Fig. 1. Dictionary visualization expression. (a) Example 1; (b) example 2
Fig. 2. Model based on SC and CNN
Fig. 3. Algorithm flowchart of landform scene classification based on SC and CNN
Fig. 4. Feature visualization of image blocks with size of 14×14 on two databases by using SC. (a) UAV landform database 3, before feature sorting; (b) UAV landform database 3, after feature sorting; (c) UC Merced LU database, before feature sorting (d) UC Merced LU database, after feature sorting
Fig. 5. Training convergence curves for UC Merced LU database. (a) Training convergence curves obtained with four different methods; (b) training convergence curves obtained with SC-CNN algorithm
Fig. 6. Confusion matrix obtained by classify landforms with SC-CNN algorithm SC-CNN
Fig. 7. Classification effect maps of complex landforms image. (a) Landform image; (b) artificial landform division; (c) post-blocking image; (d) landform classification effect map
Layer | Type | Patch size | Stride | Zero padding | Output size |
---|
x | Input | | | | 256×256×3 | h1 | Convolution | 5×5 | 5 | 2 | 128×128×64 | h2 | ReLU | | | | 128×128×64 | h3 | Mean pooling | 3×3 | 2 | | 64×64×64 | h4 | Convolution | 3×3 | 2 | 0 | 32×32×64 | h5 | ReLU | | | | 32×32×64 | h6 | Max pooling | 3×3 | 2 | | 16×16×64 | h7 | Convolution | 7×7 | 1 | 2 | 14×14×128 | h8 | ReLU | | | | 14×14×128 | h9 | Max pooling | 3×3 | 2 | | 7×7×128 | h10 | Convolution | 7×7 | 1 | 0 | 1×1×128 | h11 | ReLU | | | | 1×1×128 | h12 | Convolution | 1×1 | 1 | 0 | 1×1×20 | o | SVM | | | | v |
|
Table 1. Network structure of model based on SC and CNN
Algorithm | Training accuracy /% | Training time /h |
---|
SVM | 78.57 | 0.7 | CS-CNN[12] | 92.86 | 4.5 | PSR[13] | 89.10 | 5.2 | MS-DCNN[11] | 91.34 | 5.9 | SC-CNN | 98.14 | 4.3 |
|
Table 2. Classification accuracy of different algorithms on UC Merced LU database
Algorithm | Training accuracy /% | Training time /h |
---|
SVM | 76.96 | 2.6 | CS-CNN[12] | 92.91 | 12.9 | MS-DCNN[11] | 91.53 | 13.7 | PCANet[14] | 86.49 | 11.1 | SC-CNN | 97.50 | 10.5 |
|
Table 3. Classification accuracy of existing methods on UAV landform database 3