Figure 1.
Flow chart of building extraction based on CNN and superpixel in the PolSAR image.
Figure 1.
Flow chart of building extraction based on CNN and superpixel in the PolSAR image.
Figure 2.
The SAR and optical images, and the true ground map. (a1,b1,c1) are the PauliRGB images from E-SAR, GF-3, and RADARSAT-2, respectively. (a2,b2,c2) are the optical images. (a3,b3,c3) are the true ground map. Red is the building and green is the non-building in (a3,b3,c3).
Figure 2.
The SAR and optical images, and the true ground map. (a1,b1,c1) are the PauliRGB images from E-SAR, GF-3, and RADARSAT-2, respectively. (a2,b2,c2) are the optical images. (a3,b3,c3) are the true ground map. Red is the building and green is the non-building in (a3,b3,c3).
Figure 3.
The convolutional neural network structure for building extraction.
Figure 3.
The convolutional neural network structure for building extraction.
Figure 4.
ESAR image building extraction results. (a) Quan’s threshold extraction method; (b) PauliRGB and G0 statistical texture parameters using SVM results; (c) RVCNN; (d) PFDCNN; (e) PauliRGB and G0 Statistical texture parameters using CNN; (f) as the result of introducing superpixel constraints in (e).
Figure 4.
ESAR image building extraction results. (a) Quan’s threshold extraction method; (b) PauliRGB and G0 statistical texture parameters using SVM results; (c) RVCNN; (d) PFDCNN; (e) PauliRGB and G0 Statistical texture parameters using CNN; (f) as the result of introducing superpixel constraints in (e).
Figure 5.
GF-3 image building extraction results. (a) Quan’s threshold extraction method; (b) PauliRGB and G0 statistical texture parameters using SVM results; (c) RVCNN; (d) PFDCNN; (e) PauliRGB and G0 Statistical texture parameters using CNN; (f) as the result of introducing superpixel constraints in (e).
Figure 5.
GF-3 image building extraction results. (a) Quan’s threshold extraction method; (b) PauliRGB and G0 statistical texture parameters using SVM results; (c) RVCNN; (d) PFDCNN; (e) PauliRGB and G0 Statistical texture parameters using CNN; (f) as the result of introducing superpixel constraints in (e).
Figure 6.
RADARSAT-2 image building extraction results. (a) Quan’s threshold extraction method; (b) PauliRGB and G0 statistical texture parameters using SVM results; (c) RVCNN; (d) PFDCNN; (e) PauliRGB and G0 Statistical texture parameters using CNN; (f) as the result of introducing superpixel constraints in (e).
Figure 6.
RADARSAT-2 image building extraction results. (a) Quan’s threshold extraction method; (b) PauliRGB and G0 statistical texture parameters using SVM results; (c) RVCNN; (d) PFDCNN; (e) PauliRGB and G0 Statistical texture parameters using CNN; (f) as the result of introducing superpixel constraints in (e).
Figure 7.
Buildings and similar building features in the PolSAR image (A,B,C) images from ESAR. Group (D) images from GF-3 and Group (E) images from RADARSAT-2. Where (a) is the optical image, (b) is PauliRGB, (c) is the mask of the real building distributed on PauliRGB (red area), and (d) is the G0 statistical texture parameter. (e) Classification results obtained by using only PauliRGB as CNN input training, in which red is a building, green is a non-building, and (f) is a classification obtained by adding a G0 statistical texture parameter to PauliRGB as a CNN input training, in which red is a building and green is a non-building.
Figure 7.
Buildings and similar building features in the PolSAR image (A,B,C) images from ESAR. Group (D) images from GF-3 and Group (E) images from RADARSAT-2. Where (a) is the optical image, (b) is PauliRGB, (c) is the mask of the real building distributed on PauliRGB (red area), and (d) is the G0 statistical texture parameter. (e) Classification results obtained by using only PauliRGB as CNN input training, in which red is a building, green is a non-building, and (f) is a classification obtained by adding a G0 statistical texture parameter to PauliRGB as a CNN input training, in which red is a building and green is a non-building.
Figure 8.
Comparison of feature sets under three data using G0 texture parameters before and after comparison. (a,b) is the experimental result under ESAR data, (c,d) is the experimental result under GF-3 data, and (e,f) is the experimental result under RADASAT-2 data. Where (a,c,e) is the experimental result of the input of only PauliRGB and (b,d,f) is the experimental result of the input texture parameter of PauliRGB and G0.
Figure 8.
Comparison of feature sets under three data using G0 texture parameters before and after comparison. (a,b) is the experimental result under ESAR data, (c,d) is the experimental result under GF-3 data, and (e,f) is the experimental result under RADASAT-2 data. Where (a,c,e) is the experimental result of the input of only PauliRGB and (b,d,f) is the experimental result of the input texture parameter of PauliRGB and G0.
Figure 9.
Comparison of using the MLP and CNN methods. (a,b) is the experimental result under ESAR data, (c,d) is the experimental result under GF-3 data, and (e,f) is the experimental result under RADASAT-2 data. Where (a,c,e) is the experimental result of the MLP and (b,d,f) is the experimental result of the CNN.
Figure 9.
Comparison of using the MLP and CNN methods. (a,b) is the experimental result under ESAR data, (c,d) is the experimental result under GF-3 data, and (e,f) is the experimental result under RADASAT-2 data. Where (a,c,e) is the experimental result of the MLP and (b,d,f) is the experimental result of the CNN.
Figure 10.
