面向滑坡危险性评价的深度一维卷积神经网络方法--以四川省芦山县为例  

A deep one-dimensional convolutional neural network method for landslide risk assessment:A case study in Lushan,Sichuan,China

作　　者：张洪吉[1] 赵铮 陈建华[2] 甘先霞 谢华伟 谭小琴[1] ZHANG Hongji;ZHAO Zheng;CHEN Jianhua;GAN Xianxia;XIE Huawei;TAN Xiaoqin(Institute of Natural Resource Sciences of Sichuan Province,Chengdu 610015,China;College of Geophysics,Chengdu University of Technology,Chengdu 610059,China)

机构地区：[1]四川省自然资源科学研究院,四川成都610015 [2]成都理工大学地球物理学院,四川成都610059

出　　处：《自然灾害学报》2021年第3期191-198,共8页Journal of Natural Disasters

基　　金：四川省科技计划项目(2019YFG0187)。

摘　　要：滑坡灾害威胁着人类生命财产安全。因此,准确地评价滑坡危险性对于灾害分析至关重要。目前常见的机器学习方法不能提取更深层次的数据特征,难以获得高质量的滑坡危险性评价图。因此,论文提出一种面向滑坡危险性评价的深度一维卷积神经网络方法,以期利用卷积神经网络强大的表征学习能力获得更高精度的评价结果。以滑坡灾害多发的四川省芦山县为例,选择2016年346个历史滑坡数据用于模型的训练和验证。为了验证模型的有效性和精度,实验将支持向量机模型、信息量模型与一维卷积神经网络模型进行对比分析,并将其应用于芦山县全区滑坡危险性评价。结果上,一维卷积神经网络和支持向量机的总体精度分别为0.9015和0.842,而一维卷积神经网络、支持向量机、信息量模型的AUC值分别为0.888、0.876、0.818。结果表明:一维卷积神经网络模型精度较高,在数据拟合度和评价效果上优于支持向量机模型、信息量模型,可以有效地应用于滑坡危险性评价。Landslide disasters threaten the safety of human life and properties.It is important to accurately assess the risk of regional landslides.At present,common machine learning methods cannot extract deeper data features and obtain higher precision landslide hazard assessment maps.Therefore,this paper proposes a deep one-dimensional convolutional neural network method for landslide risk assessment,in order to extract deeper features and obtain higher precision evaluation results by based on its powerful learning ability.Taking Lushan,Sichuan,China as an example,346 historical landslide pionts of Lushan in 2016 were selected for model training and verification.In order to verify the validity and precision of the model,the support vector machine,the information value method and the proposed convolutional neural network model were compared in the experiments.The overall accuracies of the convolutional neural network and the support vector machine are 0.9015 and 0.842.The AUC values of the one-dimensional convolutional neural network,support vector machine,and information value model are 0.888,0.876,and 0.818,respectively,The results show that the proposed convolutional neural network model has a higher accuracy and is superior to the support vector machine and the information value method in data fitting ability and evaluation effect,and can be effectively applied to the regional landslide risk evaluation.

关 键 词：滑坡灾害 卷积神经网络 支持向量机 滑坡危险性评价 机器学习 

分 类 号：TU4[建筑科学—土工工程] X4[环境科学与工程—灾害防治]