结合均值漂移分割与全卷积神经网络的高分辨遥感影像分类 下载: 1437次
High Resolution Remote Sensing Image Classification Combining with Mean-Shift Segmentation and Fully Convolution Neural Network
1 中国矿业大学环境与测绘学院, 江苏 徐州 221116
2 国家测绘地理信息局卫星测绘应用中心, 北京 100048
图 & 表
图 1. FCN的基本结构
Fig. 1. Basic structure of FCN
下载图片 查看原文
图 2. 网络结构示意图
Fig. 2. Diagram of network structure
下载图片 查看原文
图 3. 区域合并示意图
Fig. 3. Diagram of regional consolidation
下载图片 查看原文
图 4. 分类实验流程图
Fig. 4. Flow chart of classification method
下载图片 查看原文
图 5. 原始影像与标签数据示例。(a)示例1;(b)示例2
Fig. 5. Original images and tag data examples. (a) Example 1; (b) example 2
下载图片 查看原文
图 6. 不同方法的分类结果。(a)原始实验影像;(b)均值漂移①分割结果;(c)均值漂移②分割结果;(d)均值漂移③分割结果;(e)真实分类图;(f) SVM分类结果;(g) ANN分类结果;(h) FCN-16分类结果;(i) FCN-8分类结果;(j) FCN分类结果;(k)加入均值漂移①分割结果的FCN分类结果;(l)加入均值漂移②分割结果的FCN分类结果;(m)加入均值漂移③分割结果的FCN分类结果
Fig. 6. Classification results of different methods. (a) Original image; (b) segmentation result of mean-shift ①;(c) segmentation result of mean-shift ②; (d) segmentation result of mean-shift ③; (e) true classification image;(f) classification result of SVM ; (g) classification result of ANN; (h) classification result of FCN-16; (i) classification result of FCN-8; (j) classification result of proposed FCN; (k) classification result of proposed FCN adding segmentation result of mean-shift ①; (l) classifi
下载图片 查看原文
图 7. 部分细节标注图。(a)真实分类图;(b) FCN-16分类结果;(c)加入均值漂移②分割结果的FCN分类结果
Fig. 7. Marked images of some details. (a) True classification image; (b) classification result of FCN-16; (c) classification result of proposed FCN adding segmentation result of mean-shift ②
下载图片 查看原文
表 1SVM分类方法的混淆矩阵与总体精度
Table1. Confusion matrice and overall accuracy of SVM classification method%
Type ofground | B | F | W | R | S | G |
---|
B | 81.7 | 8.1 | 2.1 | 45.2 | 64.1 | 45.8 | F | 3.8 | 70.0 | 15.1 | 8.8 | 7.2 | 6.9 | W | 0.5 | 0.4 | 79.6 | 0 | 0.1 | 0 | R | 11.6 | 14.4 | 2.5 | 43.6 | 13.3 | 4.5 | S | 0.7 | 0.1 | 0.1 | 0.2 | 6.7 | 0.8 | G | 1.7 | 7.1 | 0.7 | 2.1 | 8.6 | 42.0 | OA | 66.08 |
|
查看原文
表 2ANN分类方法的混淆矩阵与总体精度
Table2. Confusion matrice and overall accuracy of ANN classification method%
Type ofground | B | F | W | R | S | G |
---|
B | 82.0 | 11.2 | 2.5 | 51.1 | 72.5 | 53.0 | F | 4.4 | 68.3 | 12.0 | 7.4 | 8.1 | 32.5 | W | 0.4 | 3.7 | 83.4 | 0.6 | 1.9 | 2.8 | R | 13.2 | 16.8 | 2.2 | 40.9 | 17.5 | 11.7 | S | 0 | 0 | 0 | 0 | 0 | 0 | G | 0 | 0 | 0 | 0 | 0 | 0 | OA | 64.79 |
|
查看原文
表 3FCN-16分类方法的混淆矩阵与总体精度
Table3. Confusion matrice and overall accuracy of FCN-16 classification method%
Type ofground | B | F | W | R | S | G |
---|
B | 86.3 | 2.6 | 0 | 18.86 | 34.53 | 0.38 | F | 4.1 | 82.7 | 5.84 | 20.93 | 28.52 | 51.52 | W | 1.7 | 1.7 | 92.56 | 3.81 | 2.88 | 11.69 | R | 2.7 | 5.6 | 0.14 | 34.56 | 1.70 | 0.14 | S | 3.8 | 1.9 | 0.30 | 1.32 | 19.80 | 13.37 | G | 1.4 | 5.6 | 1.16 | 20.51 | 12.57 | 22.91 | OA | 71.6 |
