基于不同特征选择方法和随机森林法的滑坡易发性评价——以湖南中西部地区为例

1.湖南省自然资源事务中心,长沙 410118;2.遵义师范学院,贵州遵义563006;3.中南大学,长沙 410083;4.湖南容诚致远工程技术有限公司,长沙 410000

滑坡;最大互信息系数;递归特征选择;随机森林

Landslide Susceptibility Assessment Based on Different Feature Selection Methods and Random Forest Method—a Case Study of Central and Western Hunan
DUAN Zhongman1,JIA Liangliang2,3,JIANG Mingguang3,4,LEI Yaobo1,CHEN Yana1

1.Hunan Center of Natural Resources Affairs , Changsha 410118, China;2.Zunyi Normal University, Zunyi 563006, China;3.Central South University , Changsha 410083, China;4.Hunan Rongcheng Zhiyuan Engineering Technology Co., Ltd., Changsha 410000, China

Landslide;Maximal mutual information coefficient;Recursive feature selection;Random forest

DOI: 10.13512/j.hndz.2023.02.13

备注

湘中、湘西地区是湖南省滑坡地质灾害最为频发的地区,同时该区旅游资源和自然资源丰富,是滑坡管理的重点区域。为研究湘中、湘西地区滑坡易发性评价模型的适用性,以湖南中西部地区为例,在初步选取的15个滑坡致灾因子的基础上,采用最大互信息系数、递归特征选择、基于随机森林的基尼不纯度指标和平均精确度指标等方法开展滑坡致灾因子优化,分析剔除了平面曲率和剖面曲率两个不重要因子,最终提取了13个重要因子,利用随机森林模型开展了研究区易发性评价,并采用最近两年滑坡数据开展验证。结果表明:不同特征选择方法优化后的滑坡因子结合随机森林模型所得的模型结果与实际情况吻合性较好,中、较高和高易发区滑坡占比77.58%,验证结果为79.58%,该模型对湘中、湘西地区地质灾害易发性评价模型选取提供了参考与借鉴。
The central and western Hunan is the most frequent area of landslide geological disasters in Hunan Province. At the same time, the area is rich in tourism and natural resources, and is the key area for landslide management. In order to study the applicability of the landslide susceptibility evaluation model in central and western Hunan,the historical landslides points and their corresponding features in central and western Hunan are taken as analysis data. Based on the 15 landslide disaster factors initially selected, the maximum mutual information coefficient, recursive feature selection, Gini impurity index based on random forest and average accuracy index are used to optimize the landslide disaster factors. Two unimportant factors of plane curvature and profile curvature are eliminated,and 13 important factors are finally extracted. The random forest model is used to evaluate the susceptibility of the study area, and the landslide data in the last two years are used for verification. The results show that the model results obtained by combining the landslide factors optimized by different feature selection methods with the random forest model are in good agreement with the actual situation. The proportion of landslides in medium, high and high prone areas is 77.58 % , and the verification result is 79.58 % . The model provides a reference for the selection of geological disaster susceptibility evaluation models in central and western Hunan.
·