Optics and Precision Engineering, Volume. 29, Issue 9, 2222(2021)

Remote sensing image feature extraction and classification based on contrastive learning method

Xiao-dong MU1... Kun BAI1,*, Xuan-ang YOU1, Yong-qing ZHU1 and Xue-bing CHEN2 |Show fewer author(s)
Author Affiliations
  • 1College of Operational Support, Rocket Force University of Engineering, Xi’an70025, China
  • 2Unit 61068, Xi’an710100, China
  • show less
    Figures & Tables(13)
    Model structure of APCL
    Structure of Projection Head
    Structure of Prediction Head
    Structure of Asymmetric Predictor
    Example images from the four remote sensing scene classification datasets
    Illustrations of the studied data augmentation operators
    Color histograms for the studied data augmentation operators
    Training behaviors with varying structures of AP operator
    • Table 1. Comparison of the four different remote sensing scene datasets

      View table
      View in Article

      Table 1. Comparison of the four different remote sensing scene datasets

      数据集类别数

      每类样

      本数

      样本总数分辨率/m
      NWPU-RESISC454570031 5000.3~30
      EuroSAT102 000~3 00027 00010~30
      UC Merced211002 1000.3
      SIRI-WHU122002 4002
    • Table 2. Evaluation of the effectiveness of the studied data augmentation operators

      View table
      View in Article

      Table 2. Evaluation of the effectiveness of the studied data augmentation operators

      数据增强方法

      CKA

      指数

      分类准确率(%)
      Horizontal Flip0.935 163.45
      Color Jitter0.680 373.83
      Grayscale0.933 267.97
      Horizontal Flip+Color Jitter0.672 777.73
      Horizontal Flip+Grayscale0.929 466.02
      Color Jitter+Grayscale0.674 676.95
      Horizontal Flip+Color Jitter+Grayscale0.670 178.13
    • Table 3. Evaluation of the effectiveness of different AP structures

      View table
      View in Article

      Table 3. Evaluation of the effectiveness of different AP structures

      AP算子深度值准确率(%)
      AP2-233.59
      AP2-372.66
      AP2-478.52
      AP2-575.78
    • Table 4. Classification accuracy of different methods

      View table
      View in Article

      Table 4. Classification accuracy of different methods

      方法NWPU-RESISC45EuroSATUC MercedSIRI-WHU
      训练使用样本数
      5%20%5%20%5%20%5%20%
      ResNet18 from scratch21.5423.3656.6760.1131.0443.3329.5546.31
      ImageNet pre-trained ResNet1852.3158.9782.8585.4556.8177.5267.0978.91
      SimSiam73.1477.0885.6788.1542.5960.5150.563.17
      APCL73.3777.5785.0187.7046.9860.5256.4365.83
    • Table 5. Classification time of different methods

      View table
      View in Article

      Table 5. Classification time of different methods

      方法NWPU-RESISC45EuroSATUC MercedSIRI-WHU
      训练使用样本数
      5%20%5%20%5%20%5%20%
      ResNet18 from scratch0.380.370.110.110.060.050.060.06
      ImageNet pre-trained ResNet180.380.370.110.110.050.050.040.04
      SimSiam4.394.393.573.560.530.530.540.54
      APCL4.514.503.583.580.540.540.530.53
    Tools

    Get Citation

    Copy Citation Text

    Xiao-dong MU, Kun BAI, Xuan-ang YOU, Yong-qing ZHU, Xue-bing CHEN. Remote sensing image feature extraction and classification based on contrastive learning method[J]. Optics and Precision Engineering, 2021, 29(9): 2222

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category: Information Sciences

    Received: Jan. 29, 2021

    Accepted: --

    Published Online: Nov. 22, 2021

    The Author Email: BAI Kun (nudt@foxmail. com)

    DOI:10.37188/OPE.20212909.2222

    Topics