Acta Photonica Sinica, Volume. 50, Issue 11, 1128002(2021)
Scene Classification of Optical High-resolution Remote Sensing Images Using Vision Transformer and Graph Convolutional Network
Figures & Tables(6)
Table 1. Stds and overall accuracies of different methods with 20% and 50% training ratio in the AID dataset
View tableView in ArticleTable 1. Stds and overall accuracies of different methods with 20% and 50% training ratio in the AID dataset
Method 20% training ratio/% 50% training ratio/% CaffeNet［ 9 ］86.86±0.47 89.53±0.31 VGG-VD-16［ 9 ］86.59±0.29 89.64±0.36 GoogLeNet［ 9 ］83.44±0.40 86.39±0.55 Two-stream deep feature fusion［ 11 ］94.09±0.34 95.99±0.35 MSCP［ 12 ］91.52±0.21 94.42±0.17 mSmL-Gcoding［ 13 ］91.69±0.36 95.61±0.28 MDFR［ 14 ］90.62±0.27 93.37±0.29 SAFF［ 18 ］90.25±0.29 93.83±0.28 Scale-attention network［ 19 ］92.53±0.33 95.72±0.27 Proposed method 94.52 ±0.25 96.80 ±0.17
Table 2. Stds and overall accuracies of different methods with 10% and 20% training ratio in the NWPU-RESISC45 dataset
View tableView in ArticleTable 2. Stds and overall accuracies of different methods with 10% and 20% training ratio in the NWPU-RESISC45 dataset
Method 10% training ratio/% 20% training ratio/% CaffeNet［ 10 ］81.22±0.19 85.16±0.18 VGG-VD-16［ 10 ］87.15±0.45 90.36±0.18 GoogLeNet［ 10 ］82.57±0.12 86.02±0.18 Two-stream deep feature fusion［ 11 ］85.02±0.25 87.01±0.19 MSCP［ 12 ］85.33±0.17 88.93±0.14 mSmL-Gcoding［ 13 ］89.34±0.48 91.64±0.26 MDFR［ 14 ］83.37±0.26 86.89±0.17 SAFF［ 18 ］84.38±0.19 87.86±0.14 Scale-attention network［ 19 ］88.92±0.29 92.25±0.18 Proposed method 90.50 ±0.26 93.31 ±0.15
Table 3. Ablation experiments of the proposed method in the AID and NWPU-RESISC45 datasets
View tableView in ArticleTable 3. Ablation experiments of the proposed method in the AID and NWPU-RESISC45 datasets
Method AID
50% training ratio/%
NWPU-RESISC45
20% training ratio/%
Proposed method w/o ViT 89.90±0.50 88.10±0.21 Proposed method w/o GCN 94.20±0.22 91.70±0.18 Proposed method 96.80 ±0.17 93.31 ±0.15
Table 4. Overall accuracies of different methods with 20% training ratio in the GID dataset
View tableView in ArticleTable 4. Overall accuracies of different methods with 20% training ratio in the GID dataset
Method DenseNet-121 Proposed method w/o ViT Proposed method w/o GCN Proposed method OA/% 69.10 68.50 70.20 72.30
Table 5. FPS comparison of proposed method in the AID，NWPU-RESISC45 and GID datasets
View tableView in ArticleTable 5. FPS comparison of proposed method in the AID，NWPU-RESISC45 and GID datasets
Method AID
50% training ratio/%
NWPU-RESISC45
20% training ratio/%
GID
20% training ratio/%
DenseNet-121 63.4 63.5 24.4 Proposed method 135.6 138.1 139.3
Tools
Get Citation
Copy Citation Text
Jianan WANG, Yue GAO, Jun SHI, Ziqi LIU. Scene Classification of Optical High-resolution Remote Sensing Images Using Vision Transformer and Graph Convolutional Network[J]. Acta Photonica Sinica, 2021, 50(11): 1128002
Paper Information
Category: Remote Sensing and Sensors
Received: May. 25, 2021
Accepted: Jul. 26, 2021
Published Online: Dec. 2, 2021
The Author Email: GAO Yue (bjlguniversity@163.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20