首页|考虑非滑坡样本选取和集成机器学习方法的水库滑坡易发性预测

考虑非滑坡样本选取和集成机器学习方法的水库滑坡易发性预测

扫码查看
原文链接
  • NETL
  • NSTL
  • 万方数据
准确的滑坡易发性建模对预警预报和风险管控具有重要意义.针对机器学习技术建模中非滑坡样本随机选取和单个分类器存在的精度不高问题,提出了一种耦合多模型的区域滑坡易发性建模框架.以三峡库区秭归‒巴东段为例,选取高程、坡度等 12个因子构建评价指标体系,应用信息量法定量分析各指标对滑坡空间发育的影响程度.随机选取 70%的滑坡作为训练样本,剩余的 30%作为验证样本;应用逻辑回归模型(LR)制作研究区的初始易发性分区图,确定非滑坡随机采样的约束范围.随后,分别采用LR模型约束和无约束条件下随机选取的非滑坡样本,应用单个分类回归树(LR-CART和No-CART)及分类回归树‒Bagging组合模型(LR-CART-Bagging和No-CART-Bagging)开展滑坡易发性建模,并应用多个指标进行精度评估.结果发现:高程和水系等是滑坡发育的主控因素;LR-CART-Bagging模型精度为0.973,高于LR-CART模型的 0.889;相比于No-CART和No-CART-Bagging模型,LR-CART和LR-CART-Bagging模型精度分别提升了0.057和0.047.LR模型可以有效约束非滑坡样本的选取范围,提升样本的选取质量;CART-Bagging模型综合了机器学习和集成学习的优势,预测性能更强,提出的LR-CART-Bagging模型是一种准确可靠的滑坡易发性建模方法.
Reservoir Landslide Susceptibility Prediction Considering Non-Landslide Sampling and Ensemble Machine Learning Methods
Landslide susceptibility evaluation is important for its early warning and forecasting and risk management.To address the problems of a random selection of non-landslide samples and low accuracy of individual classifiers in modeling by machine learning techniques,a coupled multi-model regional landslide susceptibility modeling framework is proposed.Taking the Zigui-Badong section of the Three Gorges reservoir area as an example,12 factors such as elevation and slope were selected to construct an evaluation index system,and the information quantity method was applied to quantify the influence degree of each factor on landslide spatial development.70%of the landslides were randomly selected as training samples and the remaining 30%as validation samples;the Logistic Regression model(LR)was applied to produce an initial susceptibility zoning map of the study area and to determine the constraint range for random sampling of non-landslides.Subsequently,a single Classification and Regression Tree(LR-CART and No-CART)and combined Classification and Regression Tree-Bagging model(LR-CART-Bagging and No-CART-Bagging)were applied to model landslide susceptibility using randomly selected non-landslide samples under the constrained and unconstrained conditions of LR model,respectively,and multiple metrics were applied for accuracy assessment.The results show that elevation and water system are the main controlling factors for landslide development;the accuracy of the LR-CART-Bagging model is 0.973,higher than 0.889 of the LR-CART model;compared with No-CART and No-CART-Bagging models,the accuracy of LR-CART and LR-CART-Bagging models is improved by 0.057 and 0.047,respectively.LR model can effectively constrain the selection range of non-landslide samples and improve the quality of sample selection;the CART-Bagging model integrates the advantages of machine learning and ensemble learning with better prediction performance,and the proposed LR-CART-Bagging model is an accurate and reliable method for landslide susceptibility modeling.

machine learninglandslidessusceptibility mappingnon-landslide samplingensemble learningThree Gorges reservoir areaengineering geology

王悦、曹颖、许方党、周超、余蓝冰、吴立星、汪洋、殷坤龙

展开 >

中国地质大学工程学院,湖北武汉 430074

三峡库区地质灾害野外监测与预警示范中心,重庆 404199

浙江省第十一地质大队,浙江温州 325006

中国地质大学地理与信息工程学院,湖北武汉 430074

展开 >

机器学习 滑坡 易发性制图 非滑坡样本选取 集成学习 三峡库区 工程地质

国家自然科学基金青年基金国家自然科学基金青年基金湖北省重点研发计划

41907253417023302021BCA219

2024

10.3799/dqkx.2022.407

地球科学
中国地质大学

地球科学

CSTPCD北大核心
影响因子：1.447
ISSN：1000-2383
年,卷(期)：2024.49(5)
  • 59