中国安全科学学报 ›› 2022, Vol. 32 ›› Issue (3): 194-200.doi: 10.16265/j.cnki.issn1003-3033.2022.03.026

• 防灾减灾技术与工程 • 上一篇    下一篇

基于RBF核的多分类SVM滑塌易发性评价模型

丁茜1(), 赵晓东1,**(), 吴鑫俊1, 张泰丽2, 徐振涛1   

  1. 1大连大学 建筑工程学院,大连 116622
    2中国地质调查局 南京地质调查中心,江苏 南京210016
  • 收稿日期:2021-12-25 修回日期:2022-02-12 出版日期:2022-08-23 发布日期:2022-09-28
  • 通讯作者: 赵晓东
  • 作者简介:

    丁 茜 (1996—),女,河南濮阳人,硕士研究生,研究方向主要为防灾减灾。E-mail:
    赵晓东 教授, 张泰丽 教授级高级工程师

  • 基金资助:
    国家自然科学基金资助(51374046); 浙江丽水地区灾害地质灾害调查项目(DD20190648)

Landslide susceptibility assessment model based on multi-class SVM with RBF kernel

DING Xi1(), ZHAO Xiaodong1,**(), WU Xinjun1, ZHANG Taili2, XU Zhentao1   

  1. 1College of Civil Engineering & Architecture, Dalian University, Dalian Liaoning 116622, China
    2Nanjing Center, China Geological Survey, Nanjing Jiangsu 210016, China
  • Received:2021-12-25 Revised:2022-02-12 Online:2022-08-23 Published:2022-09-28
  • Contact: ZHAO Xiaodong

摘要:

为合理有效地评价地质灾害易发性,依托浙江省温州市飞云江流域地质灾害调查数据,通过有监督的支持向量机(SVM)机器学习模型,建立基于径向基函数(RBF)核的多分类SVM滑塌易发性评价模型。在地理信息系统(GIS)中选取9类特征数据,作为模型训练输入值,采用随机森林(RF)算法分析各指标权重,以高斯函数为RBF核函数,将SVM的输入特征数据映射到高维空间中并加以识别;同时提出高斯核函数超参数C和γ的优化方法,解决SVM模型中超参数不能优化求解的问题;训练后,极易发生类的受试者特征曲线(ROC)的曲线下面积(AUC)达0.97,宏观多分类ROC曲线下AUC值为0.87。训练得到的评价模型既保证了重要分类的精度,又避免过拟合的情况。

关键词: 滑塌, 易发性评价, 支持向量机(SVM), 径向基函数(RBF), 随机森林(RF)

Abstract:

In order to evaluate susceptibility of geological disasters reasonably and effectively, a multi-class SVM model with RBF kernel was established by using supervised machine learning model of SVM based on geological disaster survey data of Feiyun River basin in Wenzhou, Zhejiang Province. In geographic information system (GIS) platform, nine kinds of feature data for training were selected to analyze weight of each indicator by utilizing RF algorithm. Then, Gaussian function was used as RBF kernel function that mapped input feature data of SVM into high dimensional space for identification. In the meantime, an optimal method was put forward to calculate hyper parameters of C and γ in Gaussian function that were difficult to solve out in SVM model. After the model is learned from the training set, its evaluation index, i.e. area under curve (AUC), is up to 0.97 by using receiver operating characteristic curve (ROC) evaluation method, and the AUC of micro-multi-classification ROC is 0.87. The trained evaluation model not only guarantees accuracy of important classifications, but also avoids over-fitting.

Key words: landslide, susceptibility assessment, support vector machine (SVM), radial basis function (RBF), random forest (RF)