Author Affiliations
1Aviation University of Air Force, Changchun, Jilin 130022, China2School of Geographic Science, Northeast Normal University, Changchun, Jilin 130024, China395910 Troop of PLA, Jiuquan, Gansu 735000, China495795Troop of PLA, Guilin, Guangxi 541000, Chinashow less
Fig. 1. Process of two-dimensional convolution
Fig. 2. Process of three-dimensional convolution
Fig. 3. Structure of LSTM network
Fig. 4. Structure of ladder network
Fig. 5. Structure of 3D-CNN-LSTM network
Fig. 6. Model of semi-supervised classification algorithm based on improved ladder network
Fig. 7. Structure of 3D-CNN-LSTM ladder network
Fig. 8. True color image and ground truth map of Pavia University dataset. (a) True color image; (b) ground truth map
Fig. 9. True color image and ground truth map of Indian Pines dataset. (a) True color image; (b) ground truth map
Fig. 10. Classification results of different algorithms on Pavia University dataset. (a) True color image; (b) feature label map; (c) M3D-DCNN algorithm; (d) 3D-CNN-LSTM algorithm; (e) SS-CNN algorithm; (f) S4CNN algorithm; (g) 3D-CNN-LSTM-LN algorithm
Fig. 11. Classification results of different algorithms on Indian Pines dataset. (a) True color image; (b) feature label map; (c) M3D-DCNN algorithm; (d) 3D-CNN-LSTM algorithm; (e) SS-CNN algorithm; (f) S4CNN algorithm; (g) 3D-CNN-LSTM-LN algorithm
Layer | Kernel size | Input size | Output size | Number of convolution kernels | Activation function |
---|
Input | -- | -- | 11×11×30 | -- | -- | Conv-1(Conv 3D) | 3×3×1 | 11×11×30 | 9×9×30 | 4 | -- | BN-1 | -- | 9×9×30 | 9×9×30 | -- | ReLU | Pooling-1(MaxPooling) | 1×1×2 | 9×9×30 | 9×9×15 | -- | -- | Conv-2(Conv 3D) | 3×3×1 | 9×9×15 | 7×7×15 | 8 | -- | BN-2 | -- | 7×7×15 | 7×7×15 | -- | ReLU | Pooling-2(MaxPooling) | 1×1×2 | 7×7×15 | 7×7×7 | -- | -- | Conv-3(Conv 3D) | 3×3×1 | 7×7×7 | 5×5×7 | 16 | -- | BN-3 | -- | 5×5×7 | 5×5×7 | -- | ReLU | Pooling-3(MaxPooling) | 1×1×2 | 5×5×7 | 5×5×3 | -- | -- | Linear-1(flatten) | -- | 5×5×3 | 1×75 | -- | -- | RNN-1(Bi-LSTM) | -- | 1200×1 | 400×1 | -- | Tanh |
|
Table 1. Network structure of encoder in 3D-CNN-LSTM-LN model
No. | Class | M3D-DCNN | 3D-CNN-LSTM | SS-CNN | S4CNN | 3D-CNN-LSTM-LN |
---|
1 | Asphalt | 93.25 | 93.03 | 94.98 | 96.94 | 99.21 | 2 | Meadow | 98.14 | 96.99 | 96.84 | 99.11 | 99.63 | 3 | Gravel | 76.83 | 75.13 | 89.16 | 86.97 | 96.22 | 4 | Tree | 97.78 | 91.45 | 93.50 | 98.33 | 98.94 | 5 | Metal sheet | 100.00 | 96.63 | 99.33 | 99.78 | 99.66 | 6 | Bare soil | 92.87 | 90.31 | 92.28 | 97.40 | 99.18 | 7 | Bitumen | 80.56 | 86.97 | 83.21 | 91.76 | 97.79 | 8 | Brick | 83.14 | 83.95 | 90.92 | 93.50 | 98.13 | 9 | Shadow | 94.25 | 88.52 | 96.84 | 97.00 | 98.55 | OA | -- | 93.73 | 92.46 | 94.46 | 97.06 | 99.03 | AA | -- | 90.62 | 89.13 | 92.96 | 95.64 | 98.59 | Kappa | -- | 91.70 | 89.90 | 92.80 | 96.20 | 98.80 |
|
Table 2. Classification accuracy of different algorithms on Pavia University dataset unit: %
No. | Class | M3D-DCNN | 3D-CNN-LSTM | SS-CNN | S4CNN | 3D-CNN-LSTM-LN |
---|
1 | Alfalfa | 39.12 | 77.65 | 81.91 | 97.23 | 96.85 | 2 | Corn-notill | 77.95 | 82.45 | 83.24 | 91.77 | 96.93 | 3 | Corn-mintill | 88.67 | 82.36 | 74.34 | 89.34 | 93.97 | 4 | Corn | 87.02 | 66.75 | 79.67 | 96.38 | 97.15 | 5 | Grass-pasture | 95.43 | 93.17 | 97.65 | 92.93 | 95.49 | 6 | Grass-tree | 98.78 | 95.36 | 88.96 | 99.36 | 98.17 | 7 | Grass-pasture-mowed | 70.31 | 88.95 | 75.02 | 97.72 | 62.13 | 8 | Hay-windrowed | 95.33 | 97.34 | 89.61 | 98.33 | 98.88 | 9 | Oat | 53.35 | 56.02 | 82.47 | 72.03 | 57.15 | 10 | Soybean-notill | 81.63 | 90.14 | 82.35 | 92.19 | 97.31 | 11 | Soybean-mintill | 86.27 | 88.46 | 82.63 | 92.84 | 98.19 | 12 | Soybean-clean | 80.68 | 74.28 | 83.28 | 83.96 | 92.56 | 13 | Wheat | 99.45 | 97.67 | 81.36 | 97.95 | 99.70 | 14 | Wood | 94.59 | 98.32 | 87.64 | 97.71 | 99.65 | 15 | Building-grass-tree-drive | 76.95 | 90.24 | 80.27 | 89.72 | 95.12 | 16 | Stone-steel-tower | 95.27 | 46.83 | 97.79 | 90.98 | 93.35 | OA | -- | 87.20 | 88.83 | 89.56 | 93.34 | 97.04 | AA | -- | 82.51 | 83.22 | 84.26 | 82.40 | 92.03 | Kappa | -- | 85.30 | 87.20 | 88.80 | 92.30 | 96.60 |
|
Table 3. Classification accuracy of different algorithms on Indian Pines dataset unit: %