基于多层深度特征融合的极化合成孔径雷达图像语义分割 下载: 1135次
Semantic Segmentation of Polarimetric Synthetic Aperture Radar Images Based on Multi-Layer Deep Feature Fusion
空军工程大学信息与导航学院, 陕西 西安 710077
图 & 表
图 1. 原始数据。(a) Oberpfaffenhofen数据;(b) Flevoland数据
Fig. 1. Original data. (a) Oberpfaffenhofen data; (b) Flevoland data
下载图片 查看原文
图 2. Flevoland数据分类结果对比。(a)地物分布参考图[19];(b)方法1;(c)方法2;(d)方法3;(e)方法4;(f)所提方法
Fig. 2. Comparison of Flevoland data classification results. (a) Ground truth[19]; (b) method 1; (c) method 2; (d) method 3; (e) method 4; (f) proposed method in this paper
下载图片 查看原文
图 3. Oberpfaffenhofen数据分类结果对比。(a)地物分布参考图[20];(b)方法1;(c)方法2;(d)方法3;(e)方法4;(f)所提方法
Fig. 3. Comparison of Oberpfaffenhofen data classification results. (a) Ground truth[20]; (b) method 1; (c) method 2; (d) method 3; (e) method 4; (f) proposed method in this paper
下载图片 查看原文
表 1对比方法中用到的特征
Table1. Features used in comparison methods
Cloude decomposition | Freeman decomposition | Covariance matrix diagonal element |
---|
H,α,A,λ1,λ2,λ3 | Ps,Pd,Pv | C11,C22,C33 |
|
查看原文
表 2Flevoland数据下不同方法的性能对比
Table2. Performance comparison of different methods under Flevoland data
Class | Classificationaccuracy formethod 1 /% | Classificationaccuracy formethod 2 /% | Classificationaccuracy formethod 3 /% | Classificationaccuracy formethod 4 /% | Classificationaccuracy for theproposed methodin this paper /% |
---|
Beans | 97.12 | 83.33 | 96.70 | 89.84 | 87.84 | Forest | 75.91 | 94.02 | 73.33 | 92.22 | 89.22 | Potato | 68.77 | 84.08 | 82.14 | 95.55 | 87.13 | Alfalfa | 60.99 | 89.22 | 71.92 | 95.15 | 96.51 | Wheat | 93.40 | 88.15 | 0.864 | 92.95 | 98.94 | Bare land | 51.42 | 87.13 | 90.38 | 99.89 | 91.26 | Beet | 91.34 | 90.34 | 89.55 | 90.56 | 84.84 | Rapeseed | 57.23 | 78.29 | 62.70 | 93.29 | 91.40 | Pea | 58.91 | 82.14 | 82.00 | 98.79 | 95.59 | Grass | 96.28 | 77.43 | 83.89 | 83.20 | 94.26 | Water | 72.14 | 97.04 | 52.63 | 94.90 | 99.41 | OA /% | 75.15 | 87.04 | 77.87 | 93.38 | 92.22 | Training time /s | 579 | 565 | 632 | 2851 | 1266 | Test time /s | 3.20 | 3.00 | 3.40 | 13.20 | 8.50 |
|
查看原文
表 3Oberpfaffenhofen数据下不同方法的性能对比
Table3. Performance comparison of different methods under Oberpfaffenhofen data
Class | Classificationaccuracy formethod 1 /% | Classificationaccuracy formethod 2 /% | Classificationaccuracy formethod 3 /% | Classificationaccuracy formethod 4 /% | Classificationaccuracy for theproposed methodin this paper /% |
---|
Building area | 69.65 | 64.52 | 71.25 | 85.60 | 92.22 | Woodland | 89.54 | 89.60 | 70.00 | 93.13 | 83.94 | Open area | 62.23 | 84.32 | 87.41 | 95.42 | 94.86 | OA /% | 69.15 | 80.41 | 80.00 | 92.61 | 89.10 | Training time /s | 684 | 659 | 779 | 3398 | 1476 | Test time /s | 3.40 | 3.30 | 3.80 | 16.90 | 9.70 |
|
查看原文
表 4不同特征组合时的分类精度
Table4. Classification with different combination of features%
Conv3-3 | Conv4-3 | Conv5-3 | Conv3-3+ Conv4-3 | Conv3-3+ Conv5-3 | Conv4-3+ Conv5-3 | Proposed methodin this paper |
---|
80.88 | 85.47 | 87.02 | 87.66 | 89.97 | 90.11 | 92.22 |
|
查看原文
胡涛, 李卫华, 秦先祥. 基于多层深度特征融合的极化合成孔径雷达图像语义分割[J]. 中国激光, 2019, 46(2): 0210001. Tao Hu, Weihua Li, Xianxiang Qin. Semantic Segmentation of Polarimetric Synthetic Aperture Radar Images Based on Multi-Layer Deep Feature Fusion[J]. Chinese Journal of Lasers, 2019, 46(2): 0210001.