测绘学报 ›› 2017, Vol. 46 ›› Issue (12): 1969-1977.doi: 10.11947/j.AGCS.2017.20170291

• 摄影测量学与遥感 • 上一篇    下一篇

结合对抗网络与辅助任务的遥感影像无监督域适应方法

许夙晖1, 慕晓冬1, 张雄美1, 柴栋2   

  1. 1. 火箭军工程大学信息工程系, 陕西 西安 710025;
    2. 北京航空工程技术研究中心, 北京 100076
  • 收稿日期:2017-06-05 修回日期:2017-10-24 出版日期:2017-12-20 发布日期:2017-12-28
  • 作者简介:许夙晖(1989-),女,博士生,研究方向为遥感图像处理和模式识别。E-mail:xu_suhui@163.com
  • 基金资助:
    国家自然科学基金(61640007)

Unsupervised Remote Sensing Domain Adaptation Method with Adversarial Network and Auxiliary Task

XU Suhui1, MU Xiaodong1, ZHANG Xiongmei1, CHAI Dong2   

  1. 1. Department of Information Engineering, Rocket Force Engineering University, Xi'an 710025, China;
    2. Beijing Aeronautical Technology Research Center, Beijing 100076, China
  • Received:2017-06-05 Revised:2017-10-24 Online:2017-12-20 Published:2017-12-28
  • Supported by:
    The National Natural Science Foundation of China (No. 61640007)

摘要: 使用机器学习进行遥感影像标注的一个重要前提是有足够的训练样本,而样本的标注是非常耗时的。本文采用了域适应的方法来解决遥感影像场景分类中小样本量的无监督学习问题,提出了结合对抗网络与辅助任务的遥感影像域适应方法。首先建立了基于深度卷积神经网络的遥感影像分类框架;其次,为了学习到域不变特征,在标签分类器的基础上增加域分类器,并使域损失函数在其反射传播时的梯度与标签损失的梯度相反,从而保证域分类器不能区分样本来自于哪个域;最后引入了辅助分类任务,扩充了样本的同时使网络更具泛化能力。试验结果表明,本文方法优于主流的无监督域适应方法,在小样本遥感影像无监督分类中得到了较好的效果。

关键词: 遥感影像, 场景分类, 域适应, 深度卷积神经网络, 对抗网络, 多任务学习

Abstract: An important prerequisite when annotating the remote sensing images by machine learning is that there are enough training samples for training, but labeling the samples is very time-consuming. In this paper, we solve the problem of unsupervised learning with small sample size in remote sensing image scene classification by domain adaptation method. A new domain adaptation framework is proposed which combines adversarial network and auxiliary task. Firstly, a novel remote sensing scene classification framework is established based on deep convolution neural networks. Secondly, a domain classifier is added to the network, in order to learn the domain-invariant features. The gradient direction of the domain loss is opposite to the label loss during the back propagation, which makes the domain predictor failed to distinguish the sample's domain. Lastly, we introduce an auxiliary task for the network, which augments the training samples and improves the generalization ability of the network. The experiments demonstrate better results in unsupervised classification with small sample sizes of remote sensing images compared to the baseline unsupervised domain adaptation approaches.

Key words: remote sensing image, scene classification, domain adaptation, deep convolutional neural network, adversarial network, multi-task learning

中图分类号: