摘要
遥感影像中农村道路矢量中心线的准确提取对乡村规划和地理信息数据库建设具有重要意义。针对现有深度学习方法未能充分利用上下文信息,且在下采样过程中易造成图像分辨率下降和局部特征丢失的问题,该文改进U-Net网络模型以提高提取结果的准确性。首先,网络结构设计两次下采样处理,并将上下文两处特征信息用跳跃层连接,使输出的道路细节清晰;其次,为避免样本不均衡导致训练效果不理想,采用交叉熵损失函数与广义骰子损失函数叠加的方式提升训练效果;最后,采用邻域质心投票算法和融合算法对提取的道路进行矢量化和中心线优化,得到高精度的农村道路矢量中心线。试验结果表明:改进方法在复杂场景的农村道路矢量中心线提取中准确率达95.03%,较4种对比算法(U-Net、DC-Net、PA-Net、SM-Net)具有明显优势。
Accurate extraction of rural road vector centerlines from remote sensing images is of great significance to rural planning and the construction of geographic information database.The existing image segmentation algorithms,such as U-Net network,can not effectively solve the influence of background information such as terrain shadows and trees on road extraction,so this paper improves the U-Net network to improve the accuracy of segmentation results.Firstly,the designed network structure is downsampled twice,and the two information features in the context are connected by a jump layer,so that the output road has clear detail expression ability.Secondly,in order to reduce the influence of poor training effect caused by unbalanced samples,this paper adopts the superposition of cross entropy loss function and generalized dice loss function to improve the training effect.Finally,in order to get the high-precision rural road vector centerline,the existing neighborhood centroid voting algorithm and fusion algorithm are used to vectorize and optimize the extracted road centerline.The experimental results show that the improved U-Net network combined with fusion algorithm is feasible in extracting the centerline of rural road vector,with an accuracy rate of 95.03%,which is obviously improved compared with the existing algorithms.The algorithm proposed in this paper provides a new idea and method for high-precision extraction of rural road vector centerline from remote sensing images.
作者
王怡君
李旺平
柴成富
尉文博
邓灵芝
WANG Yijun;LI Wangping;CHAI Chengfu;WEI Wenbo;DENG Lingzhi(Gansu Provincial Map Institute,Lanzhou 730099;School of Civil Engineering,Lanzhou University of Technology,Lanzhou 730050;Gansu Emergency Surveying and Mapping Engineering Research Center,Lanzhou 730050;Aerial Photogrammetry and Remote Sensing Group Co.,Ltd.(ARSC),Xi′an 710199;Enterprise Technology Center of Aerial Photogrammetry and Remote Sensing Group Co.,Ltd,Xi′an 710199,China)
出处
《地理与地理信息科学》
CSCD
北大核心
2024年第4期34-39,共6页
Geography and Geo-Information Science
基金
甘肃省自然科学基金项目(22JR5RA247)
甘肃省青年科技基金项目(23JRRA830)。
关键词
改进U-Net网络
遥感影像
网络分割
农村道路提取
矢量线优化
improved U-Net network
remote sensing image
network segmentation
rural road extraction
vector line optimization
