基于卷积神经网络的福建省区域滑坡灾害预警模型

董力豪; 刘艳辉; 黄俊宝; 刘海宁

doi:10.16030/j.cnki.issn.1000-3665.202211018

基于卷积神经网络的福建省区域滑坡灾害预警模型

doi: 10.16030/j.cnki.issn.1000-3665.202211018

董力豪^{1, 2,},
刘艳辉^1, ,,
黄俊宝³,
刘海宁²

1.
中国地质环境监测院（自然资源部地质灾害技术指导中心），北京　100081，中国
2.
华北水利水电大学，郑州　450045，中国
3.
福建省地质环境监测中心，福州　350002，中国

基金项目: 国家自然科学基金（42077440；41202217）；国家重点研发计划（2018YFC1505503）

详细信息

作者简介:
董力豪（1998—），男，硕士，硕士研究生，主要从事地质灾害预警相关研究工作。E-mail：1365234358@qq.com

通讯作者:
刘艳辉（1978—），女，博士，正高级工程师，主要从事地质灾害预警与防治、工程地质等方面的研究工作。E-mail：392990563@qq.com

中图分类号: 中图分类号: 文献标志码: 文章编号:
计量
- 文章访问数: 6682
- HTML全文浏览量: 2110
- PDF下载量: 178
- 被引次数: 0
出版历程
- 收稿日期: 2022-11-07
- 录用日期: 2023-03-13
- 修回日期: 2023-03-06

An early prediction model of regional landslide disasters in Fujian Province based on convolutional neural network

DONG Lihao^{1, 2
,},
LIU Yanhui^{1
, ,},
HUANG Junbao³,
LIU Haining²

1.
China Institute of Geological Environment Monitoring （Technical Guidance Center for Geological Hazards, Ministry of Natural Resources）, Beijing　100081
2.
North China University of Water Resources and Electric Power, Zhengzhou, Henan　450045, China
3.
Fujian Geological Environment Monitoring Center, Fuzhou, Fujian　350002, China

摘要

摘要: 福建省滑坡灾害频发，开展区域尺度上的滑坡灾害预警是有效防灾减灾的重要手段，由于滑坡成灾机理复杂，传统的区域滑坡预警方法存在精度不足等问题。深度学习主要是指通过构建神经网络模型进行特征的提取、抽象、表示与学习的技术，是机器学习的一种。卷积神经网络作为一种经典的深度学习算法，具有比传统机器学习更强大的分类能力与表征能力。文章以福建省为研究区，将卷积神经网络引入滑坡灾害预警领域，构建福建省区域滑坡预警模型，过程如下：①采用SMOTE优化算法对2010—2018年福建省滑坡灾害样本库进行优化，扩充正样本的个数，将正负样本比例从1∶3.4扩充到1∶2，样本总量达到19151个。②构建卷积神经网络模型结构，模型结构包括一个输入层、两个卷积层、两个最大池化层和一个全连接层以及一个输出层。③使用卷积神经网络对优化后的样本（2010—2018年样本的80%作为训练集）进行训练，并用贝叶斯优化算法优化模型超参数，得到福建省区域滑坡预警模型。④以2010—2018年样本的20%作为测试集对模型进行校验，采用混淆矩阵、ROC曲线进行模型校验，结果显示模型准确度为0.94−0.97，AUC值达到0.975，模型精度与泛化能力良好。⑤以2019年汛期滑坡灾害实况作为正样本，通过时空采样的方法采集负样本，构建2019年区域滑坡样本校验集（样本数603个），对模型进行进一步实况校验，采用混淆矩阵、ROC曲线进行模型校验，结果显示模型准确度为0.75−0.80，AUC值为0.852。虽然仅用了2019年一年汛期的滑坡实况样本进行校验，但也达到较好的效果。本文将卷积神经网络算法应用到区域滑坡预警中，为建立区域滑坡预警模型提供了一种新的途径，初步校验表明，模型效果良好，今后将在福建省对模型进行进一步的应用与校验。
- 滑坡灾害 /
- 预警模型 /
- 深度学习 /
- 卷积神经网络 /
- 模型构建
Abstract: Landslide disasters occur frequently in Fujian Province, and early warning of landslide disasters on a regional scale is an important means of effective disaster prevention and mitigation. Due to the complex mechanism of landslide disasters, the traditional regional landslide early warning methods have such problems as insufficient accuracy. Deep learning mainly refers to the technology of feature extraction, abstraction, representation and learning by constructing the neural network model, which is a kind of machine learning. As a classical deep learning algorithm, convolutional neural network has more powerful classification and representation ability than traditional machine learning. Taking Fujian Province as the research area, this paper introduces the convolution neural network into the field of landslide disaster early warning and constructs a regional landslide early warning model of Fujian Province. The process is as follows: ① the SMOTE optimization algorithm is used to optimize the sample database of landslide disasters in Fujian Province from 2010 to 2018, enlarging the number of positive samples and expanding the proportion of positive and negative samples from 1∶3.4 to 1∶2, and the total number of samples reaches 19151. ② Construct a convolution neural network model structure, which includes an input layer, two convolution layers, two maximum pooling layers, a full connection layer and an output layer. ③ Use the convolution neural network to train the optimized samples (80% of the samples from 2010 to 2018 as the training set), and use the Bayesian optimization algorithm to optimize the model parameters to obtain the regional landslide early warning model of Fujian Province. ④ The model is verified with 20% of the samples from 2010 to 2018 as the test set, and the confusion matrix and ROC curve are used to verify the model. The results show that the accuracy of the model ranges from 0.94 to 0.97, the AUC value is 0.975, indicating that the model accuracy and generalization ability are good. ⑤ The actual situation of the landslide disaster in the flood season of 2019 is taken as a positive sample, negative samples are collected through the method of time-space sampling, and the 2019 regional landslide sample verification set (603 samples) is constructed. The model is further verified by using the confusion matrix and ROC curve. The results show that the accuracy of the model ranges from 0.75 to 0.80, and the AUC value is 0.852. Although only the actual landslide samples in the flood season of 2019 is used for verification, good results is also achieved. In this paper, the convolution neural network algorithm is applied to the regional landslide early warning, which provides a new way to establish the regional landslide early warning model. The preliminary verification shows that the model is effective and will be further applied and verified in Fujian Province in the future.
- landslide disaster /
- early warning model /
- deep learning /
- convolutional neural network /
- model building

