欢迎访问《地震》,

地震 ›› 2022, Vol. 42 ›› Issue (3): 124-140.doi: 10.12196/j.issn.1000-3274.2022.03.009

• • 上一篇    下一篇

基于Deeplab V3+的2019年长宁M6.0地震建筑物震害信息提取研究

王鑫1, 窦爱霞1, 郭红梅2, 袁小祥1   

  1. 1.中国地震局地震预测研究所, 北京 100036;
    2.四川省地震局, 四川 成都 610041
  • 收稿日期:2021-07-27 修回日期:2022-01-07 出版日期:2022-07-31 发布日期:2023-03-29
  • 通讯作者: 窦爱霞, 正研级高工。 E-mail: axdothy@163.com
  • 作者简介:王鑫(1994-), 男, 河南永城人, 硕士研究生, 主要从事遥感与GIS应用研究。
  • 基金资助:
    中国地震局地震预测研究所基本科研业务费专项重点项目(2018IEF010106)

Research on Extraction of Building Damage Information in the 2019 Changning M6.0 Earthquake Based on Deeplab V3+

WANG Xin1, DOU Ai-xia1, GUO Hong-mei2, YUAN Xiao-xiang1   

  1. 1. Key Laboratory of Earthquake Prediction, Institute of Earthquake Forecasting, CEA, Beijing 100036, China;
    2. Sichuan Earthquake Agency, Chengdu 610041, China
  • Received:2021-07-27 Revised:2022-01-07 Online:2022-07-31 Published:2023-03-29

摘要: 快速获取地震所造成的建筑物震害信息是灾后救援及灾害损失评估的关键, 以语义分割网络为代表的深度学习技术的发展, 为震后建筑物震害信息的获取提供了一种新的手段。 本文将Deeplab V3+神经网络模型应用于建筑物震害信息的提取, 以2019年长宁M6.0地震震后双河镇0.06 m分辨率的无人机影像作为建筑物震害信息提取试验数据源, 将试验区建筑物划分为无明显破坏、 破坏两类, 其余作为背景信息, 人工进行建筑物震害信息的分类与标注, 构建样本数据集并进行网络模型的训练。 为了提高模型针对不同场景下地物类别分类的鲁棒性与泛化能力, 进行了样本增强处理, 开展了样本增强对模型精度的影响实验与训练样本数量对模型精度的影响实验, 并以九寨沟漳扎镇震后无人机0.4 m分辨率影像进行模型的可迁移性实验。 实验结果表明, 本次研究所建立的神经网络模型在建筑物震害信息的提取中可以达到较高的提取精度; 样本增强后较样本增强前模型各类别提取精度都有不同程度的提高, 特别是破坏建筑物召回率提高约0.15, 精确率提高约0.20; 该神经网络模型在未做任何修改的情况下, 具有一定的可迁移性。

关键词: 深度学习, Deeplab V3+, 建筑物, 震害提取, 样本增强

Abstract: Rapid acquisition of building damage information caused by earthquakes is the key to post-disaster rescue and disaster loss assessment. The development of deep learning technology represented by semantic segmentation network provides a new means for the acquisition of building damage information after earthquakes. In this paper, Deeplab V3 + neural network model is applied to the extraction of earthquake damage information of buildings. The UAV image with 0.06 meters resolution in Shuanghe Town after Chang Ning earthquake is used as the data source for the extraction of earthquake damage information of buildings. The buildings in the test area are divided into two categories: no obvious damage and damage, and the rest are used as background information. The seismic damage information of buildings is classified and annotated manually, and the sample data set is constructed and the network model is trained. In order to improve the robustness and generalization ability of the model for the classification of ground objects in different scenarios, the sample enhancement process was carried out, and the experiment of the effect of sample enhancement on the model accuracy and the effect of the number of training samples on the model accuracy were carried out. After the earthquake, the 0.4-meter-resolution UAV image was used to conduct the model transferability experiment. The experimental results show that the neural network model established in this study can achieve higher extraction accuracy in the extraction of building earthquake damage information; the extraction accuracy of each category of the model after sample enhancement is improved to varying degrees, especially it is that the recall rate of the destroyed building is increased by about 0.15, and the precision rate is increased by about 0.20; the neural network model has a certain transferability without any modification.

Key words: Deep learning, Deeplab V3+, Building, Earthquake damage extraction, Sample enhancement

中图分类号: