摘要
针对传统的高分影像建筑物提取方法存在分割精度低和分割边界模糊等问题,提出基于U-net神经网络的高分辨率光学遥感影像建筑物提取方法。方法包含U-net神经网络和全连接CRFs网络,依据图像特征进行语义分割并优化分割结果;在实现U-net神经网络和全连接CRFs模型的基础上,根据U-net的特点以及本文数据特性调试出不同数据量的增强扭曲数据集进行测试,以达到最高鲁棒的分割效果;调试全连接CRFs模型,使得后处理结果更加贴近影像中的真实情况。实验结果表明,与利用传统分割方法的分割效果相比,该方法分割精度及地物边缘分割完整度都得到了显著提升,对高分辨率遥感影像中建筑物的实验分割精度达到了87.64%。
Aiming at the problems of low segmentation accuracy and fuzzy segmentation boundary of traditional high-resolution image extraction methods for buildings,this paper proposes a high-resolution optical remote sensing image extraction method based on U-net neural network.The method includes U-net neural network and fully connected CRFs network.On the basis of U-net neural network and fully connected CRFs model,according to the characteristics of U-net and the data characteristics of this paper,the enhanced distorted data sets of different data volumes are debutted for testing,so as to achieve the most robust segmentation effect.And debug the fully connected CRFs model,making the post-processing results more close to the real situation in the image.The experimental results show that compared with the traditional segmentation method,the segmentation accuracy and the segmentation integrity of the edge of ground objects are significantly improved,and the experimental segmentation accuracy of buildings in high-resolution remote sensing images reaches 87.64%.
作者
张浩然
赵江洪
张晓光
ZHANG Haoran;ZHAO Jianghong;ZHANG Xiaoguang(Institute of Mapping and City Space Information,Beijing University of Civil Engineering and Architecture,Beijing 102616,China;Key Laboratory of National Bureau of Surveying,Mapping and Geographic Information of Modern City,Beijing 102616,China;Beijing Key Laboratory of Fine Reconstruction and Health Monitoring of Architectural Heritage,Beijing 102616,China)
出处
《遥感信息》
CSCD
北大核心
2020年第3期143-150,共8页
Remote Sensing Information
基金
国家自然科学基金项目(41601409)
国家自然科学基金青年基金项目(41501495)
北京市自然科学基金项目(8172016)
城市空间信息工程北京市重点实验室开放研究基金项目(2018210)。
关键词
深度学习
高分辨率影像处理
图像语义分割
建筑物提取
全连接CRFs
deep learning
high resolution image processing
image semantic segmentation
building extraction
fully connected CRFs
