Fig. 1. Flow chart of our method
Fig. 2. Segmentation results with different number of superpixels
Fig. 3. Extraction result of saliency map. (a) Original image; (b) K-means clustering; (c) superpixel segmentation; (d) fusion result
Fig. 4. Extraction result of the ROI. (a) Saliency map;(b) gray enhancement map; (c) binarization map; (d) ROI
Fig. 5. Structure of the DCNN
Fig. 6. Expanded results of the data. (a) Original image; (b) horizontal flip; (c) vertical flip; (d) brightness adjustment
Fig. 7. Loss function and classification accuracy of different models. (a) UC-Merced data set; (b) WHU-RS data set
Fig. 8. Images in different data sets. (a) UC-Merced data set; (b) WHU-RS data set
Fig. 9. Experimental results of the UC-Merced data set. (a) Original image; (b) saliency map; (c) gray enhancement map; (d) binarization map; (e) ROI
Fig. 10. Confusion matrix of our method on the UC-Merced data set
Fig. 11. Experimental results in the WHU-RS data set. (a) Original image; (b) saliency map; (c) gray enhancement map; (d) binarization map; (e) ROI
Fig. 12. Confusion matrix of our method on the WHU-RS data set
No. | Scene | Accuracy | No. | Scene | Accuracy | No. | Scene | Accuracy |
---|
1 | agricultural | 96 | 8 | forest | 95 | 15 | overpass | 100 | 2 | airplane | 93 | 9 | freeway | 100 | 16 | parking lot | 100 | 3 | baseball diamond | 100 | 10 | golf course | 100 | 17 | river | 100 | 4 | beach | 100 | 11 | harbor | 100 | 18 | runway | 96 | 5 | buildings | 84 | 12 | intersection | 100 | 19 | sparse residential | 100 | 6 | chaparral | 100 | 13 | medium residential | 84 | 20 | storage tanks | 88 | 7 | dense residential | 92 | 14 | mobile home park | 100 | 21 | tennis court | 90 | Average accuracy | 96.10 |
|
Table 1. Classification accuracy of the UC-Merced data set unit: %
Method | Accuracy /% | Time /s |
---|
Saliency + Multi-CNN | 92.86 | 2.27 | MS-DCNN | 91.34 | 5.90 | JMCNN | 88.30 | 0.81 | Our method | 96.10 | 1.95 |
|
Table 2. Classification accuracies of different methods on the UC-Merced data set
No. | Scene | Accuracy | No. | Scene | Accuracy | No. | Scene | Accuracy |
---|
1 | airport | 100 | 8 | football field | 100 | 15 | pond | 100 | 2 | beach | 100 | 9 | forest | 92 | 16 | port | 100 | 3 | bridge | 100 | 10 | industrial | 93 | 17 | rail way station | 93 | 4 | commercial | 91 | 11 | meadow | 86 | 18 | residential | 91 | 5 | desert | 100 | 12 | mountain | 100 | 19 | river | 92 | 6 | farmland | 91 | 13 | park | 100 | | | | 7 | viaduct | 92 | 14 | parking | 100 | | | | Average accuracy | 95.84 |
|
Table 3. Classification accuracy of the WHU-RS data set unit: %
Method | Accuracy /% | Time /s |
---|
Saliency + Multi-CNN | 91.80 | 3.13 | MS-DCNN | 90.05 | 7.33 | JMCNN | 87.63 | 0.98 | Our method | 95.84 | 2.32 |
|
Table 4. Classification accuracies of different methods on the WHU-RS data set