Remote Sensing Technology and Application, Volume. 39, Issue 3, 527(2024)

Topic Model for High Resolution Remote Sensing Data Interpretation: A Review

Zhen LI, Qiqi ZHU*, Yang LEI, Jiangqin WAN, Linlin WANG, and Lei XU
Author Affiliations
  • School of Geography and Information Engineering, China University of Geosciences, National Engineering Research Center of GIS, China University of Geosciences, Wuhan, China
  • show less
    Figures & Tables(11)
    Directed graph model of pLSA
    Directed graph model of LDA
    Directed graph model of FSTM
    Experimental data of semantic segmentation
    • Table 1. Comprehensive applications of topic model for remote sensing images

      View table
      View in Article

      Table 1. Comprehensive applications of topic model for remote sensing images

      主要应用最新进展
      场景分类Lienou等[1];Zhong等[18]
      语义分割Tang等[29];Ruben等[32]
      目标识别Zhang等[30]
      灾难监测Cheng等[33]
      光谱解混Jaramago等[10]
      LiDAR点云分类Kang等[31]
      超分辨率重建Polatkan等[34]
    • Table 2. Topic numbers K and word numbers V values for the different methods for the UC Merced Dataset

      View table
      View in Article

      Table 2. Topic numbers K and word numbers V values for the different methods for the UC Merced Dataset

      实验方式词典数V主题数K
      pLSA[42]1 000270
      LDA[37]1 000240
      SAL-LDA[18]3 000210
      FSSTM[38]2 800840
    • Table 3. Classification Accuracy (%) of the Different Methods for the UC Merced Dataset

      View table
      View in Article

      Table 3. Classification Accuracy (%) of the Different Methods for the UC Merced Dataset

      实验方式精度/%
      主题模型单特征pLSA[42]89.51±1.31
      LDA[37]81.27±2.01
      多特征SAL-LDA[18]88.33±1.76
      FSSTM[38]95.71±1.01
      与深度学习结合AlexNet[44]94.37±0.00
      VGG-19[44]93.15±0.00
      RNBFE[44]97.92±0.00
      ADSSM[45]99.76±0.24
    • Table 4. Topic numbers K and word numbers V values for the different methods for the QuickBird image

      View table
      View in Article

      Table 4. Topic numbers K and word numbers V values for the different methods for the QuickBird image

      实验方式词典数V主题数K
      msLDA[29]2 0007
      ssLDA[41]2 00010
    • Table 5. Classification accuracy (%) of different methods for the QuickBird image

      View table
      View in Article

      Table 5. Classification accuracy (%) of different methods for the QuickBird image

      实验方式msLDA[29]ssLDA[41]
      Kappa系数62.8474.70
      水体/97.23
      建筑/81.45
      空地/92.08
      道路/34.09
      阴影/64.26
      树木/77.62
    • Table 6. the definition of the acronyms in the literature

      View table
      View in Article

      Table 6. the definition of the acronyms in the literature

      缩写中文名称英文名称
      BoVW视觉词袋模型Bag of Visual Words
      pLSA潜在概率语义分析probabilistic latent semantic analysis
      LDA潜在狄利克雷分布Latent Dirichlet allocation
      SVM支持向量机Support Vector Machine
      FSTM全稀疏主题模型Fully Sparse Topic Model
      GTM图主题模型Graph Topic Model
      sDTM监督贝叶斯深度主题模型Supervised Bayesian Deep Topic Model
      NTM神经主题建模Neural Topic Modeling
      RNN循环神经网络Recurrent Neural Networks
      LiDAR激光探测及测距系统Light Detection and Ranging
      SAR合成孔径雷达Synthetic Aperture Radar
      SIFT尺度不变特征变换Scale-invariant Feature Transform
      MSI多光谱影像Multispectral Image
      DEpLSA双深度稀疏概率潜在语义分析Dual-Depth Sparse Probabilistic Latent Semantic Analysis
      SAL-LDA基于语义分布的潜在狄利克雷分布Semantic Allocation Level- Latent Dirichlet allocation
      FSSTM全稀疏语义主题模型Fully Sparse Semantic Topic Model
      RNBFE基于残差网络的特征提取Residual Network-based Features Extraction
      ADSSM自适应深度稀疏语义建模框架Adaptive Deep Sparse Semantic Modeling Framework
      msLDA多尺度潜在狄利克雷分布Multiscale Latent Dirichlet Allocation
      ssLDA半监督潜在狄利克雷分布Semisupervised Latent Dirichlet Allocation
    • Table 7. Summary of the strengths and weaknesses of different topic models

      View table
      View in Article

      Table 7. Summary of the strengths and weaknesses of different topic models

      主题模型优点缺点
      经典模型BoVW使用了恰当的方式选取了具有代表性的因素来线性地表征影像信息基于底层特征的描述无法很好地描述复杂场景
      pLSA更清楚地再现了从特征到场景再到影像的整个过程容易过拟合
      LDA克服了过拟合现象,使得整个模型更贴合真实情况无法很好地处理多源遥感数据
      FSTM能生成场景的稀疏又有代表性的特征表达主题数需人为确定
      与深度学习结合利用深度学习网络加强主题建模,更易贴合实际场景需要大量的训练样本训练模型
      传统模型优化融合不同的主题模型来实现模型间的互补难以提取与理解的场景高层特征
    Tools

    Get Citation

    Copy Citation Text

    Zhen LI, Qiqi ZHU, Yang LEI, Jiangqin WAN, Linlin WANG, Lei XU. Topic Model for High Resolution Remote Sensing Data Interpretation: A Review[J]. Remote Sensing Technology and Application, 2024, 39(3): 527

    Download Citation

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

    Category:

    Received: Sep. 11, 2020

    Accepted: --

    Published Online: Dec. 9, 2024

    The Author Email: ZHU Qiqi (zhuqq@cug.edu.cn)

    DOI:10.11873/j.issn.1004-0323.2024.3.0527

    Topics