HTML全文

图 1 SMOTE算法合成数据示意图

Figure 1. SMOTE algorithm synthesis data diagram

下载: 全尺寸图片幻灯片

图 2 福建省训练样本集分布

Figure 2. Training sample set of Fujian Province

下载: 全尺寸图片幻灯片

图 3 典型CNN网络结构图

Figure 3. Typical network structure of CNN

下载: 全尺寸图片幻灯片

图 4 训练过程中准确率的变化

Figure 4. Changes in accuracy during training

下载: 全尺寸图片幻灯片

图 5 CNN模型ROC曲线

Figure 5. ROC curves of the CNN model

下载: 全尺寸图片幻灯片

图 6 福建省2019年汛期滑坡灾害分布图

Figure 6. Distribution map of landslide hazards in 2019 flood season in Fujian Province

下载: 全尺寸图片幻灯片

图 7 2019年样本CNN模型校验结果的ROC曲线

Figure 7. ROC curves of the 2019 sample CNN model verification results

下载: 全尺寸图片幻灯片

表 1 本文CNN模型超参数设置

Table 1. Hyperparameter settings of the CNN models

超参数	意义	优化后参数
units1 units2 dropout rate activation1 activation2 activation3 lr	第一层卷积核的大小第二层卷积核的大小每层神经元丢弃率第一层激活函数第二层激活函数全连接层激活函数学习率	512$ \times 1 $ 32$ \times 1 $ 0.1 relu relu elu 0.002

下载: 导出CSV

表 2 不同阈值下的 CNN分类结果混淆矩阵

Table 2. Confuse matrix of the results of the CNN classification under different thresholds

阈值			实际值
阈值			非滑坡	滑坡
	预测值	非滑坡	2325	72	特异度：0.963
0.25	预测值	滑坡	73	1138	灵敏度：0.939
		召回率：0.969		精确率：0.941	总精度：0.959
	预测值	非滑坡	2344	53	特异度：0.978
0.5	预测值	滑坡	78	1133	灵敏度：0.935
		真反例率：0.967		精确率：0.955	准确率：0.964
	预测值	非滑坡	2353	44	特异度：0.982
0.75	预测值	滑坡	86	1125	灵敏度：0.929
		真反例率：0.964		精确率：0.962	总精度：0.964

