首页 >  , Vol. , Issue () : -

摘要

全文摘要次数: 267 全文下载次数: 376
引用本文:

DOI:

10.11834/jrs.20210021

收稿日期:

2020-01-20

修改日期:

2020-09-03

PDF Free   EndNote   BibTeX
基于空间信息感知语义分割模型的高分辨率遥感影像道路提取方法
吴强强1, 王帅2, 王彪1, 吴艳兰3,1,4
1.安徽大学资源与环境工程学院;2.武汉市测绘研究院,湖北,武汉,430022;3.信息材料与智能感知安徽省实验室;4.安徽省地理信息智能技术工程研究中心,安徽 合肥 230000
摘要:

道路信息自动化提取已经成为遥感领域热门的研究方向,而基于深度学习的遥感影像道路信息提取方法已经取得了许多成果。但由于受到网络中卷积和池化等操作的影响,基于深度学习的道路提取方法存在着空间特征和地物细节信息丢失等问题,造成许多误提现象。针对此问题,本文设计了一种改进的道路提取语义分割网络模型,该网络以改进的ResNet网络为主体,并引入坐标卷积和全局信息增强模块,用于增强空间信息和全局上下文信息的感知能力,突出道路边缘特征进而确保道路分类的精确性。本文方法在公开道路数据集和高分数据集上获得了显著的提取效果,与其它方法相比取得了明显提高。并且,在一定程度上减少了树木、建筑阴影等自然场景因素遮挡的影响,可以完整准确地提取出道路。此外,模型对多尺度道路也可以实现有效地提取。

An road extraction method of high-resolution remote sensing image based on spatial information perception semantic segmentation model
Abstract:

With the rapid development of remote sensing satellite technology, the automatic extraction of high-resolution remote sensing images has become a hot research direction in the field of remote sensing.Deep learning method has been applied in remote sensing image road information extraction and achieved certain results. However, due to the network convolution and pooling and other operations, road extraction method based on the deep learning has some problems such as the loss of spatial features and ground object details, and a lot of false extraction occurs in the process of road extraction. To solve this problem, this paper designs an improved road extraction semantic segmentation network model to mitigate the impact of the above network structure. The paper method is based on ResNet network, and introduces coordinate convolution and global information enhancement module before and after the coding structure. First of all, the network structure is mainly composed of residual units of ResNet, which has powerful feature extraction and feature multiplexing capabilities, and extracts road features of different scales and levels. Secondly, coordinate convolution has the characteristics of reducing spatial information loss and enhancing edge information. The coordinate convolution before the coding structure introduces spatial coordinate information, which is beneficial to enhance the extraction of effective spatial information.Finally, global pooling has the ability to improve global context awareness. The global information enhancement module after coding structure can effectively extract global context information through global pooling, so as to improve the accuracy of road classification and reduce the influence of natural scene factors such as houses and tree shadows to a certain extent. In this paper, Massachusetts Roads dataset was used to conduct the experiment, and the results obtained good accuracy. Among them, the Recall rate was 71.02%, the comprehensive evaluation index (F1 score) was 76.35%, and the IoU reached 62.18%. Compared with other methods, the F1 Score and IoU indicators are about 1% higher than U-net and D-LinkNet, and far exceed the DeeplabV3+ and Segnet, which is lower than the D-LinkNet only in the Recall index. By comparing the experimental results, the paper method can effectively alleviate the spatial feature loss and context information loss based on the deep learning road extraction method, and can completely extract the roads in the remote sensing image. Moreover, the method in this paper can also effectively extract the road in the case of trees and building shadows. In addition, multi-scale road can also be accurately extracted.

本文暂时没有被引用!

欢迎关注学报微信

遥感学报交流群 分享按钮