滑坡点采样模式对易发性评价结果的对比分析  

Comparative Study of Landslide Susceptibilities Using the Machine Learning MaxEnt Model on Different Sampling Strategies

作　　者：史进昌 SHI Jinchang(Shaanxi Institute of Engineering Prospecting,Xi’an 710068,China)

机构地区：[1]陕西工程勘察研究院有限公司,陕西西安710068

出　　处：《灾害学》2023年第4期50-55,共6页Journal of Catastrophology

基　　金：国家自然科学基金青年科学基金项目(42201011)。

摘　　要：线性工程是我国区域均衡发展的生命线,对线性工程沿线进行滑坡易发性评价对线路维护和防灾减灾有着巨大的支撑意义。该文以标志性工程滇中引水工程大理段Ⅱ段为例,采用常用于生态位评价中的机器学习方法最大熵模型(Maximum entropy model,MaxEnt)评估模型,以30 m分辨率DEM提取的高程、地形湿度指数、坡度、平/剖面曲率、坡向、地面粗糙度,1∶50万地质图提取的工程地质岩组、距断层距离,哨兵2号遥感影像提取的距道路距离、土地利用类型和植被覆盖度指数NDVI等共计12个评价指标,基于GIS平台和开源MaxEnt软件,对滑坡中部点采样和滑坡后壁点采样模式分别完成了该线路工程沿线滑坡易发性评价。两种采样模式的模型训练和预测精度均证明了该模型的有效性。同时,该模型自带的Jackknife检验模型揭示了NDVI、高程和距道路距离对滑坡贡献率最高,点采样模式影响地形指标的贡献度。ROC曲线证明,滑坡中部采样的易发性评价训练精度较高,LR Class指数证明易发性分区结果优于滑坡后壁采样模式;但滑坡后壁采样模式的在验证精度上远优于前者,且评价结果精度与前者相比很是接近。Linear engineering is the lifeline of balanced regional development in China,and the assessment of landslide susceptibility of the linear engineering has tremendous support significance for line maintenance and disaster prevention and mitigation.we take the Dali Section II of the Central Yunnan Water Diversion Project,a landmark project,as an example.Taking the maximum entropy model(MaxEnt),a machine learning method which commonly used in ecological potential evaluation,we complete the landslide susceptibility assessment on zone of study regions using 12 index as elevation,slope,slope direction,terrain moisture index,flat/profile curvature,roughness,engineering geological rock group,distance from faults,distance from roads,normalized vegetation index(NDVI)and land use type.Based on a GIS platform and open source MaxEnt software,two landslide susceptibility zoning map is conducted using two sampling types of centroid of landslide scarp and centroid of landslide body or boundary.The model training and prediction accuracies all proved the validity of the model.Meanwhile,the model built-in Jackknife test model revealed that the distance from the road,elevation and NDVI all are heavy contributed to the landslides happen and sampling ways do influence the results by influencing the topography index.ROC curve shows that a higher training accuracy using centroid of landslide body or boundary is appeared,and the LR class index is better than results using centroid of landslide scarp.While the latter performed better on test accuracy than the former one.

关 键 词：滑坡 MaxEnt模型 LR Class指数 易发性评价 滇中引水工程 

分 类 号：X43[环境科学与工程—灾害防治] X915.5[天文地球—地质学] P694