can provide guidance for environmental protection, agricultural production and town construction. [Methods] In this paper, DeepLabV3+ network is built using Pytorch framework. The encoder uses ResNet101 and atrous spatial pyramid pooling module for feature extraction, and the decoder uses the bilinear interpolation method for feature map resizing. During the training process, a data augmentation strategy is specifically designed for the remote sensing image parcel segmentation task, so as to enhance the generalization ability of the model. A loss function with joint Lovasz loss and Softmax loss is used to solve the problem of unbalanced distribution of sample categories. [Results] The mean Intersection over Union is chosen as the evaluation index for the experimental results. The mean Intersection over Union of the final model can reach 70.3%, which is 7.6% higher than the UNet, a commonly used method for remote sensing image segmentation. [Limitations] The region of some segmented images is not complete, and the connectivity of segmented images needs to be further improved. [Conclusions] The remote sensing image parcel segmentation method proposed in this paper can achieve fine segmentation results of high-resolution remote sensing images and provide a reference for the research of remote sensing image segmentation.

Key words: remote sensing image, image segmentation, deep learning