下载: 导出CSV

表 3 2019年样本CNN模型校验结果的混淆矩阵

Table 3. Confusion matrix of the 2019 sample CNN model verification results

阈值			实际值
阈值			非滑坡	滑坡
	预测值	非滑坡	434	76	特异度：0.851
0.25	预测值	滑坡	18	75	灵敏度：0.806
		召回率：0.960		精确率：0.497	总精度：0.844
	预测值	非滑坡	450	120	特异度：0.789
0.5	预测值	滑坡	2	31	灵敏度：0.939
		真反例率：0.996		精确率：0.205	准确率：0.798
	预测值	非滑坡	452	151	特异度：0.750
0.75	预测值	滑坡	0	0	灵敏度：0
		真反例率：1		精确率：0	总精度：0.750

下载: 导出CSV

参考文献(36)

[1]	KAVZOGLU T,SAHIN E K,COLKESEN I. Landslide susceptibility mapping using GIS-based multi-criteria decision analysis,support vector machines,and logistic regression[J]. Landslides,2014,11(3):425 − 439. doi: 10.1007/s10346-013-0391-7
[2]	HUANG Lu,XIANG Luyang. Method for meteorological early warning of precipitation-induced landslides based on deep neural network[J]. Neural Processing Letters,2018,48(2):1243 − 1260. doi: 10.1007/s11063-017-9778-0
[3]	何永金. 福建省主要地质灾害的特点、成因及其对策[J]. 福建地质,1995,14(4):263 − 271. [HE Yongjin. Characteristics and mechanism of major geological hazards in Fujian Province and protection and controlling method against them[J]. Geology of Fujian,1995,14(4):263 − 271. (in Chinese with English abstract)
[4]	CANNON S H, ELLEN S D. Rainfall conditions for abundant debris avalanches, San Francisco Bay region, California[J]. California Geology. 1985, 38（12）: 267–272
[5]	兰恒星,周成虎,王苓涓,等. 地理信息系统支持下的滑坡-水文耦合模型研究[J]. 岩石力学与工程学报,2003,22(8):1309 − 1314. [LAN Hengxing,ZHOU Chenghu,WANG Lingjuan,et al. GIS based landslide stability and hydrological distribution coupled model[J]. Chinese Journal of Rock Mechanics and Engineering,2003,22(8):1309 − 1314. (in Chinese with English abstract) doi: 10.3321/j.issn:1000-6915.2003.08.015
[6]	史中发. 哀牢山地区典型降雨型滑坡稳定性研究[D]. 北京: 中国地质大学（北京）, 2014 SHI Zhongfa. Stability analysis of a rainfall-induced landslide in the area of ailao mountain[D]. Beijing: China University of Geosciences （Beijing）, 2014. （in Chinese with English abstract）
[7]	李媛. 区域降雨型滑坡预报预警方法研究[D]. 北京: 中国地质大学（北京）, 2005 LI Yuan. Method for the warning of precipitation-induced landslides[D]. Beijing: China University of Geosciences （Beijing）, 2005. （in Chinese with English abstract）
[8]	刘传正,李铁锋,程凌鹏,等. 区域地质灾害评价预警的递进分析理论与方法[J]. 水文地质工程地质,2004,31(4):1 − 8. [LIU Chuanzheng,LI Tiefeng,CHENG Lingpeng,et al. A method by to analyses four parameters for assessment and early warning on the regional geo-hazards[J]. Hydrogeology & Engineering Geology,2004,31(4):1 − 8. (in Chinese with English abstract) doi: 10.3969/j.issn.1000-3665.2004.04.001
[9]	CAINE N. The rainfall intensity - duration control of shallow landslides and debris flows[J]. Geografiska Annaler:Series A,Physical Geography,1980,62(1/2):23 − 27.
[10]	BRAND E W, PREMCHITT J, PHILLIPSON H B. Relationship between rainfall and landslides in Hong Kong[C]//Proceedings of the 4th International Symposium on Landslides. Toronto: Canadian Geotechnical Society, 1984, 1（1）: 276 − 284.
[11]	HONG Yong,HIURA H,SHINO K,et al. The influence of intense rainfall on the activity of large-scale crystalline schist landslides in Shikoku Island,Japan[J]. Landslides,2005,2(2):97 − 105. doi: 10.1007/s10346-004-0043-z
[12]	刘传正,刘艳辉,温铭生,等. 中国地质灾害气象预警实践:2003—2012[J]. 中国地质灾害与防治学报,2015,26(1):1 − 8. [LIU Chuanzheng,LIU Yanhui,WEN Mingsheng,et al. Early warning for regional geo-hazards during 2003-2012,China[J]. The Chinese Journal of Geological Hazard and Control,2015,26(1):1 − 8. (in Chinese with English abstract)
