Author Affiliations
1College of Software, Liaoning Technical University, Huludao 125105, Liaoning, China2Quanzhou Institute of Equipment Manufacturing Haixi Institutes, Fujian Institute of Research on the Structure, Chinese Academy of Sciences, Quanzhou 362216, Fujian, Chinashow less
Fig. 1. Diagram of dictionary learning
Fig. 2. Flowchart of the proposed method
Fig. 3. Batch normalization and layer normalization
Fig. 4. Schematic of multilayer perceptron
Fig. 5. Flowchart of attention module method
Fig. 6. Attention module based on dictionary learning
Fig. 7. RSSCN7 dataset
Fig. 8. NWPU-RESISC45 dataset
Fig. 9. AID dataset
Fig. 10. Rate of change of classification accuracy on Gaussian noise images
Dataset | Number of scene classes | Number of total images | Image size | Spatial resolution /m | Year |
---|
RSSCN7 | 7 | 2800 | 400×400 | | 2015 | NWPU-RESISC45 | 45 | 31500 | 256×256 | ~30-0.2 | 2016 | AID | 30 | 10000 | 600×600 | ~8-0.5 | 2017 |
|
Table 1. Introduction of datasets
Laboratory environment | Environment configuration |
---|
Language | Python3.8.6 | Tool | PyCharm11.0.11 | Framework | PyTorch1.9.1 | CUDA | 10.2 |
|
Table 2. Laboratory environment
Network | Accuracy /% |
---|
AlexNet | 82.230 | VGG | 80.833 | ResNet50 | 89.048 | TNT | 84.833 | ViT | 89.643 | Proposed network | 91.406 |
|
Table 3. Accuracy of different networks on RSSCN7 dataset
Network | Accuracy /% |
---|
Fine-tuned AlexNet | 85.160 | Fine-tuned VGGNet-16 | 90.360 | Fine-tuned GoogLeNet | 86.020 | TNT | 85.031 | ViT | 90.255 | Proposed network | 91.576 |
|
Table 4. Accuracy of different networks on NWPU-RESISC45 dataset
Network | Accuracy /% |
---|
CaffeNet | 86.860 | VGG-VD-16 | 86.590 | ResNet152 | 89.130 | GoogLeNet | 83.440 | TNT | 80.450 | ViT | 85.514 | Proposed network | 89.218 |
|
Table 5. Accuracy of different networks on AID dataset
Parameter | RSSCN7 | NWPU-RESISC45 | AID |
---|
ViT | Proposed method | ViT | Proposed method | ViT | Proposed method |
---|
kappa | 0.900 | 0.916 | 0.934 | 0.947 | 0.883 | 0.909 | F1 | 86.222 | 90.890 | 88.927 | 90.207 | 84.202 | 87.768 | recall | 85.986 | 91.142 | 88.984 | 90.286 | 84.147 | 87.662 | precision | 86.417 | 91.002 | 89.039 | 90.317 | 84.558 | 88.004 |
|
Table 6. Parameter indicators of two methods on three datasets
Network | Number of parameter /106 |
---|
AlexNet | 6 | VGG | 13.3 | ResNet50 | 2.55 | TNT | 2.25 | ViT | 2.6 | Proposed method | 1.84 |
|
Table 7. Parameters of different classification frameworks