|
查看原文
表 4FCN-8分类方法的混淆矩阵与总体精度
Table4. Confusion matrice and overall accuracy of FCN-8 classification method%
Type ofground | B | F | W | R | S | G |
---|
B | 83.54 | 3.68 | 0.14 | 21.24 | 35.77 | 8.49 | F | 8.13 | 89.59 | 8.93 | 38.10 | 28.27 | 36.16 | W | 1.91 | 0.57 | 89.92 | 2.16 | 2.90 | 2.18 | R | 2.46 | 3.25 | 0.46 | 22.96 | 5.54 | 2.40 | S | 1.06 | 0.19 | 0.01 | 1.67 | 9.77 | 1.61 | G | 2.90 | 2.72 | 0.54 | 13.87 | 17.75 | 49.16 | OA | 68.8 |
|
查看原文
表 5所提FCN分类方法的混淆矩阵与总体精度
Table5. Confusion matrice and overall accuracy of proposed FCN classification method%
Type ofground | B | F | W | R | S | G |
---|
B | 86.2 | 0.83 | 0 | 10.61 | 14.3 | 1.8 | F | 1.97 | 82.41 | 2.34 | 11.15 | 19.4 | 26.2 | W | 0.23 | 0.30 | 96.77 | 3.96 | 2.7 | 0 | R | 4.59 | 13.39 | 0.55 | 69.51 | 10.4 | 1.1 | S | 6.36 | 1.81 | 0.32 | 3.36 | 52.7 | 4.7 | G | 0.62 | 1.26 | 0.02 | 1.42 | 0.4 | 66.3 | OA | 80.90 |
|
查看原文
表 6加入均值漂移①分割结果的FCN分类方法的混淆矩阵与总体精度
Table6. Confusion matrice and overall accuracy of proposed FCN adding segmentation result of mean-shift ①%
Type ofground | B | F | W | R | S | G |
---|
B | 86.7 | 0.8 | 0 | 14.1 | 18.5 | 7.4 | F | 1.2 | 84.7 | 2.6 | 9.9 | 17.3 | 10.0 | W | 0 | 0.6 | 96.2 | 2.9 | 1.7 | 0 | R | 4.5 | 11.5 | 1.0 | 68.9 | 9.3 | 4.5 | S | 5.3 | 2.0 | 0.2 | 4.1 | 53.1 | 9.6 | G | 0.3 | 0.4 | 0 | 0.1 | 0.1 | 68.6 | OA | 82.1 |
|
查看原文
表 7加入均值漂移②分割结果的FCN分类方法的混淆矩阵与总体精度
Table7. Confusion matrice and overall accuracy of proposed FCN adding segmentation result of mean-shift ②%
Type ofground | B | F | W | R | S | G |
---|
B | 91.0 | 1.8 | 0 | 17.9 | 16.5 | 8.9 | F | 1.0 | 82.6 | 2.9 | 8.7 | 14.3 | 14.3 | W | 0 | 0.8 | 95.0 | 2.1 | 1.6 | 0 | R | 5.4 | 13.1 | 1.6 | 69.5 | 9.0 | 0.3 | S | 2.1 | 1.7 | 0.6 | 1.8 | 58.5 | 0 | G | 0.5 | 0.1 | 0 | 0 | 0.1 | 76.5 | OA | 83.5 |
|
查看原文
表 8加入均值漂移分割结果③的FCN分类方法的混淆矩阵与总体精度
Table8. Confusion matrice and overall accuracy of proposedFCN adding segmentation result of mean-shift ③%
Type ofground | B | F | W | R | S | G |
---|
B | 85.5 | 1.0 | 0 | 18.2 | 18.9 | 12.8 | F | 0.9 | 79.1 | 2.0 | 8.2 | 12.9 | 0 | W | 0 | 0.6 | 94.5 | 2.5 | 1.2 | 0 | R | 8.7 | 17.8 | 3.2 | 68.0 | 15.3 | 5.7 | S | 3.9 | 1.3 | 0.3 | 2.3 | 51.1 | 0 | G | 1.0 | 0.1 | 0 | 0.8 | 0.7 | 81.5 | OA | 79.4 |
|
查看原文
方旭, 王光辉, 杨化超, 刘慧杰, 闫立波. 结合均值漂移分割与全卷积神经网络的高分辨遥感影像分类[J]. 激光与光电子学进展, 2018, 55(2): 022802. Xu Fang, Guanghui Wang, Huachao Yang, Huijie Liu, Libo Yan. High Resolution Remote Sensing Image Classification Combining with Mean-Shift Segmentation and Fully Convolution Neural Network[J]. Laser & Optoelectronics Progress, 2018, 55(2): 022802.