Author Affiliations
School of Surveying and Land Information Engineering, Henan Polytechnic University, Jiaozuo,Henan 454000, Chinashow less
Fig. 1. Flow chart of the small sample point cloud classification based on transfer learning
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
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
Fig. 4. Schematic diagram of the multi-projection. (a) X direction; (b) Y direction
Fig. 5. Deep feature extraction based on transfer learning
Fig. 6. Point cloud classification based on FCN
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
Fig. 8. F1 scores when classifying different feature combinations
Fig. 9. F1 scores when classifying different pre-training models
Fig. 10. Classification results when K=4
Fig. 11. Comparison of the misclassification results. (a) Misclassification result before graph-cuts optimization; (b) misclassification result after graph-cuts optimization
K | F1 score | OA | Avg F1 |
---|
Low_veg | Imp_sur | Car | Fe/He | Roof | Facade | Shrub | Tree |
---|
0 | 86.11 | 93.51 | 66.24 | 65.77 | 94.46 | 61.09 | 68.75 | 89.08 | 87.76 | 78.13 | 1 | 86.50 | 93.64 | 68.49 | 67.12 | 94.69 | 62.05 | 69.72 | 89.43 | 88.18 | 78.96 | 2 | 87.14 | 94.15 | 75.89 | 68.96 | 95.30 | 64.57 | 71.19 | 89.99 | 89.12 | 80.90 | 3 | 87.15 | 93.99 | 78.44 | 71.00 | 95.76 | 65.91 | 72.60 | 90.69 | 89.60 | 81.94 | 4 | 87.42 | 94.09 | 79.91 | 71.78 | 96.06 | 67.49 | 72.04 | 90.80 | 89.91 | 82.45 | 5 | 87.15 | 93.69 | 79.62 | 70.66 | 96.24 | 67.23 | 70.73 | 90.69 | 89.76 | 82.00 | 6 | 86.58 | 93.19 | 73.60 | 70.22 | 96.03 | 65.22 | 68.44 | 90.53 | 89.27 | 80.48 |
|
Table 1. Influence of different K on the classification results unit: %
Method | F1 score | OA |
---|
Pow | Low_veg | Imp_surf | Car | Fe/He | Roof | Facade | Shrub | Tree |
---|
UM | 46.1 | 79.0 | 89.1 | 47.7 | 5.2 | 92.0 | 52.7 | 40.9 | 77.9 | 80.8 | WhuY2 | 31.9 | 80.0 | 88.9 | 40.8 | 24.5 | 93.1 | 49.4 | 41.1 | 77.3 | 81.0 | WhuY3 | 37.1 | 81.4 | 90.1 | 63.4 | 23.9 | 93.4 | 47.5 | 39.9 | 78.0 | 82.3 | LUH | 59.6 | 77.5 | 91.1 | 73.1 | 34.0 | 94.2 | 56.3 | 46.6 | 83.1 | 81.6 | BIJ_W | 13.8 | 78.5 | 90.5 | 56.4 | 36.3 | 92.2 | 53.2 | 43.3 | 78.4 | 81.5 | RIT_1 | 37.5 | 77.9 | 91.5 | 73.4 | 18.0 | 94.0 | 49.3 | 45.9 | 82.5 | 81.6 | NANJ2 | 62.0 | 88.8 | 91.2 | 66.7 | 40.7 | 93.6 | 42.6 | 55.9 | 82.6 | 85.2 | WhuY4 | 42.5 | 82.7 | 91.4 | 74.7 | 53.7 | 94.3 | 53.1 | 47.9 | 82.8 | 84.9 | Ours | -- | 87.4 | 94.1 | 79.9 | 71.8 | 96.1 | 67.5 | 72.0 | 90.8 | 89.9 |
|
Table 2. F1 scores and overall classification accuracy of different methods unit: %
Method | F1 score | OA | Avg F1 |
---|
Low_veg | Imp_surf | Car | Fe/He | Roof | Facade | Shrub | Tree |
---|
NANJ2 | 86.9 | 93.1 | 69.5 | 70.9 | 95.2 | 73.0 | 65.3 | 92.1 | 89.1 | 80.8 | Ours | 88.6 | 93.8 | 68.7 | 75.5 | 95.6 | 74.0 | 68.8 | 94.3 | 90.1 | 82.4 |
|
Table 3. Classification results of our method and NANJ2 method unit: %
Method | F1 score | OA | Avg F1 |
---|
Low_veg | Imp_surf | Car | Fe/He | Roof | Facade | Shrub | Tree |
---|
DRN | 83.3 | 92.5 | 52.4 | 62.5 | 95.2 | 62.8 | 64.9 | 88.7 | 86.8 | 75.3 | DRN-1 | 87.0 | 94.3 | 42.2 | 65.7 | 96.0 | 66.0 | 71.7 | 91.8 | 89.5 | 78.2 | Ours | 87.4 | 94.1 | 79.9 | 71.8 | 96.1 | 67.5 | 72.0 | 90.8 | 89.9 | 82.4 |
|
Table 4. Classification results of different transfer learning methods unit: %