基于随机森林权重的滑坡危险性评价:以东江流域为例  

A landslide hazard assessment based on random forest weight:a case study in the Dongjiang River Basin

作　　者：吴孝情 赖成光[3,2] 陈晓宏[2,4] 任秀文 

机构地区：[1]环境保护部华南环境科学研究所,广东广州510535 [2]中山大学地理科学与规划学院水资源与环境研究中心,广东广州510275 [3]华南理工大学土木与交通学院,广东广州510641 [4]广东省华南地区水安全调控工程技术研究中心,广东广州510275

出　　处：《自然灾害学报》2017年第5期119-129,共11页Journal of Natural Disasters

基　　金：国家自然科学基金(51479216;51579105;91547202);环保公益性行业科研专项经费项目(201509027)~~

摘　　要：对滑坡潜在区域进行危险性评价是滑坡防治管理的一种有效手段。在评价过程中,指标权重关乎评价结果是否合理。基于随机森林(RF)算法提出一种客观赋权方法,选取8个指标变量构建随机森林权重(RFW)的滑坡危险性评价模型,并对东江流域进行评估,结果表明:增加RF分类树的数目可一定程度降低模型的训练和测试误差;RFW可根据RF分类过程产生的Gini指数减少值计算,其中高程(DEM)与坡度(SL)被识别为最重要的两个指标,而土壤可蚀性指数(SEF)和岩石抗剪能力(SRC)被识别为最不重要的2个指标;评估结果表明危险区域主要位于东江流域的东南部、西部及东北部山区,验证精度高达76.81%,分别高于熵权和CRITIC权的44.93%和42.03%;当去掉RFW中最重要和最不重要的2个指标并分别生成危险图,验证精度分别为75.36%和36.23%,进一步验证基于RF生成的权重科学合理。基于RF算法提出的客观赋权方法应用效果良好,可应用于滑坡危险性评价领域,也可为相关领域权重计算提供新思路。Landslide hazard assessment in potentially dangerous areas is an effective measure for landslide prevention and control management. Index weights play a significant role during the assessment process,determining whether the assessment results are reasonable or not. Based on random forest（ RF）,a method that computes objective weights is proposed. Taking the Dongjiang River Basin as a case study,eight indices are selected to construct the landslide hazard assessment model using the random forest weight（ RFW）. The results show that the training and testing error can be partly reduced by increasing the number of the classification trees of RF; the RFW can be computed by the Gini mean decrease; the indices digital elevation model（ DEM） and slope（ SL） are identified as the most important indices,while soil erodability factor（ SEF） and shear resistance capacity（ SRC） are the leastconsequential; the most dangerous areas are predominantlylocated in the southeastern,western,and northeastern parts of the basin with a verification accuracy of 76. 81%,which is higher than the results of entropy and CRITIC,respectively; and the verification accuracies will be 75. 36% and 36. 23% by removing the two most important and least consequential indices,respectively,further proving the scientificity and rationality of the RFW. The application of objective weights computed by RF presents a good effect,and the method can be used in the field of landslide hazard assessment and provides a new technique for relative fields.

关 键 词：滑坡 权重 危险性评估 随机森林 东江流域 

分 类 号：P642.22[天文地球—工程地质学] X4[天文地球—地质矿产勘探]