[13]	PENNINGTON C, DASHWOOD C, FREEBOROUGH K. The National Landslide Database and GIS for Great Britain: construction, development, data acquisition, application and communication[C]//EGU General Asse-mbly Conference Abstracts. 2014: 3638.
[14]	陈香,王俪儒. 福建省滑坡灾害气象预警的研究[J]. 防灾科技学院学报,2015,17(4):68 − 75. [CHEN Xiang,WANG Liru. A study on landslide hazard meteorological early warning in Fujian Province[J]. Journal of Institute of Disaster Prevention,2015,17(4):68 − 75. (in Chinese with English abstract) doi: 10.3969/j.issn.1673-8047.2015.04.011
[15]	方然可,刘艳辉,苏永超,等. 基于逻辑回归的四川青川县区域滑坡灾害预警模型[J]. 水文地质工程地质,2021,48(1):181 − 187. [FANG Ranke,LIU Yanhui,SU Yongchao,et al. A early warning model of regional landslide in Qingchuan County,Sichuan Province based on logistic regression[J]. Hydrogeology & Engineering Geology,2021,48(1):181 − 187. (in Chinese with English abstract) doi: 10.16030/j.cnki.issn.1000-3665.201911034
[16]	杜国梁,杨志华,袁颖,等. 基于逻辑回归-信息量的川藏交通廊道滑坡易发性评价[J]. 水文地质工程地质,2021,48(5):102 − 111. [DU Guoliang,YANG Zhihua,YUAN Ying,et al. Landslide susceptibility mapping in the Sichuan-Tibet traffic corridor using logistic regression-information value method[J]. Hydrogeology & Engineering Geology,2021,48(5):102 − 111. (in Chinese with English abstract)
[17]	Paraskevas,Tsangaratos. Comparison of a logistic regression and Naïve Bayes classifier in landslide susceptibility assessments:the influence of models complexity and training dataset size[J]. CATENA,2016,145:164 − 179. doi: 10.1016/j.catena.2016.06.004
[18]	YOUSSEF A M,POURGHASEMI H R,POURTAGHI Z S,et al. Landslide susceptibility mapping using random forest,boosted regression tree,classification and regression tree,and general linear models and comparison of their performance at Wadi Tayyah Basin,Asir Region,Saudi Arabia[J]. Landslides,2016,13(5):839 − 856. doi: 10.1007/s10346-015-0614-1
[19]	CHEN Wei,XIE Xiaoshen,WANG jiale,et al. A comparative study of logistic model tree,random forest,and classification and regression tree models for spatial prediction of landslide susceptibility[J]. CATENA,2017,151:147 − 160. doi: 10.1016/j.catena.2016.11.032
[20]	冉光静,李晓,陈刚. 福建省滑坡发育强度分布规律及影响因素分析[J]. 西部探矿工程,2009,21(2):20 − 22. [RAN Guangjing,LI Xiao,CHEN Gang. Distribution law and influencing factors of landslide development intensity in Fujian Province[J]. West-China Exploration Engineering,2009,21(2):20 − 22. (in Chinese with English abstract) doi: 10.3969/j.issn.1004-5716.2009.02.009
[21]	刘艺梁,殷坤龙,刘斌. 逻辑回归和人工神经网络模型在滑坡灾害空间预测中的应用[J]. 水文地质工程地质,2010,37(5):92 − 96. [LIU Yiliang,YIN Kunlong,LIU Bin. Application of logistic regression and artificial neural networks in spatial assessment of landslide hazards[J]. Hydrogeology & Engineering Geology,2010,37(5):92 − 96. (in Chinese with English abstract) doi: 10.3969/j.issn.1000-3665.2010.05.017
[22]	刘福臻,王灵,肖东升. 机器学习模型在滑坡易发性评价中的应用[J]. 中国地质灾害与防治学报,2021,32(6):98 − 106. [LIU Fuzhen,WANG Ling,XIAO Dongsheng. Application of machine learning model in landslide susceptibility evaluation[J]. The Chinese Journal of Geological Hazard and Control,2021,32(6):98 − 106. (in Chinese with English abstract)
[23]	方然可,刘艳辉,黄志全. 基于机器学习的区域滑坡危险性评价方法综述[J]. 中国地质灾害与防治学报,2021,32(4):1 − 8. [FANG Ranke,LIU Yanhui,HUANG Zhiquan. A review of the methods of regional landslide hazard assessment based on machine learning[J]. The Chinese Journal of Geological Hazard and Control,2021,32(4):1 − 8. (in Chinese with English abstract)
