Author Affiliations
1National Key Laboratory of Remote Sensing Information and Image Analysis Technology, Beijing Research Institute of Uranium Geology, Beijing 100029, China2Iflytek Intelligent Information Technology Co., Ltd., Hefei, Anhui 230094, China3Zachry Department of Civil and Environmental Engineering, Texas A & M University, Texas 77843, USA;show less
Fig. 1. Diagrams of (a) traditional convolution and (b) dilated convolution. (a) Traditional convolution; (b) dilated convolution
Fig. 2. Structural diagram of PPM
Fig. 3. Structural diagram of CBAM
Fig. 4. Illustration of network architecture
Fig. 5. Fusion of features with different scales
Fig. 6. Illustration of data augmentation
Fig. 7. Comparison of results before and after post-processing by proposed method
Fig. 8. Comparison of examples of network ablation models. (a) Image 1; (b) image 2; (c) ground truth; (d) proposed network; (e) ablation model A; (f) ablation model B; (g) ablation model C; (h) enlarged drawing in box
Fig. 9. Results of detection for ordered building change. (a) Image 1; (b) image 2; (c) ground truth; (d) UNet; (e) DeepLab; (f) CSCDNet; (g) UPerNet; (h) proposed network
Fig. 10. Results of detection for multi-scale building change. (a) Ground truth; (b) UNet;(c) DeepLab; (d) CSCDNet; (e) UPerNet; (f) proposed network
ResNet101 convolutional layer | Stage name | Output feature | Output scale |
---|
7×7,64,stride 2 | conv1 | C1 | 1/2 | 3×3,max pooling,stride 2 | | | | ×3 | conv2 | C2 | 1/4 | ×4 | conv3 | C3 | 1/8 | ×23(DC) | conv4 | C4 | 1/8 | ×3 (DC) | conv5 | C5 | 1/8 |
|
Table 1. Architecture of feature extraction network in encoding stage
Ablation experiment | DC | CBAM | PPM | P /% | R /% | F1 /% |
---|
A | O | P | P | 85.16 | 83.10 | 84.11 | B | P | O | P | 83.94 | 87.07 | 85.47 | C | P | P | O | 84.54 | 86.71 | 85.61 | Ours | P | P | P | 84.47 | 88.10 | 86.25 |
|
Table 2. Ablation experiment analysis of network
Method | P /% | R /% | F1 /% |
---|
UNet | 73.09 | 42.84 | 54.02 | DeepLab | 77.73 | 51.41 | 61.89 | CSCDNet | 81.08 | 69.60 | 74.90 | UPerNet | 78.66 | 71.99 | 75.18 | Ours | 84.47 | 88.10 | 86.25 |
|
Table 3. Accuracy assessment of building change detection by different methods