测绘学报 ›› 2022, Vol. 51 ›› Issue (5): 691-702.doi: 10.11947/j.AGCS.2022.20210270

• 摄影测量学与遥感 • 上一篇    下一篇

融合半监督学习的无监督遥感影像场景分类

白坤1, 慕晓冬1, 陈雪冰2, 朱永清1, 尤轩昂1   

  1. 1. 火箭军工程大学作战保障学院, 陕西 西安 710025;
    2. 61068部队, 陕西 西安 710100
  • 收稿日期:2021-05-12 修回日期:2021-12-30 出版日期:2022-05-20 发布日期:2022-05-28
  • 作者简介:白坤(1993-),男,硕士生,研究方向为遥感图像处理与计算机视觉。E-mail:baikun.nudt@foxmail.com
  • 基金资助:
    国家自然科学基金(61773389)

Unsupervised remote sensing image scene classification based on semi-supervised learning

BAI Kun1, MU Xiaodong1, CHEN Xuebing2, ZHU Yongqing1, YOU Xuanang1   

  1. 1. College of Operational Support, Rocket Force University of Engineering, Xi'an 710025, China;
    2. Troops 61068, Xi'an 710100, China
  • Received:2021-05-12 Revised:2021-12-30 Online:2022-05-20 Published:2022-05-28
  • Supported by:
    The National Nature Science Foundation of China (No. 61773389)

摘要: 自监督学习可以不依赖样本标签对遥感影像进行特征提取,但是特征分类仍然依赖有监督方法。为了克服有监督特征分类过程的不足,实现遥感影像特征的无监督自动分类,本文提出一种融合半监督学习的无监督语义聚类方法。首先,使用自监督学习提取遥感影像特征,抽象出图像包含的高层语义信息;然后,基于特征相似度寻找每个样本最相似的近邻,使用在线聚类将相似样本聚为一类,训练一个线性分类器;最后,根据聚类结果为高置信度样本生成伪标签,构造标注样本集,使用半监督方法对模型微调。在4个公开遥感影像场景分类数据集EuroSAT、GID、AID和NWPU-RESISC45上进行验证,分类精度分别达到了94.84%、63.55%、76.42%和86.24%。本文方法结合了在线聚类和半监督学习的优点,缓解了已有方法存在的误差积累和样本利用不充分的问题,在完全不使用标注样本的情况下,充分利用自监督特征训练分类模型,对遥感影像进行场景分类,达到接近有监督学习的分类效果,具有良好的应用价值。

关键词: 遥感影像, 场景分类, 自监督, 在线聚类, 半监督

Abstract: Self-supervised learning can extract features from remote sensing images without relying on sample labels, but supervised method is still required to complete feature classification. To overcome the shortcomings of feature classification process and automatically classify remote sensing image features, an unsupervised semantic clustering method based on semi-supervised learning is proposed. First, the features of remote sensing images are extracted using self-supervised learning to abstract the high-level semantic information contained in the images. Then, each sample's closest neighbors are found based on the feature similarity, and a linear classifier is trained by clustering similar samples into one class using the online clustering method. Finally, based on the clustering results, pseudo labels are generated for samples with high confidence to construct a label sample set, and the model is fine-tuned using a semi-supervised method. Four public remote sensing image scene classification datasets, EuroSAT, GID, AID and NWPU-RESISC45, were validated, and the classification accuracy reached 94.84%, 63.55%, 76.42% and 86.24%, respectively. The method presented in this paper combines the advantages of online clustering and semi-supervised learning, alleviates the problems of error accumulation and insufficient sample utilization in existing methods. It makes full use of the self-supervised features to train the classification model and complete the scene classification of remote sensing images without using labeled samples at all. It achieves the classification effect close to supervised learning and has good application value.

Key words: remote sensing, scene classification, self-supervised, online clustering, semi-supervised

中图分类号: