Table 3.
The classification accuracy (OA (%)) of the linear classification of fused features with Principal Component Analysis (PCA) transformation, with 80% and 50% of UC-Merced dataset as a training set.
| Method | 80% of UCM Dataset as a Training Set | 50% of UCM Dataset as a Training Set |
|---|---|---|
| ResNet50 last conv layer (PCA) + InceptionV3 avg pooling | 97.14 | 97.33 |
| ResNet50 last conv layer (PCA) + Xception avg pooling | 97.62 | 97.43 |
| DenseNet121 conv5_block16_concat (PCA) + Xception avg pooling | 97.86 | 96.67 |
| DenseNet121 conv4_block24_concat (PCA) + Xception avg pooling | 97.86 | 96.57 |
| InceptionV3 mixed_10 (PCA) + ResNet50 avg pooling | 97.62 | 96.57 |
| InceptionV3 mixed_8 (PCA) + ResNet50 avg pooling | 98.33 | 97.43 |
| InceptionV3 mixed_10 (PCA) + Xception avg pooling | 95.95 | 95.14 |
| InceptionV3 mixed_8 (PCA) + Xception avg pooling | 98.57 | 97.62 |
| DenseNet121 conv5_block16_concat (PCA) + ResNet50 avg pooling | 97.14 | 96.67 |
| DenseNet121 conv4_block24_concat (PCA) + ResNet50 avg pooling | 96.9 | 95.24 |
| Xception block14_sepconv2_act (PCA) + DenseNet121 avg pooling | 96.67 | 96.48 |
| Xception block14_sepconv1_act (PCA) + DenseNet121 avg pooling | 98.57 | 96.29 |