摘要
帕隆藏布然乌—通麦段是连接西藏与内地的战略要道,区域内冻融侵蚀严重,滑坡灾害频发。本研究基于Google Earth解译与野外核查构建滑坡编目,选取冻融强度指数(FTI)等12个特征变量,采用随机森林(RF)、支持向量机(SVM)、多层感知器(MLP)模型完成了区域滑坡的易发性评价和主控因素分析。结果表明:区域内共发育以中小型为主的296处滑坡,主要分布在海拔3500~4500 m、优势坡向为阳坡(67.5°~202.5°)和FTI值大于0.6的区域;RF模型的滑坡灾害预测性能最优(AUC=0.90),SVM与MLP模型次之;主控因素分析结果表明,坡向、高程、夏季降雨量和FTI均为区域滑坡的主要控制因素,其中高程和坡向控制着滑坡发育的物质基础和能量条件,降雨和冻融是滑坡发育的外部触发因素;研究区可分为古乡—通麦段、然乌—玉普段两大防治靶区,其中古乡—通麦段需做好排水措施,然乌—玉普段需做好冻融防护。本研究可为沿线潜在滑坡灾害风险防控与后续工程防灾减灾提供参考。
The Ranwu-Tongmai section of the Parlung Zangbo serves as a critical strategic corridor connecting Xizang with inland China.This area is characterized by complex landforms,deeply incised valleys,widely distributed maritime glaciers and seasonally frozen soil,and high sensitivity to climate change.In the context of global climate change,the region experiences severe freeze-thaw erosion and frequent landslide hazards,which often damage roads,disrupt traffic,and cause significant losses.With China’s continuously increasing investment in Xizang’s infrastructure development,the exchange of personnel and goods through this corridor has undergone rapid growth.Consequently,there is an urgent need to systematically identify potential landslide hazards and implement specifically targeted risk prevention measures to effectively reduce regional landslide risks.This study utilized multi-temporal sub-meter high-resolution remote sensing images archived in Google Earth from 2015 to 2023.Based on interpretation indicators such as color variations,drainage patterns,landform characteristics,and vegetation changes,landslides were identified and delineated through visual interpretation.The final landslide inventory was determined through a field investigation conducted in April 2024,with remote sensing interpretations rigorously validated via on-site inspections to ensure accuracy.Twelve landslide-related characteristic variables,including the freeze-thaw intensity index(FTI)and summer rainfall,were selected.Correlation analysis was conducted to remove variables with high collinearity,and the remaining variables were used to construct the dataset.A 90 m grid cell was selected as the evaluation unit.The landslides involved 3271 grid cells in total,which were set as positive samples with a label value of 1.An equal number of non-landslide grid cells were randomly selected within the study area as negative samples,resulting in a total of 6542 samples.Seventy percent of both positive and negative samples were randomly allocated to construct the training dataset,while the remaining 30%were reserved exclusively for the testing dataset.Both training and testing datasets maintained an identical 1∶1 ratio of positive to negative samples.Three machine learning models—random forest(RF),support vector machine(SVM),and multilayer perceptron(MLP)—were trained and validated using these datasets to assess regional landslide susceptibility.The importance of each characteristic variable obtained from the RF algorithm and correlation analysis was used to analyze the main controlling factors.The results showed that the study area had abundant rainfall and contained 296 landslides in total,predominantly small to medium in size.These landslides were mainly distributed within the elevation range of 3500~4500 m,on sunny slopes(67.5°~202.5°),and in regions with FTI values greater than 0.6.All three algorithms—RF,SVM,and MLP—demonstrated robust landslide prediction performance.Among these,RF achieved superior performance indicators:Accuracy(0.82),Recall(0.85),F1-score(0.83),Jaccard Index(0.70),and AUC(0.90),outperforming SVM(AUC=0.87)and MLP(AUC=0.86).Feature importance analysis revealed four main controlling factors on landslide susceptibility:slope aspect(25.63%),elevation(18.82%),summer rainfall(14.05%),and freeze-thaw intensity index(9.83%).Among them,elevation and slope aspect determined the material basis and energy conditions for slope landslides,while rainfall and freeze-thaw cycles were external triggering factors for landslide development.The susceptibility assessment results from the RF,SVM,and MLP models showed that the distribution of high and very high susceptibility zones was basically consistent,and the areas of very high susceptibility zones were nearly identical.The study area was divided into two target zones for landslide hazard mitigation.For the Guxiang-Tongmai section,special attention should be given to the impact of rainfall,with landslide mitigation focusing on improving water-catching structures and drainage systems.Meanwhile,for the Ranwu-Yupu section,landslide mitigation should prioritize the effects of freeze-thaw cycles,implementing measures such as insulation and seepage prevention to enhance the frost resistance of the geotechnical materials.These findings not only contribute to the risk prevention and control of potential landslide hazards along the Ranwu-Tongmai section of the Parlung Zangbo in the high-altitude alpine mountainous areas,but also provide a reference for subsequent engineering decision-making in disaster mitigation and prevention for regional landslide hazards.
作者
张智
江耀
鲁兴生
任彤捷
路昊天
赵巍
刘云辉
ZHANG Zhi;JIANG Yao;LU Xingsheng;REN Tongjie;LU Haotian;ZHAO Wei;LIU Yunhui(Institute of Mountain Hazards and Environment,Chinese Academy of Sciences,Chengdu 610299,China;China-Pakistan Joint Research Center on Earth Sciences,CAS-HEC,Islamabad 45320,Pakistan;University of Chinese Academy of Sciences,Beijing 100049,China)
出处
《冰川冻土》
2025年第5期1401-1415,共15页
Journal of Glaciology and Geocryology
基金
国家自然科学基金项目(U20A20112)
中国科学院、水利部成都山地灾害与环境研究所科研项目(IMHE-ZDRW-07)资助。
关键词
滑坡易发性
机器学习
主控因素
冻融强度指数
帕隆藏布
landslide susceptibility
machine learning
main controlling factors
freeze-thaw intensity index
Parlung Zangbo