Comparison of results before and after superpixel constraints under three data. (a,b) is the experimental result under ESAR data; (c,d) is the experimental result under GF-3 data; and (e,f) is the experimental result under RADASAT-2 data. (a,c,e) are experimental results without superpixel constraints and (b,d,f) are experimental results after using superpixel constraints.
Figure 10.
Comparison of results before and after superpixel constraints under three data. (a,b) is the experimental result under ESAR data; (c,d) is the experimental result under GF-3 data; and (e,f) is the experimental result under RADASAT-2 data. (a,c,e) are experimental results without superpixel constraints and (b,d,f) are experimental results after using superpixel constraints.
Figure 11.
Sample selection and building extraction results. (a) is the samples from the true surface map; (b) is the result of building extraction using (a) samples; (c) consists of positive and negative samples (dark green); and (d) is the result of building extraction using (c) samples.
Figure 11.
Sample selection and building extraction results. (a) is the samples from the true surface map; (b) is the result of building extraction using (a) samples; (c) consists of positive and negative samples (dark green); and (d) is the result of building extraction using (c) samples.
Figure 12.
Accuracy of building extraction results under different image block sizes.
Figure 12.
Accuracy of building extraction results under different image block sizes.
Figure 13.
Building elements used for training account for the accuracy of building extraction results at different percentages of all building pixels.
Figure 13.
Building elements used for training account for the accuracy of building extraction results at different percentages of all building pixels.
Table 1.
Different methods of building extraction effects under ESAR data.
Table 1.
Different methods of building extraction effects under ESAR data.
Feature | Method | AR (%) | FAR (%) | F1-Score (%) |
---|
Eigenvalue | Threshold | 22.20 | 39.23 | 32.52 |
PauliRGB + G0 | SVM | 61.85 | 58.15 | 49.92 |
6D-Vector [38] | CNN | 85.18 | 27.89 | 78.10 |
Polarimetric Features [39] | CNN | 85.64 | 45.54 | 66.58 |
PauliRGB + G0 | CNN | 88.05 | 25.01 | 80.99 |
PauliRGB + G0 | CNN + Superpixel | 86.14 | 17.61 | 84.22 |
Table 2.
Different methods of building extraction effects under GF-3 data.
Table 2.
Different methods of building extraction effects under GF-3 data.
Feature | Method | AR (%) | FAR (%) | F1-Score (%) |
---|
Eigenvalue | Threshold | 67.02 | 41.35 | 62.55 |
PauliRGB + G0 | SVM | 78.95 | 41.63 | 67.11 |
6D-Vector [38] | CNN | 94.69 | 24.91 | 83.75 |
Polarimetric Features [39] | CNN | 95.33 | 24.09 | 84.51 |
PauliRGB + G0 | CNN | 95.56 | 15.45 | 89.71 |
PauliRGB + G0 | CNN + Superpixel | 94.97 | 12.2 | 91.24 |
Table 3.
Different methods of building extraction effects under RADARSAT-2 data.
Table 3.
Different methods of building extraction effects under RADARSAT-2 data.
Feature | Method | AR (%) | FAR (%) | F1-Score (%) |
---|
Eigenvalue | Threshold | 64.11 | 25.13 | 69.07 |
PauliRGB + G0 | SVM | 84.03 | 45.06 | 66.44 |
6D-Vector [38] | CNN | 93.62 | 29.99 | 80.11 |
Polarimetric Features [39] | CNN | 94.29 | 30.82 | 79.80 |
PauliRGB + G0 | CNN | 94.37 | 21.76 | 85.55 |
PauliRGB + G0 | CNN + Superpixel | 93.64 | 17.89 | 87.49 |
Table 4.
Building extraction accuracy table using G0 statistical texture parameters.
Table 4.
Building extraction accuracy table using G0 statistical texture parameters.
| E-SAR | GF-3 | RADASAT-2 |
---|
| AR (%) | FAR (%) | F1-Score (%) | AR (%) | FAR (%) | F1-Score (%) | AR (%) | FAR (%) | F1-Score (%) |
---|
PauliRGB | 80.05 | 30.49 | 74.41 | 93.29 | 20.54 | 85.82 | 93.99 | 29.4 | 80.63 |
PauliRGB + G0 | 88.05 | 25.01 | 80.99 | 95.56 | 15.45 | 89.71 | 94.37 | 21.76 | 85.55 |
Table 5.
Building extraction accuracy table using the MLP and CNN methods.
Table 5.
Building extraction accuracy table using the MLP and CNN methods.
| E-SAR | GF-3 | RADASAT-2 |
---|
| AR (%) | FAR (%) | F1-Score (%) | AR (%) | FAR (%) | F1-Score (%) | AR (%) | FAR (%) | F1-Score (%) |
---|
MLP | 75.36 | 57.67 | 54.21 | 81.37 | 40.94 | 0.6844 | 79.93 | 39.63 | 68.78 |
CNN | 88.05 | 25.01 | 80.99 | 95.56 | 15.45 | 0.8971 | 94.37 | 21.76 | 85.55 |
Table 6.
Building extraction accuracy table using the superpixel methods.
Table 6.
Building extraction accuracy table using the superpixel methods.
| E-SAR | GF-3 | RADASAT-2 |
---|
| AR (%) | FAR (%) | F1-Score (%) | AR (%) | FAR (%) | F1-Score (%) | AR (%) | FAR (%) | F1-Score (%) |
---|
Non-superpixel | 88.05 | 25.01 | 80.99 | 94.37 | 21.76 | 85.55 | 95.56 | 15.45 | 89.71 |
Superpixel | 86.14 | 17.61 | 84.22 | 93.64 | 17.89 | 87.49 | 94.97 | 12.2 | 94.12 |