摘要
城市建筑物阴影的存在增加了基于高分影像信息提取的复杂度,影响了高分影像的行业应用。以GF-2影像为数据源,以影像覆盖的两个典型阴影区域为实验区,设计了一种基于深度学习技术的城市建筑物阴影提取方法。借助于全卷积神经网络(FCN)对图像进行语义分割,获取阴影信息,通过建立光照方向相邻对象的空间关系,去除建筑物阴影信息中的干扰项。将提取的建筑物阴影结果与面向对象方法提取的结果进行对比,在实验区1和实验区2中,DBSE提取的总体精度分别为97.5%和98.1%,面向对象的总体精度分别为96.2%和97.2%。提升精度的同时,在视觉效果上,DBSE方法在阴影完整性与真实阴影一致性,以及减少阴影错分等方面明显优于OBSE方法。
The existence of urban building shadows increases the difficulties of information extraction for high-resolution images and thus limits its industrial applications.This article proposes a new approach for urban building shadow extraction based on deep learning technology.The full convolutional neural network(FCN)is used for semantic segmentation of images and extraction of shadow information.The spatial relationships between any two adjacent objects in the direction of illumination are established,and the interference to the shadow information of buildings are removed.We apply the approach to two typical shadow areas of GF-2 images.Results indicate that the overall accuracies of the object-oriented method for experimental area 1 and 2 are 97.5%and 98.1%respectively,and that of DBSE are 96.2%and 97.2%respectively.Our approach works better than OBSE as it improves extraction accuracy and visualizes better,especially for shadow integrity and complex terrain.
作者
任慧群
蔡国印
李志强
REN Huiqun;CAI Guoyin;LI Zhiqiang(Department of Survey and Urban Spatial Information,Beijing University of Civil Engineering and Architecture,Beijing 102616,China;Key Laboratory for Urban Geomatics of National Administration of Surveying,Mapping and Geoinformation,Beijing 102616,China;Twenty First Century Aerospace Technology Co.,Ltd.,Beijing 100096,China)
出处
《地理信息世界》
2020年第2期81-86,共6页
Geomatics World
基金
高分辨率对地观测系统重大专项应用共性关键技术预研项目(06-Y20A17-9001-17/18)
北京建筑大学市属高校基本科研业务费专项资金(NO.2018N060301)资助。
关键词
高分二号影像
阴影提取
全卷积神经网络
语义分割
GF-2
shadow extraction
full convolutional neural network
semantic segmentation
