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

Road extraction method of high-resolution remote sensing image on the basis of the spatial information perception semantic segmentation model

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。

With the rapid development of remote sensing satellite technology,the automatic extraction of high-resolution remote sensing images has become a popular research direction in the field of remote sensing.Deep learning methods have been applied in remote sensing image road information extraction and achieved significant results.However,due to convolution and pooling and other operations in network,road extraction methods based on the deep learning have some problems,such as the loss of spatial features and ground object details and the frequent occurrence of false extraction during road extraction.In order to solve these problems,this paper designs an improved road extraction semantic segmentation network model to mitigate the impact of the above network structure.The proposed method is based on ResNet and introduces coordinate convolution and a global information enhancement module before and after the coding structure,respectively.First,the network structure is mainly composed of residual units of ResNet,which has powerful feature extraction and multiplexing capabilities,and extracts road features of different scales and levels.Second,coordinate convolution reduces the spatial information loss and enhances the edge information.The coordinate convolution before the coding structure introduces spatial coordinate information,which is beneficial to enhancing the extraction of effective spatial information.Finally,global pooling can improve global context awareness.The global information enhancement module after the coding structure can effectively extract global context information through global pooling,thereby improving the accuracy of road classification and reducing the influence of natural scene factors,such as houses and tree shadows,to a certain extent.In this paper,the Massachusetts Roads dataset was used in the experiment,and the results obtained exhibited good accuracy.The Recall rate was 71.02%,the comprehensive evaluation index(F1 Score) was 76.35%,and the IoU reached 62.18%.The F1 Score and IoU indicators of the proposed method are approximately 1% higher than those in U-net and D-LinkNet and exceed those of DeeplabV3+ and Segnet,which are lower than D-LinkNet in the recall index only.The comparison of the experimental results indicates that the proposed method can effectively alleviate the spatial feature and context information losses on the basis of the deep learning road extraction method and completely extract the roads in remote sensing images.Moreover,the proposed method can effectively extract the road in the case of trees and building shadows,and multi-scale roads can also be accurately extracted.