基于EfficientNet对地震受灾建筑物破坏评估  

Damaged buildings assessment based on EfficientNet network

作　　者：丁子林 姚新强[1] 李雅静[1] 张勇[2] DING Zilin;YAO Xinqiang;LI Yajing;ZHANG Yong(Tianjin Earthquake Agency,Tianjin 300201,China;Yunnan Earthquake Agency,Kunming 650041,China)

机构地区：[1]天津市地震局,天津300201 [2]云南省地震局,云南昆明650041

出　　处：《自然灾害学报》2023年第2期26-31,共6页Journal of Natural Disasters

基　　金：天津市重点研发计划科技支撑重点项目(20YFZCSN01010);天津市地震局重点项目(Zd202008)。

摘　　要：为了实现对建筑物工程震害的图像数据进行智能化分析,以汶川地震、玉树地震等震后受损建筑物图片数据为研究对象,划分了4个建筑物受损类型,建立了震害数据集。通过使用深度学习EfficientNet模型,对图片中建筑物的受损类型进行识别,并通过使用数据增强技术扩大数据容量,同时采用迁移学习技术提高识别准确率,实验得到的模型对震害建筑物图片进行评估识别,达到87.45%的准确度。最后,用训练后的模型对青海玛多7.4级地震的震害数据进行识别,达到了85.71%的准确率,制作了灾情分布图。研究表明,本技术可以为地震后的建筑物受损程度和相关的灾情,提供快速智能化识别,对实现地震后自动化、智能化地提供受灾建筑物的破坏信息有一定实用价值。In order to realize the intelligent analysis of image data of earthquake damage to building projects,four types of building damage were classified and a seismic damage dataset was established,which was based on the dataset of damaged building images after earthquake,such as Wenchuan earthquake and Yushu earthquake.A deep learning model(EfficientNet model)was used to identify and assess the building damage and collapse images.Data augmentation technology was used to expand the data capacity.We also improve the model accuracy by transfer learning and get the neural network model with 87.45%accuracy.At last,through using model to identify the images of Maduo M s7.4 earthquake in Maduo,Qinghai Province,the accuracy was 85.71%,and the seismic damage distribution map was obtained.The result reveals that this technology can provide rapid and intelligent identification of disaster situations related to the degree of damage to buildings after an earthquake.It is of practical value in achieving automated and intelligent provision of information on the damage to affected buildings after an earthquake.It also shows the potential uses of deep learning in structural damage recognition automatically and intelligently,and the value of providing disaster information.

关 键 词：建筑物震害图像 EfficientNet 深度学习 灾情获取 图像分类 

分 类 号：P315[天文地球—地震学] X43[天文地球—固体地球物理学]