Fig. 1. Polarimetric SAR images from Radarsat-2 in San Francisco. (a) Pauli pseudo color image; (b) corresponding map; (c) ground truth
Fig. 2. Polarimetric SAR images from Radarsat-2 in Flevoland. (a) Pauli pseudo color image; (b) corresponding map; (c) ground truth
Fig. 3. Polarimetric SAR images from Radarsat-2 in Vancouver. (a) Pauli pseudo color image; (b) corresponding map; (c) ground truth
Fig. 4. Fitting results of each feature at different distributions. (a) Water; (b) forest; (c) farmland; (d) urban
Fig. 5. Flow chart of constrained distance estimation algorithm
Fig. 6. Distance function at different parameters
Fig. 7. Fitting results at different parameters. (a) k=1; (b) k=5; (c) k=8
Fig. 8. Flow chart of polarimetric SAR image classification algorithm based on GMM model
Fig. 9. Polarimetric SAR classification results in San Francisco. (a) KNN; (b) SVM; (c) RF; (d) WHRT; (e) GMM
Fig. 10. Polarimetric SAR classification results in Flevoland. (a) KNN; (b) SVM; (c) RF; (d) WHRT; (e) GMM
Fig. 11. Polarimetric SAR classification results in Vancouver. (a) KNN; (b) SVM; (c) RF; (d) WHRT; (e) GMM
Fig. 12. Overall accuracy vs. number of training samples
Distribution | Water | Forest | Farmland | Urban |
---|
Gamma | 1.414 | 0.691 | 0.436 | 1.221 | Log-normal | 1.277 | 0.796 | 0.400 | 1.115 | K | 1.764 | 1.134 | 0.784 | 1.515 | GMM | 0.245 | 0.257 | 0.283 | 0.175 |
|
Table 1. KS distance mean at different distributions
Distribution | Water | Forest | Farmland | Urban |
---|
Gamma | 2.544 | 0.941 | 0.537 | 2.087 | Log-normal | 2.233 | 1.387 | 0.506 | 1.890 | K | 3.135 | 1.419 | 0.882 | 2.551 | GMM | 0.305 | 0.322 | 0.401 | 0.250 |
|
Table 2. Maximum KS distance at different distributions
Place | Algorithm | Water | Forest | Farmland | Urban | Overall | Kappa |
---|
San Francisco | KNN | 99.82 | 98.28 | - | 69.17 | 94.34 | 91.16 | SVM | 99.68 | 96.70 | - | 78.26 | 95.03 | 91.61 | RF | 99.78 | 98.24 | - | 79.73 | 95.97 | 92.76 | WHRT | 99.42 | 95.44 | - | 73.01 | 93.01 | 89.24 | GMM | 99.50 | 98.98 | - | 95.84 | 96.90 | 93.23 | Flevoland | KNN | 86.77 | 91.10 | 75.29 | 74.91 | 83.85 | 76.23 | SVM | 82.47 | 90.42 | 72.88 | 75.29 | 82.71 | 75.15 | RF | 85.98 | 86.03 | 73.69 | 75.88 | 83.01 | 76.60 | WHRT | 85.60 | 86.96 | 72.05 | 72.19 | 80.89 | 72.37 | GMM | 87.69 | 85.81 | 92.70 | 88.19 | 89.70 | 82.11 | Vancouver | KNN | 99.28 | 87.15 | 73.88 | 48.19 | 81.02 | 66.77 | SVM | 99.18 | 84.72 | 63.20 | 66.10 | 80.57 | 65.45 | RF | 99.28 | 86.86 | 79.56 | 72.52 | 84.16 | 69.35 | WHRT | 99.21 | 88.72 | 66.49 | 86.21 | 80.21 | 66.22 | GMM | 99.66 | 86.88 | 80.08 | 96.71 | 89.97 | 74.22 |
|
Table 3. Experimental accuracy in classification of polarimetric SAR%
Feature sets | Algorithm | Water | Forest | Farmland | Urban | Overall | Kappa |
---|
Cloude-Pottierdecomposition | KNN | 82.73 | 95.54 | 74.59 | 15.59 | 69.87 | 61.74 | SVM | 77.40 | 92.23 | 79.83 | 15.67 | 69.73 | 61.63 | RF | 84.07 | 94.87 | 83.04 | 33.84 | 77.79 | 68.75 | GMM | 88.00 | 76.48 | 82.67 | 82.42 | 79.86 | 70.58 | Freemandecomposition | KNN | 92.56 | 90.39 | 87.51 | 37.41 | 78.87 | 70.70 | SVM | 70.20 | 89.51 | 77.13 | 34.24 | 75.38 | 69.22 | RF | 93.01 | 86.06 | 88.88 | 51.15 | 81.00 | 71.58 | GMM | 94.53 | 67.57 | 89.77 | 85.03 | 82.35 | 72.78 | Yamaguchidecomposition | KNN | 86.61 | 89.05 | 69.76 | 70.43 | 80.09 | 70.78 | SVM | 89.31 | 88.31 | 67.78 | 65.60 | 75.89 | 68.65 | RF | 86.38 | 87.68 | 77.03 | 72.00 | 80.69 | 71.31 | GMM | 87.58 | 84.50 | 70.71 | 92.95 | 82.23 | 72.61 |
|
Table 4. Experimental accuracy in classification of polarimetric SAR under different features in Flevoland%