• Chinese Journal of Lasers
  • Vol. 48, Issue 16, 1610001 (2021)
Xiangda Lei, Hongtao Wang*, and Zongze Zhao
Author Affiliations
  • School of Surveying and Land Information Engineering, Henan Polytechnic University, Jiaozuo,Henan 454000, China
  • show less
    DOI: 10.3788/CJL202148.1610001 Cite this Article Set citation alerts
    Xiangda Lei, Hongtao Wang, Zongze Zhao. Small-Sample Airborne LiDAR Point Cloud Classification Based on Transfer Learning and Fully Convolutional Network[J]. Chinese Journal of Lasers, 2021, 48(16): 1610001 Copy Citation Text show less
    Flow chart of the small sample point cloud classification based on transfer learning
    Fig. 1. Flow chart of the small sample point cloud classification based on transfer learning
    Generation process of the point feature map. (a) Two-dimensional coordinates of the feature map; (b) point features in the cube neighborhood; (c) point feature map
    Fig. 2. Generation process of the point feature map. (a) Two-dimensional coordinates of the feature map; (b) point features in the cube neighborhood; (c) point feature map
    Generation of the multi-scale feature maps. (a) Grid size of 0.1 m; (b) grid size of 0.3 m; (c) grid size of 0.5 m
    Fig. 3. Generation of the multi-scale feature maps. (a) Grid size of 0.1 m; (b) grid size of 0.3 m; (c) grid size of 0.5 m
    Schematic diagram of the multi-projection. (a) X direction; (b) Y direction
    Fig. 4. Schematic diagram of the multi-projection. (a) X direction; (b) Y direction
    Deep feature extraction based on transfer learning
    Fig. 5. Deep feature extraction based on transfer learning
    Point cloud classification based on FCN
    Fig. 6. Point cloud classification based on FCN
    Experimental datasets. (a) Training dataset displayed by normalized height; (b) aerial image corresponding to training dataset; (c) testing dataset displayed by normalized height; (d) aerial image corresponding to testing dataset
    Fig. 7. Experimental datasets. (a) Training dataset displayed by normalized height; (b) aerial image corresponding to training dataset; (c) testing dataset displayed by normalized height; (d) aerial image corresponding to testing dataset
    F1 scores when classifying different feature combinations
    Fig. 8. F1 scores when classifying different feature combinations
    F1 scores when classifying different pre-training models
    Fig. 9. F1 scores when classifying different pre-training models
    Classification results when K=4
    Fig. 10. Classification results when K=4
    Comparison of the misclassification results. (a) Misclassification result before graph-cuts optimization; (b) misclassification result after graph-cuts optimization
    Fig. 11. Comparison of the misclassification results. (a) Misclassification result before graph-cuts optimization; (b) misclassification result after graph-cuts optimization
    KF1 scoreOAAvg F1
    Low_vegImp_surCarFe/HeRoofFacadeShrubTree
    086.1193.5166.2465.7794.4661.0968.7589.0887.7678.13
    186.5093.6468.4967.1294.6962.0569.7289.4388.1878.96
    287.1494.1575.8968.9695.3064.5771.1989.9989.1280.90
    387.1593.9978.4471.0095.7665.9172.6090.6989.6081.94
    487.4294.0979.9171.7896.0667.4972.0490.8089.9182.45
    587.1593.6979.6270.6696.2467.2370.7390.6989.7682.00
    686.5893.1973.6070.2296.0365.2268.4490.5389.2780.48
    Table 1. Influence of different K on the classification results unit: %
    MethodF1 scoreOA
    PowLow_vegImp_surfCarFe/HeRoofFacadeShrubTree
    UM46.179.089.147.75.292.052.740.977.980.8
    WhuY231.980.088.940.824.593.149.441.177.381.0
    WhuY337.181.490.163.423.993.447.539.978.082.3
    LUH59.677.591.173.134.094.256.346.683.181.6
    BIJ_W13.878.590.556.436.392.253.243.378.481.5
    RIT_137.577.991.573.418.094.049.345.982.581.6
    NANJ262.088.891.266.740.793.642.655.982.685.2
    WhuY442.582.791.474.753.794.353.147.982.884.9
    Ours--87.494.179.971.896.167.572.090.889.9
    Table 2. F1 scores and overall classification accuracy of different methods unit: %
    MethodF1 scoreOAAvg F1
    Low_vegImp_surfCarFe/HeRoofFacadeShrubTree
    NANJ286.993.169.570.995.273.065.392.189.180.8
    Ours88.693.868.775.595.674.068.894.390.182.4
    Table 3. Classification results of our method and NANJ2 method unit: %
    MethodF1 scoreOAAvg F1
    Low_vegImp_surfCarFe/HeRoofFacadeShrubTree
    DRN83.392.552.462.595.262.864.988.786.875.3
    DRN-187.094.342.265.796.066.071.791.889.578.2
    Ours87.494.179.971.896.167.572.090.889.982.4
    Table 4. Classification results of different transfer learning methods unit: %
    Xiangda Lei, Hongtao Wang, Zongze Zhao. Small-Sample Airborne LiDAR Point Cloud Classification Based on Transfer Learning and Fully Convolutional Network[J]. Chinese Journal of Lasers, 2021, 48(16): 1610001
    Download Citation