Author Affiliations
1 School of Electronics and Information Engineering, Liaoning Technical University, Huludao, Liaoning 125105, China2 Graduate School, Liaoning University of Engineering and Technology, Huludao, Liaoning 125105, China3 Liaoning Unicom Fuxin Branch, Fuxin, Liaoning 123100, Chinashow less
Fig. 1. Schematic of G-CNN classification model
Fig. 2. Relationship between spectral and spatial information of hyperspectral images
Fig. 3. Super-edge construction based on spectral features
Fig. 4. Hypergraph construction of spatial relationship. (a) 4 neighborhood; (b) 8 neighborhood; (c) 16 neighborhood; (d) 24 neighborhood
Fig. 5. Schematic of CNN classification model
Fig. 6. Effect of number of training samples on accuracy of algorithm
Fig. 7. Influence of experimental parameter change on experimental precision. (a) P; (b) U; (c) δ
Fig. 8. Classification results by different algorithms on Indian Pines dataset. (a) Ground-truth; (b) SVM; (c) G-SVM; (d) Shallower CNN; (e) Contextual Deep CNN; (f) SPPF CNN; (g) G-CNN; (h) label
No. | Class | Number oftraining samples | Number ofteat samples |
---|
1 | Corn-notill | 200 | 1228 | 2 | Corn-mintill | 200 | 630 | 3 | Grass-pasture | 200 | 283 | 4 | Hay-windrowed | 200 | 278 | 5 | Soybean-notill | 200 | 772 | 6 | Soybean-mintill | 200 | 2255 | 7 | Soybean-clean | 200 | 393 | 8 | Woods | 200 | 1065 | Total | 1600 | 6904 |
|
Table 1. Number of training samples and test samples in Indian Pines dataset
Dataset | t1 | s1 | t2 | s2 | t3 | s3 | t4 | s4 | N6 | N7 |
---|
Indian Pines | 21 | 1 | 2 | 2 | 21 | 1 | 2 | 2 | 100 | 8 | University of Pavia | 10 | 1 | 2 | 2 | 10 | 1 | 2 | 2 | 100 | 9 | Salinas | 21 | 1 | 2 | 2 | 21 | 1 | 2 | 2 | 100 | 16 |
|
Table 2. Parameter setting of convolutional neural network
Item | SVM | G-SVM | Shallower CNN | Contextual Deep CNN | SPPF CNN | G-CNN |
---|
Training time | 0.21 | 4.74 | 287.16 | 431.71 | 2704.34 | 337.08 | Testing time | 1.12 | 0.92 | 0.23 | 3.35 | 19.57 | 0.29 |
|
Table 3. Training time and test time for each algorithm on Indian Pines datasets
Algorithm | IndianPines | Universityof Pavia | Salinas |
---|
SVM | 85.86 | 86.83 | 87.12 | G-SVM | 89.40 | 90.91 | 90.97 | Shallower CNN[12] | 90.12 | 92.34 | 92.18 | Contextual Deep CNN[13] | 93.30 | 95.01 | 95.47 | SPPF CNN[14] | 94.87 | 91.74 | 94.21 | G-CNN | 96.63 | 97.42 | 97.27 |
|
Table 4. Overall classification accuracy of each model on three datasets%
Parameter | IndianPines | Universityof Pavia | Salinas |
---|
OA | 96.63 | 97.42 | 97.27 | AA | 96.87 | 97.83 | 97.44 | Kappa | 96.54 | 97.18 | 97.12 |
|
Table 5. G-CNN classification results%