Author Affiliations
1 Graduate School, Space Engineering University, Beijing 101416, China2 63883 Troops, Luoyang, Henan 471000, China3 School of Space Information, Space Engineering University, Beijing 101416, China4 Department of Electronic and Optical Engineering, Space Engineering University, Beijing 101416, Chinashow less
Fig. 1. Comparison of FCN and CNN based on image blocks. (a) CNN; (b) FCN
Fig. 2. Architecture of model system
Fig. 3. Class distribution of DSTL dataset
Fig. 4. Labels of image in DSTL dataset
Fig. 5. Training process. (a) Accuracy; (b) loss function; (c) Jaccard_coef; (d) Jaccard_coef_int
Fig. 6. Results predicted by model
Fig. 7. Relationship between evaluation value and threshold of each class
Fig. 8. Experimental results by proposed method. (a)-(c) Ground truths; (d)-(f) results with adaptive threshold; (g)-(i) results without adaptive threshold
Fig. 9. Experimental results by proposed method. (a)-(c) Results of Patch-based CNN model; (d)-(f) results with adaptive threshold and without data augmentation; (g)-(i) results without adaptive threshold and without data augmentation
Fig. 10. Experimental results of small class. (a)(b) Original images; (c)(d) ground truths; (e)(f) results of proposed method; (g)(h) results of basic U-Net model
Fig. 11. Comparison of algorithm performance
Band | Radiometric resolution /bit | Spatial resolution /m | Size /(pixel×pixel) |
---|
RGB+P (450-690 nm) | 11 | 0.31 | 3348×3392 | M band (400-1040 nm) | 11 | 1.24 | 837×848 | A band (1195-2365 nm) | 14 | 7.50 | 134×136 |
|
Table 1. Specifications of DSTL dataset at different bands
Class | buildings | misc | road | track | trees | crops | waterway | standing water | vehicle large | vehicle small |
---|
Threshold | 0.36 | 0.22 | 0.51 | 0.26 | 0.45 | 0.39 | 0.57 | 0.61 | 0.31 | 0.18 |
|
Table 2. Best threshold of each class
Algorithm | buildings | misc | road | track | trees | crops | waterway | standing water | vehicle large | vehicle small | Average |
---|
Binary logistic | 0.484 | 0.015 | 0.500 | 0.201 | 0.393 | 0.539 | 0.575 | 0 | 0 | 0 | 0.271 | Patch-based CNN | 0.623 | 0.146 | 0.756 | 0.258 | 0.671 | 0.958 | 0.873 | 0.800 | 0.001 | 0.030 | 0.512 | FCN+DA | 0.698 | 0.152 | 0.863 | 0.455 | 0.728 | 0.969 | 0.917 | 0.830 | 0.359 | 0.166 | 0.614 | FCN+WCE+DA | 0.701 | 0.181 | 0.863 | 0.467 | 0.728 | 0.969 | 0.918 | 0.834 | 0.366 | 0.183 | 0.621 | FCN+AT+DA | 0.702 | 0.228 | 0.863 | 0.465 | 0.728 | 0.969 | 0.918 | 0.832 | 0.362 | 0.226 | 0.629 | FCN+WCE+AT | 0.599 | 0.138 | 0.568 | 0.219 | 0.451 | 0.878 | 0.829 | 0.707 | 0.042 | 0.047 | 0.448 | Proposed method | 0.705 | 0.258 | 0.864 | 0.472 | 0.728 | 0.969 | 0.919 | 0.835 | 0.369 | 0.238 | 0.636 |
|
Table 3. Comparison of algorithm performance (Jaccard index)