Fig. 1. The framework of Generative Adversarial Networks
Fig. 2. The overall framework of the proposed method
Fig. 3. Illustration of the ISPRS 2D Vaihingen Labeling dataset (a) the entire remote sensing image, including near-infrared, red and green bands, (b) partial remote sensing image numbered 2, and (c) corresponding label map and its legend
Fig. 4. Illustration of cropping the entire image
Fig. 5. Illustration of confusion matrix
Fig. 6. Visual comparison of segmentation results among the proposed method and other state-of-the-art models on test set: (a) image for segmentation, (b) ground truth label map, (c) UPB, (d) ETH_C, (e) CAS_Y3, (f) ITC_B2 (g) VNU4, (h) CASZX1, (i) UFMG_3, and (j) the proposed method
类型 | 比例 | Method | Imp Surf | Building | Low_veg | Tree | Car | OA | Mean F1 |
---|
有 监 督 | 1 | Baseline | 86.3 | 92.6 | 70.7 | 85.6 | 69.9 | 85.3 | 81.0 | | 88.8 | 93.9 | 75.6 | 87.9 | 77.6 | 87.7 | 84.7 | | 89.0 | 94.1 | 75.6 | 88.0 | 78.4 | 88.0 | 85.0 | 半 监 督 | 1/4 | Baseline | 83.0 | 90.7 | 62.7 | 83.7 | 62.0 | 82.0 | 76.4 | | 86.9 | 93.2 | 71.7 | 86.8 | 74.9 | 86.2 | 82.7 | | 87.8 | 93.8 | 73.2 | 87.3 | 75.0 | 86.8 | 83.4 | | 87.8 | 93.9 | 73.8 | 87.4 | 76.2 | 87.2 | 83.8 | 1/8 | Baseline | 80.0 | 87.9 | 60.1 | 81.0 | 49.1 | 79.1 | 71.8 | | 85.5 | 91.3 | 70.2 | 86.0 | 69.3 | 84.5 | 80.5 | | 86.4 | 92.1 | 71.6 | 86.6 | 72.9 | 85.5 | 81.9 | | 86.7 | 92.1 | 74.2 | 87.0 | 73.1 | 86.1 | 82.6 |
|
Table 1. 不同标签样本比例下各部分提升效果比较
比例 | Method | Imp Surf | Building | Low_veg | Tree | Car | OA | Mean F1 |
---|
1/4 | Baseline | 83.0 | 90.7 | 62.7 | 83.7 | 62.0 | 82.0 | 76.4 | SSGAN[19] | 83.6 | 91.0 | 63.9 | 84.1 | 65.7 | 82.9 | 77.7 | Semi-SegGAN[17] | 87.3 | 93.3 | 73.5 | 87.1 | 75.7 | 86.5 | 83.4 | 本文方法 | 87.8 | 93.9 | 73.8 | 87.4 | 76.2 | 87.2 | 83.8 | 1/8 | Baseline | 80.0 | 87.9 | 60.1 | 81.0 | 49.1 | 79.1 | 71.8 | SSGAN[19] | 81.1 | 88.3 | 62.5 | 82.0 | 54.3 | 81.6 | 73.6 | Semi-SegGAN[17] | 85.9 | 91.5 | 70.8 | 86.1 | 72.3 | 84.9 | 81.3 | 本文方法 | 86.7 | 92.1 | 74.2 | 87.0 | 73.1 | 86.1 | 82.6 |
|
Table 2. 与其它半监督语义分割方法在验证集上的结果对比
λadv | λsemi | Tsemi | OA | Mean F1 |
---|
0.01 | 0 | N/A | 85.5 | 81.9 | 0.01 | 0.01 | 0.2 | 85.6 | 82.0 | 0.01 | 0.1 | 0.2 | 86.1 | 82.6 | 0.01 | 0.2 | 0.2 | 85.8 | 82.3 | 0.01 | 0.1 | 0.1 | 85.7 | 82.1 | 0.01 | 0.1 | 0.2 | 86.1 | 82.6 | 0.01 | 0.1 | 0.3 | 85.9 | 82.4 | 0.001 | 0.1 | 0.2 | 85.4 | 81.8 | 0.01 | 0.1 | 0.2 | 86.1 | 82.6 | 0.1 | 0.1 | 0.2 | 84.9 | 80.9 |
|
Table 3. 超参数、和取值分析
| OA | Mean F1 |
---|
0 | 85.1 | 81.6 | 0.5 | 85.3 | 81.8 | 1 | 86.1 | 82.6 | 2 | 85.5 | 82.0 | 5 | 85.5 | 81.8 |
|
Table 4. 超参数取值分析
Method | Imp Surf | Building | Low_veg | Tree | Car | OA | Mean F1 |
---|
UPB[9] | 87.5 | 89.3 | 77.3 | 85.8 | 77.1 | 85.1 | 83.4 | ETH_C[25] | 87.2 | 92.0 | 77.5 | 87.1 | 54.4 | 85.9 | 79.6 | CAS_Y3[6] | 89.6 | 91.5 | 82.0 | 88.3 | 68.4 | 87.8 | 84.0 | ITC_B2[7] | 90.1 | 93.5 | 82.1 | 88.3 | 77.1 | 88.4 | 86.2 | VNU4[26] | 91.2 | 93.6 | 81.5 | 88.5 | 77.7 | 89.0 | 86.5 | CASZX1[27] | 91.3 | 93.9 | 81.9 | 88.3 | 77.6 | 89.0 | 86.6 | UFMG_3[28] | 90.7 | 94.3 | 82.5 | 88.5 | 77.4 | 89.0 | 86.7 | 所提议方法 | 92.7 | 95.1 | 84.3 | 89.4 | 86.2 | 90.6 | 89.5 |
|
Table 5. 与其它性能优异方法的测试集结果对比