• HomeAboutInstructionsSubscribeEditoral BoardJournals cooperationQuery Retrieve中文

View Full Text   View/Add Comment   Download reader  

  2019,Vol.14no.4 ,page:810-820

The study on extraction of seismic damage of buildings from remote sensing image based on fully convolutional neural network
Abstract:      In order to solve the problem that the automation degree of extracting damaged Buildings caused by earthquake is not very high, in this paper a fully convolutional neural network is applied to extract the remote sensing information of earthquake damage to buildings. The 0.2m-resolution aerial image of the Yushu County urban area obtained after the Yushu earthquake was used as the data source to test the result of convolutional neural network. The objects in the test area were classified into collapsed buildings, uncollapsed buildings, and background. Classify and label 427 sub-images of 500×500 pixels manually, 393 of them were selected as training sample set, and others as verification sample set. The training sample set is used to train the full convolutional neural network and the trained network is used to extract the building seismic damage information and evaluate the accuracy based on the verification sample. The result shows that the overall classification accuracy is 82.3%, and the fully convolutional neural network can improve the automation of information extraction and has a better ability to extract building seismic damage information.
Author NameAffiliation
Chen Meng Institute of Earthquake Forecasting, CEA, Beijing 100036, China 
Wang Xiaoqing Institute of Earthquake Forecasting, CEA, Beijing 100036, China 
Keywords:    Deep learning  Fully convolutional neural network  Buildings  Seismic damage information  Remote sensing
Received on:  January 28, 2019
全卷积神经网络在建筑物震害遥感提取中的应用研究
摘要:    为解决建筑物震害信息提取自动化程度不高的问题,本文将全卷积神经网络应用于建筑物震害遥感信息提取。以玉树地震后获取的玉树县城区0.2m分辨率航空影像作为建筑物震害信息提取试验数据源,将试验区地物划分为倒塌建筑物、未倒塌建筑物和背景3类。对427个500×500像素的子影像进行人工分类与标注,选取393个组成训练样本集,34个用于验证。利用训练样本集对全卷积神经网络进行训练,采用训练后的网络对验证样本进行建筑物震害信息提取及精度评价。研究结果表明:建筑物震害遥感信息提取总体分类精度为82.3%,全卷积神经网络方法能提高信息提取自动化程度,具有较好的建筑物震害信息提取能力。
关  键  词:深度学习  全卷积神经网络  建筑物  震害信息  遥感
基金项目:科技部重点研发课题(2017YFB0504104)
作者简介:陈梦,男,生于1992年。硕士研究生。研究方向为机器学习,深度学习,遥感图像处理。E-mail:hpu_cm@163.com
引用本文格式:陈梦,王晓青.全卷积神经网络在建筑物震害遥感提取中的应用研究[J].震灾防御技术英文版,2019,14(4):810-820  DOI:10.11899/zzfy20190412
Close