Acta Photonica Sinica, Volume. 50, Issue 7, 79(2021)
Remote Sensing Image Scene Classification Based on Supervised Contrastive Learning
Figures & Tables(13)
Fig. 2. The size of the feature maps on each key node in the model
Fig. 6. Histogram of classification results on AID dataset with training ratio of 20% and 50%
Fig. 7. Confusion matrix generated under 20% training ratio on the AID dataset
Fig. 8. Confusion matrix generated under 50% training ratio on the AID dataset
Fig. 9. Confusion matrix generated under 20% training ratio on the NWPU-RESISC45 dataset
Table 1. Data set description
View tableView in ArticleTable 1. Data set description
Dataset Images per category Number of categories Total images Size Training ratio AID 220~420 30 10 000 600×600 20%/50% NWPU-RESISC45 700 45 31 500 256×256 10%/20%
Table 2. Performance comparisons of different methods on AID and NWPU-RESISC45 datasets
View tableView in ArticleTable 2. Performance comparisons of different methods on AID and NWPU-RESISC45 datasets
Methods Year AID dataset NWPU-RESISC45 dataset 20% training ratio 50% training ratio 10% training ratio 10% training ratio D-CNN［ 2 ］2018 90.82±0.16 96.89±0.10 88.07±0.18 88.07±0.18 MSCP［ 18 ］2018 92.21±0.17 96.56±0.18 89.03±0.21 89.03±0.21 SF-CNN［ 23 ］2019 93.60±0.12 96.66±0.11 90.87±0.15 90.87±0.15 CNN-CapsNet［ 24 ］2019 93.79±0.13 96.63±0.12 91.48±0.19 91.48±0.19 MF2Net［ 25 ］2019 93.82±0.26 95.93±0.23 91.63±0.15 91.63±0.15 MG-CAP（Bil）［ 26 ］2020 92.11±0.15 95.14±0.12 88.48±0.21 88.48±0.21 DDRL-AM［ 27 ］2020 92.36±0.10 96.25±0.05 89.22±0.50 89.22±0.50 SCCov［ 28 ］2020 93.12±0.25 96.10±0.16 89.42±0.19 89.42±0.19 MG-CAP（Sqr）［ 26 ］2020 93.34±0.18 96.12±0.12 89.89±0.16 89.89±0.16 ResNet50EAM［ 29 ］2020 93.64±0.25 96.62±0.13 91.08±0.24 91.08±0.24 ResNet101EAM［ 29 ］2020 94.26±0.11 97.06±0.19 91.91±0.22 91.91±0.22 Ours -- 95.66 ±0.19 97.29 ±0.22 92.86 ±0.20 92.86 ±0.20
Table 3. Performance comparisons of different variants on two datasets
View tableView in ArticleTable 3. Performance comparisons of different variants on two datasets
Methods AID dataset NWPU-RESISC45 dataset 20% training ratio 50% training ratio 10% training ratio 20% training ratio SCRSISC-SC 94.52±0.13 96.26±0.11 91.83±0.11 93.78±0.12 SCRSISC-GSA 94.36±0.17 96.14±0.18 91.38±0.24 93.51±0.19 SCRSISC-Inc_v3 94.16±0.15 96.05±0.25 91.16±0.10 93.28±0.13 Ours（SCRSISC） 95.66 ±0.19 97.29 ±0.22 92.86 ±0.20 94.73 ±0.19
Table 4. Comparisons of efficiency and accuracy for two datasets at 20% training ratio
View tableView in ArticleTable 4. Comparisons of efficiency and accuracy for two datasets at 20% training ratio
Methods AID dataset NWPU-RESISC45 dataset Times/min Accuracy Times/min Accuracy ResNet50EAM［ 29 ］61 93.64±0.25 168 93.49±0.17 ResNet101EAM［ 29 ］109 94.26±0.11 282 94.29±0.09 Ours （SCRSISC） 45 95.66 ±0.19 126 94.73 ±0.19
Tools
Get Citation
Copy Citation Text
Dongen GUO, Ying XIA, Xiaobo LUO, Jiangfan FENG. Remote Sensing Image Scene Classification Based on Supervised Contrastive Learning[J]. Acta Photonica Sinica, 2021, 50(7): 79
Paper Information
Category: Image Processing
Received: Dec. 12, 2020
Accepted: Mar. 4, 2021
Published Online: Sep. 1, 2021
The Author Email:
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20