[24]	刘艳辉,方然可,苏永超,等. 基于机器学习的区域滑坡灾害预警模型研究[J]. 工程地质学报,2021,29(1):116 − 124. [LIU Yanhui,FANG Ranke,SU Yongchao,et al. Machine learning based model for warning of regional landslide disasters[J]. Journal of Engineering Geology,2021,29(1):116 − 124. (in Chinese with English abstract) doi: 10.13544/j.cnki.jeg.2020-533
[25]	王毅,方志策,牛瑞卿,等. 基于深度学习的滑坡灾害易发性分析[J]. 地球信息科学学报,2021,23(12):2244 − 2260. [WANG Yi,FANG Zhice,NIU Ruiqing,et al. Landslide susceptibility analysis based on deep learning[J]. Journal of Geo-Information Science,2021,23(12):2244 − 2260. (in Chinese with English abstract) doi: 10.12082/dqxxkx.2021.210057
[26]	王毅,方志策,牛瑞卿. 融合深度神经网络的三峡库区滑坡灾害易发性预测[J]. 资源环境与工程,2021,35(5):652 − 660. [WANG Yi,FANG Zhice,NIU Ruiqing. Prediction of landslide susceptibility in Three Gorges Reservoir area based on integrating deep neural network[J]. Resources Environment & Engineering,2021,35(5):652 − 660. (in Chinese with English abstract) doi: 10.16536/j.cnki.issn.1671-1211.2021.05.013
[27]	WANG Yi,FANG Zhice,HONG Haoyuan. Comparison of convolutional neural networks for landslide susceptibility mapping in Yanshan County,China[J]. Science of the Total Environment,2019,666:975 − 993. doi: 10.1016/j.scitotenv.2019.02.263
[28]	林经纬. 福建省滑坡灾害特征及驱动因素分析[J]. 莆田学院学报,2015,22(5):83 − 88. [LIN Jingwei. Characteristics and driving factors of landslide hazard in Fujian Province[J]. Journal of Putian University,2015,22(5):83 − 88. (in Chinese with English abstract)
[29]	刘艳辉,黄俊宝,肖锐铧,等. 基于随机森林的福建省区域滑坡灾害预警模型研究[J]. 工程地质学报,2022,30(3):944 − 955. [LIU Yanhui,HUANG Junbao,XIAO Ruihua,et al. Study on early warning model for regional landslides based on random forest in Fujian Province[J]. Journal of Engineering Geology,2022,30(3):944 − 955. (in Chinese with English abstract) doi: 10.13544/j.cnki.jeg.2021-0625
[30]	刘艳辉, 肖锐铧, 陈春利, 等. 区域滑坡预警中训练样本集的构建方法、系统及存储介质: 20201082-9816.0[P] LIU Yanhui, XIAO Ruihua, CHEN Chunli, et al. Construction method system and storage medium of training sample set in regional landslide early warning: 202010829816.0[P]. 2020-08-18. （in Chinese with Eng-lish abstract）
[31]	CHAWLA N V,BOWYER K W,HALL L O,et al. SMOTE:synthetic minority over-sampling technique[J]. Journal of Artificial Intelligence Research,2002,16:321 − 357. doi: 10.1613/jair.953
[32]	LECUN Y,BOSER B,DENKER J S,et al. Backpropagation applied to handwritten zip code recognition[J]. Neural Computation,1989,1(4):541 − 551.
[33]	GOODFELLOW I, BENGIO Y, COURVILLE A. Deep learning[M]. Cambridge, Massachusetts: The MIT Press, 2016
[34]	SNOEK J, LAROCHELLE H, ADAMS R P. Practical Bayesian optimization of machine learning algorithms[C]//Proceedings of the 25th International Conference on Neural Information Processing Systems-Volume 2. December 3 − 6, 2012, Lake Tahoe, Nevada. New York: ACM, 2012: 2951 − 2959.
[35]	SAMEEN M I,PRADHAN B,LEE S. Application of convolutional neural networks featuring Bayesian optimization for landslide susceptibility assessment[J]. CATENA,2020,186:104249. doi: 10.1016/j.catena.2019.104249
[36]	李亭,田原,邬伦,等. 基于随机森林方法的滑坡灾害危险性区划[J]. 地理与地理信息科学,2014,30(6):25 − 30. [LI Ting,TIAN Yuan,WU Lun,et al. Landslide susceptibility mapping using random forest[J]. Geography and Geo-Information Science,2014,30(6):25 − 30. (in Chinese with English abstract) doi: 10.3969/j.issn.1672-0504.2014.06.006