基于环境因子优化TSES法选择负样本及其在滑坡易发性评价中的应用  

Optimizing TSES method based on the environmental factors to select negative samples and its application in landslide susceptibility evaluation

作　　者：崔玉龙[1] 朱路路 徐敏 缪海波[1] CUI Yulong;ZHU Lulu;XU Min;MIAO Haibo(School of Civil Engineering and Architecture,Anhui University of Science and Technology,Huainan Anhui 232001,China)

机构地区：[1]安徽理工大学土木建筑学院,安徽淮南232001

出　　处：《地质科技通报》2024年第3期192-199,共8页Bulletin of Geological Science and Technology

基　　金：国家自然科学基金项目(42277136);安徽省高校自然科学优秀青年科研项目(2023AH030041)。

摘　　要：滑坡易发性评价是滑坡灾害防治的重要手段之一,而不合理的滑坡负样本会影响滑坡易发性评价,从而影响到滑坡灾害的防治,因此提供一种合理的负样本选取方法变得尤为关键。以西藏米林市的古滑坡为例,选择高程、坡度、坡向、坡位、距道路距离、距断层距离、距水系距离、地形起伏度、地层岩性、土地利用类型10类环境因子,使用Relief算法计算环境因子的贡献值并依据贡献值优化选择环境因子;基于环境因子优化的目标空间外向化采样法(target space exteriorization sampling,简称TSES)选择负样本,作为性能优异的随机森林模型的输入变量;之后结合优化的环境因子和正或负样本预测米林市的滑坡易发性,并用混淆矩阵和ROC曲线评价构建模型的性能。为检验环境因子优化的TSES法的有效性和先进性,采用耦合信息量法和TSES法选择滑坡负样本并构建随机森林模型,与环境因子优化的TSES法构建的随机森林模型进行对比研究。结果表明,环境因子优化的TSES法构建的随机森林模型的评价效果较好,其ACC为93.7%、AUC为0.987,均高于耦合信息量、TSES法构成的模型。环境因子优化的TSES法能够提高模型的精度,解决多因子作为约束条件取样中因子选取的问题,为滑坡易发性评价采集负样本提供了新的思路。[Objective]Landslide susceptibility evaluation is an important means for landslide disaster prevention and control.Unreasonable negative landslide samples will affect landslide susceptibility evaluation,thereby affecting landslide disaster prevention and control.Therefore,it is particularly critical to provide a reasonable negative sample selection method.[Methods]In view of the selection of negative landslide samples,the ancient landslides in Milin City,Xizang are taken as an example,and 10 environmental factors,including elevation,slope,slope aspect,slope position,distance to road,distance to fault,distance to water system,topographic relief,lithology and land use type,are selected.The Relief algorithm is used to calculate the contribution values of environmental factors and to select the optimized environmental factors.The target space exteriorization sampling(TSES)method based on the optimization of environmental factors is applied to select negative samples as the input variables of the random forest(RF)model with excellent performance.Then,the optimized environmental factors and positive/negative samples are combined to predict the landslide susceptibility of Milin city,and the confusion matrix and receiver operating characteristic(ROC)curve are used to evaluate the prediction performance.To test the effectiveness and advancement of the TSES method optimized by environmental factors,the coupled information method and TSES method are respectively used to select negative landslide samples and constructs RF models to conduct comparative research with the RF model constructed by the TSES method optimized by environmental factors.[Results]The results show that the evaluation effect of the RF model constructed by the optimized TSES method based on environmental factors is better with an ACC value of 93.7%and an AUC value of 0.987,both of which are greater than those of the RF models constructed by the coupling information method and the TSES method.[Conclusion]The TSES optimized by environmental factors can impro

关 键 词：滑坡易发性评价 RELIEF算法 负样本 环境因子优化TSES法 随机森林 古滑坡 

分 类 号：P642.22[天文地球—工程地质学]