滑坡危险度评价对BCS负样本采样的敏感性  

Sensitivity of BCS for Sampling Landslide Absence Data in Landslide Susceptibility Assessment

作　　者：缪亚敏[1] 朱阿兴[1,2,3] 杨琳[2] 白世彪[1] 刘军志[1] 邓永翠[1] 

机构地区：[1]虚拟地理环境教育部重点实验室(南京师范大学),江苏省地理环境演化国家重点实验室培育建设点,江苏省地理信息资源开发与利用协同创新中心,江苏南京210023 [2]中国科学院地理科学与资源研究所,资源与环境信息系统国家重点实验室,北京100101 [3]Department of Geography,University of Wisconsin-Madison

出　　处：《山地学报》2016年第4期432-441,共10页Mountain Research

基　　金：国家自然科学基金项目(41431177;41471178);江苏省高校自然科学研究重大项目(14KJA170001);国家重点基础研究发展计划973项目(2015CB954102)~~

摘　　要：滑坡负样本在基于统计方法的滑坡危险度制图中具有重要作用,能够抑制统计方法对滑坡危险度的高估。缓冲区控制采样（Buffer controlled sampling,BCS）是一种广泛使用的负样本采样方法,其原理是认为滑坡点附近一定范围内的地理环境与滑坡点所在的地理环境相似,易发生滑坡,因而应当在灾害点一定缓冲区以外的区域采集负样本。目前缓冲区大小主要是根据专家对研究区的经验知识确定,具有主观性。缓冲区大小对基于统计方法的滑坡危险度制图的影响研究较少。因此,有必要分析缓冲区大小与滑坡危险度制图精度之间的关系,探究适宜的缓冲区大小。以陇南山区的油坊沟流域为研究区,基于BCS负样本采样方法,探究不同缓冲区大小对基于支持向量机（Support vector machine,SVM）的滑坡危险度制图结果的影响。结果表明：缓冲区过小会导致与滑坡点地理环境相似的假的负样本的存在,从而导致滑坡危险度的低估;缓冲区过大会导致负样本在环境特征空间中太局限,负样本集的全局代表性差,从而导致滑坡危险度的高估。在本研究区基于SVM的滑坡危险度制图中,200~500m是使用BCS采集负样本的较理想的缓冲区大小。Landslide absence data plays an important role in data-driven models for landslide susceptibility mapping. It can constrain the overestimation of predicted landslide susceptibility values. Buffer controlled sampling（ BCS） is widely used in sampling landslide absence data. It is based on the general principle that the area near the landslide occurrences has similar geo-environment with landslides,resulting it prone to landslides. Thus landslide absence data should be sampled from the areas beyond the buffer zones of the landslide sites. Currently the buffer size is decided subjectively based on the experts＇ knowledge of the study area. The study of the effect of buffer size on data-driven models for landslide susceptibility mapping is rare. It is important to study the general relationships between buffer size and mapping accuracy and find an appropriate buffer size for an given area. In this study,BCS sampling strategy was used in the Youfang ravine in the south Gansu of China for sampling landslide absence data and Support Vector Machine（ SVM） was used to deliniute landslide susceptibility across the study area. Results show that if the buffer size be small,false absence data would be included in the generated absence datasets and result in the underestimation of the predicted landslide susceptibility values. If the buffer size be large,the representativeness of generated landslide absence data for the whole study area is low,resulting in the overestimation of the predicted landslide susceptibility values. In the Youfang ravine,the appropriate range of buffer size in BCS for sampling landslide absence data is from 200 m to 500 m in SVM for landslide susceptibility mapping.

关 键 词：SVM 负样本 缓冲区控制采样 缓冲区大小 滑坡危险度制图 

分 类 号：P642.22[天文地球—工程地质学]