Author Affiliations
1College of Missile Engineering, Rocket Force Engineering University, Xi’an 710025, Shaanxi , China2The 32023 Unit of the People’s Liberation Army, Dalian 116085, Liaoning , Chinashow less
Fig. 1. Network structure of DeepLabv3+ basic model
Fig. 2. Model structure of improved DeepLabv3+
Fig. 3. Hollow convolution of fusion of different receptive fields. (a) Channel stitching; (b) sampling point distribution of r=12 convolutional layer in original feature map; (c) sampling point distribution of r=12 convolutional layer in r=6 feature map
Fig. 4. Optimization of intermediate flow structure of backbone network
Fig. 5. Structure of channel attention module
Fig. 6. Part of training samples
Fig. 7. Basic structure of deep convolutional generative confrontation network
Fig. 8. Changes in different scenarios and time periods. (a) Scene 1; (b) scene 2
Fig. 9. Accuracy curve and loss curve of DeepLabv3+ network. (a) Accuracy curve; (b) loss curve
Fig. 10. Accuracy curve and loss curve of improved DeepLabv3+ network. (a) Accuracy curve; (b) loss curve
Fig. 11. Change detection results of DeepLabv3+ (left) and improved DeepLabv3+ (right). (a) Scene 1; (b) scene 2
Fig. 12. Landsat 8 test images. (a) Scene 3; (b) scene 4
Fig. 13. DeepLabv3+ (left) and improved DeepLabv3+ (right) detection results. (a) Scene 3; (b) scene 4
Fig. 14. Part of images of OSCD dataset. (a) Scene 5; (b) scene 6
Fig. 15. DeepLabv3+ (left) and improved DeepLabv3+ (right) change detection results. (a) Scene 5; (b) scene 6
Expansion rate r | Effective operation element | Receptive field | Information utilization |
---|
6 | 9 | 13 | 5.33 | 12 | 9 | 25 | 1.44 | 18 | 9 | 37 | 0.66 | 12+6 | 49 | 37 | 3.58 | 18+12 | 81 | 61 | 2.18 |
|
Table 1. Effect of fusion of different receptive fields on convolution of holes
Evaluation index(EI) | Improved DeepLabv3+ | DeepLabv3+ | Literature[11] | Literature[12] | Literature[13] |
---|
Kappa coefficient(Kappa) | 0.75 | 0.64 | 0.58 | 0.19 | 0.35 | Overall accuracy(OA) | 95.1% | 93.6% | 92.8% | 95.1% | 87.7 | Omission rate(OR) | 4.6% | 5.2% | 3.4% | 78.1% | 5.0% | Error rate(ER) | 5.6% | 18.0% | 21.2% | 2.5% | 37.2% | Sensitivity(SS) | 94.4% | 81.2% | 68.6% | 21.8% | 65.9% | Specificity(SP) | 95.2% | 94.9% | 94.0% | 97.4% | 94.2% | Balance accuracy(BA) | 94.8% | 88.1% | 81.3% | 59.6% | 80.1% | F1-score(F1) | 77.8% | 69.7% | 70.1% | 21.5% | 64.3% | Time required /s | 12.71 | 12.65 | 16.06 | 1.10 | 0.26 |
|
Table 2. Comparison of change detection results of different methods
Evaluation index(EI) | Kappa | OA% | OR% | ER% | SP% | BA% | FI% |
---|
DeepLabv3+ | 0.41 | 95.7 | 48.6 | 3.1 | 96.9 | 74.1 | 43.1 | Improved DeepLabv3+ | 0.56 | 96.4 | 20.0 | 3.1 | 96.8 | 88.4 | 57.7 |
|
Table 3. Change detection results of deep convolution method based on Landsat 8 data
Evaluation index(EI) | Kappa | OA% | OR% | ER% | SP% | BA% | FI% |
---|
DeepLabv3+ | 0.39 | 75.2 | 8.7 | 35.5 | 83.7 | 76.8 | 39.0 | Improved DeepLabv3+ | 0.44 | 83.6 | 5.3 | 27.2 | 89.1 | 80.3 | 44.8 |
|
Table 4. Comparison of depth change detection results based on OSCD data