基于RBF核的多分类SVM滑塌易发性评价模型  

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

作　　者：丁茜 赵晓东 吴鑫俊 张泰丽[2] 徐振涛 DING Xi;ZHAO Xiaodong;WU Xinjun;ZHANG Taili;XU Zhentao(College of Civil Engineering&Architecture,Dalian University,Dalian Liaoning 116622,China;Nanjing Center,China Geological Survey,Nanjing Jiangsu 210016,China)

机构地区：[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

出　　处：《中国安全科学学报》2022年第3期194-200,共7页China Safety Science Journal

基　　金：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)。

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。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.

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) 

分 类 号：X915.5[环境科学与工程—安全科学]