0 引言
1 农村公路遥感影像特征
2 BFG-U-Net架构
2.1 U-Net
2.2 BFG-U-Net架构
2.3 二值化阈值滤波
3 实验与验证
3.1 实验方案
3.2 实验环境与条件
3.3 算法性能评价指标
3.4 实验结果及性能比较
表1 算法相对交并比 |
| 路网提取算法 | BFG-U-Net | U-Net | Res-U-Net |
|---|---|---|---|
| 相对交并比(%) | 100 | 97.74 | 55.11 |
Convolutional Neural Network Architecture for Road Extraction from High-Resolution Remote Sensing Images of Rural Road
Received date: 2021-02-08
Online published: 2021-11-23
Copyright
In order to improve the accuracy of high-resolution remote sensing images of road networks in rural areas, an architecture named BFG(Block of Focusing Globe) is proposed, which is a kind of skip connection of convolutional neural network on high-resolution remote sensing images. BFG architecture gives a convoluting operation on feature maps of down sampling during the time of skip connection, and links to the size of convolution kernel and the stages of down sampling, finally concatenates the original feature maps to output the link of up sampling. Thus the skip connection can offer more detailed global and contextual information for original feature maps. Some comparative experiments are carried on to evaluate the performance of the new architecture. The results show that, in classic U-Net, the intersection over union (IoU) without BFG is 97.74% of BFG equipped, and the IoU of Res-U-Net is 55.11% of BFG-U-Net; in high-resolution remote sensing images, the interference of road extraction can be reduced effectively while the road is passing through various types of roadside terrain, and the extraction results are clearly visible and continuous; the roads can also be identified correctly and extracted even covered by cloud shadows or trees. It proves that BFG architecture has good applicability.
MA Xiao , ZHANG Xiao-zheng , FAN Wen-tao , LIU Liu-yang , SHAN Fei , SUN Zhuo . Convolutional Neural Network Architecture for Road Extraction from High-Resolution Remote Sensing Images of Rural Road[J]. Transport Research, 2021 , 7(5) : 91 -98 . DOI: 10.16503/j.cnki.2095-9931.2021.05.011
表1 算法相对交并比 |
| 路网提取算法 | BFG-U-Net | U-Net | Res-U-Net |
|---|---|---|---|
| 相对交并比(%) | 100 | 97.74 | 55.11 |
| [1] |
中共中央,国务院. 交通强国建设纲要[Z]. 北京:中共中央,国务院, 2019.
|
| [2] |
袁鹏飞, 黄荣刚, 胡平波, 等. 基于多光谱LiDAR数据的道路中心线提取[J]. 地球信息科学学报, 2018,20(4):452-461.
|
| [3] |
查中亮. 基于多源遥感数据的道路、居民点提取及布局优化研究[D]. 成都:四川师范大学, 2018.
|
| [4] |
李杰, 冯魁祥, 朱玲玲, 等. 基于阈值标记的分水岭算法遥感图像道路提取[J]. 长春大学学报, 2019,232(6):10-14.
|
| [5] |
符喜优, 张风丽, 王国军, 等. 基于模糊连接度的高分辨率SAR图像道路自动提取[J]. 计算机应用, 2015,35(2):523-527.
|
| [6] |
|
| [7] |
|
| [8] |
连仁包, 王卫星, 李娟. 自适应圆形模板及显著图的高分辨遥感图像道路提取[J]. 测绘学报, 2018,47(7):950-958.
|
| [9] |
|
| [10] |
|
| [11] |
|
| [12] |
|
| [13] |
|
| [14] |
|
| [15] |
|
| [16] |
|
| [17] |
|
| [18] |
交通运输部. 交通运输部办公厅关于推进交通运输行业数据资源开放共享的实施意见(交办科技〔2016〕113号)[Z]. 北京:交通运输部, 2016.
|
/
| 〈 |
|
〉 |