基于卷积神经网络的滑坡易发性评价:以辽南仙人洞国家级自然保护区为例  

Susceptibility Assessment of Landslides Based on Convolutional Neural Network Model:A Case Study from Xianrendong National Nature Reserve in Southern Liaoning Province

作　　者：郑德凤[1,2] 高敏[2] 闫成林 李媛媛[2] 年廷凯[3] Zheng Defeng;Gao Min;Yan Chenglin;Li Yuanyuan;Nian Tingkai(Liaoning Key Laboratory of Physical Geography and Geomatics,Liaoning Normal University,Dalian 116029,China;School of Geography,Liaoning Normal University,Dalian 116029,China;School of Infrastructure Engineering,Dalian University of Technology,Dalian 116024,China)

机构地区：[1]辽宁师范大学辽宁省自然地理与空间信息科学重点实验室,辽宁大连116029 [2]辽宁师范大学地理科学学院,辽宁大连116029 [3]大连理工大学建设工程学院,辽宁大连116024

出　　处：《地球科学》2024年第5期1654-1664,共11页Earth Science

基　　金：国家自然科学基金项目(Nos.42077272,42377185);辽宁师范大学高端科研成果培育资助计划项目(No.23GDL007)。

摘　　要：为了解决滑坡易发性评价过程中存在的滑坡编录数据不足,主观或者随机地选取非滑坡栅格单元而导致模型准确率较低等问题,以辽南仙人洞国家级自然保护区为研究区,首先,从地形地貌、地质条件、水文气象条件和人类工程活动等方面选取了12个评价因子构建滑坡评价体系;其次,利用SMOTETomek综合采样方法解决滑坡与非滑坡样本类别的比例失衡问题,进而建立滑坡易发性评价模型的数据集;最后,针对研究区东西两侧(A区和B区)的非线性滑坡数据,通过构建卷积神经网络(Convolutional Neural Networks,CNN)模型进行滑坡易发性评价,并精准绘制了研究区滑坡易发性分布图.结果表明:CNN模型具有良好的适应性,绘制的滑坡易发性分区图显示出合理的空间分布,A区和B区的测试集AUC面积分别为91.2%和94.3%;70%的滑坡数据分布在较高及以上等级的易发区,68.7%的非滑坡数据分布在较低及以下等级的易发区;滑坡高易发区主要位于研究区东北部猫岭北沟山一带、冰峪沟风景区的北部山区和碧流河水库沿岸区.研究成果为辽南仙人洞国家级自然保护区的地质灾害防治规划、应急预案制定等提供了重要的科学依据.In order to solve the problems of insufficient landslide catalog data in the process of landslide susceptibility evaluation,and low model accuracy due to subjective or random selection of non-landslide raster cells,12 evaluation factors were selected from the aspects of topography and geomorphology,geological conditions,hydrometeorological conditions and human engineering activities to construct a landslide evaluation system for Xianrendong National Nature Reserve in southern Liaoning Province in this paper;Furthermore,the imbalance of sample categories between landslide and non-landslide was solved based on SMOTETomek comprehensive sampling method,and then a dataset of landslide susceptibility evaluation was established;Finally,for the nonlinear landslide data in the east and west sides of the study area(zones A and B),the convolutional neural networks(CNN)model was constructed to evaluate the landslide susceptibility,and the distribution map of landslide susceptibility in the study area was accurately drawn.The results show that the CNN model had good adaptability,and the zoning map of landslide susceptibility shows a reasonable spatial distribution.The AUC area of the test set in part A and part B of the study area was 91.2%and 94.3%,respectively.70%of landslides were distributed in higher and above grade prone areas,and 68.7%of non-landslides were distributed in lower and below grade prone areas.The high prone area of landslide is mainly located in the area of Maoling Beigou Mountain in the northeast of the study area,the northern mountainous area of the Bingyugou Scenic Area,and the coastal area of the Biliuhe Reservoir.The research results can provide an important scientific basis for the planning of geological disaster prevention and control,and the formulation of emergency plans in Xianrendong National Nature Reserve in southern Liaoning Province.

关 键 词：滑坡 卷积神经网络 综合采样方法 易发性 仙人洞国家级自然保护区 工程地质 

分 类 号：P642[天文地球—工程地质学]