Fig. 1. Overall flowchart of the proposed method
Fig. 2. Dilated convolution with different dilation rate
Fig. 3. Multiscale dilated convolution module
Fig. 4. Multi-level attention feature fusion module
Fig. 5. Spatial attention module
Fig. 6. CM on UCM dataset with a training ratio of 80%
Fig. 7. Typical samples of forest and parking lot categories and misclassification samples
Fig. 8. CM on AID dataset with a training ratio of 50%
Fig. 9. Samples from river、port and bridge scene categories of AID dataset
Fig. 10. CM on PatternNet dataset with a training ratio of 20%
Datasets | Images per class | Total images | Images size | Scene classes | Spatial resolution/m |
---|
UCM | 100 | 2 100 | 256256 | 21 | 0.3 | AID | 220~420 | 10 000 | 600600 | 30 | 1~8 | PatternNet | 800 | 30 400 | 256256 | 38 | 0.062~0.493 |
|
Table 1. Scene dataset information
d | OA(20%) |
---|
128 | 93.02 | 256 | 93.08 | 512 | 93.11 | 1 024 | 93.70 | 2 048 | 93.23 |
|
Table 2. Influence of parameter d on classification performance of AID dataset
Architecture | Methods | OA(20%) |
---|
1 | Without MDC module | 93.280.33 | 2 | Without MAFF module | 91.230.25 | 3 | Without CBF | 92.910.22 | 4 | ResNet50+MDC+MAFF+CBF(Ours) | 93.700.11 |
|
Table 3. Ablation experiment on AID dataset with a training ratio of 20%
Methods | OA(80%) | OA(50%) |
---|
GBNet[16] | 98.570.48 | 97.050.19 | ARCNet[17] | 99.120.40 | 96.810.14 | Fusion by Addition[18] | 94.721.79 | | Fine-tuned ResNet50[19] | 91.90 | 89.43 | Siamese ResNet50[19] | 94.29 | 90.95 | Two-stream fusion[20] | 98.021.03 | 96.970.75 | CNN-CapsNet[14] | 99.050.24 | 97.590.16 | Ours | 99.320.20 | 98.130.18 |
|
Table 4. Classification result comparison on UCM dataset
Methods | OA(50%) | OA(20%) |
---|
ResNet50[21] | 91.310.58 | 88.230.70 | GBNet[16] | 94.580.12 | 92.200.23 | ARCNet[17] | 93.100.55 | 88.750.40 | Two-stream fusion[20] | 94.580.25 | 92.320.41 | Ours | 95.840.26 | 93.700.11 |
|
Table 5. Classification result comparison on AID dataset
Methods | OA(50%) | OA(20%) |
---|
GLANet(SVM)[22] | 99.400.21 | 98.91±0.19 | SDAResNet[23] | 99.580.10 | 99.30±0.08 | VGGNet(SVM)[24] | | 97.5±0.02 | ResNet101(SVM)[24] | | 98.6±0.02 | Inception-V3(SVM)[24] | | 97±0.02 | Ours | 99.600.06 | 99.42±0.05 |
|
Table 6. Classification result comparison on PatternNet dataset