ISSN 1003-8035 CN 11-2852/P
    尹超,李仲波,刘新良,等. 基于滑坡分类和改进卷积神经网络的滑坡敏感性区划[J]. 中国地质灾害与防治学报,2023,34(0): 1-14. DOI: 10.16031/j.cnki.issn.1003-8035.202301003
    引用本文: 尹超,李仲波,刘新良,等. 基于滑坡分类和改进卷积神经网络的滑坡敏感性区划[J]. 中国地质灾害与防治学报,2023,34(0): 1-14. DOI: 10.16031/j.cnki.issn.1003-8035.202301003
    YIN Chao,LI Zhongbo,LIU Xinliang,et al. Landslide susceptibility assessment and zonation based on landslide classification and improved Convolutional Neural Networks[J]. The Chinese Journal of Geological Hazard and Control,2023,34(0): 1-14. DOI: 10.16031/j.cnki.issn.1003-8035.202301003
    Citation: YIN Chao,LI Zhongbo,LIU Xinliang,et al. Landslide susceptibility assessment and zonation based on landslide classification and improved Convolutional Neural Networks[J]. The Chinese Journal of Geological Hazard and Control,2023,34(0): 1-14. DOI: 10.16031/j.cnki.issn.1003-8035.202301003

    基于滑坡分类和改进卷积神经网络的滑坡敏感性区划

    Landslide susceptibility assessment and zonation based on landslide classification and improved Convolutional Neural Networks

    • 摘要: 滑坡敏感性区划是根据研究区域的滑坡调查数据和地质环境条件,分析致灾因子的组合特征对滑坡发生的影响并将研究区域划分为不同等级的敏感区,为土地利用规划和滑坡防治政策制定提供理论依据。对山东省淄博市博山区开展了工程岩组和地质灾害调查,结合多种数据源和地质调查资料建立了博山区滑坡数据库;采用Pearson相关系数法筛选了滑坡致灾因子,基于信息量法分别对全部滑坡、自然滑坡和工程滑坡的致灾因子进行分级;建立了包含2轮卷积和池化的7层改进卷积神经网络,分别对全部滑坡、自然滑坡和工程滑坡进行了评价模型的训练和验证,采用AUC法验证了模型的准确性;基于ArcGIS12.0计算了博山区所有栅格的滑坡敏感性概率,绘制了博山区滑坡敏感性区划图。研究结果表明:滑坡致灾因子包括高程、坡度、坡向、剖面曲率、平面曲率、距河流距离、地表流沙输送量、地形湿度指数、距道路距离、土地利用类型、距断层距离、工程岩组和归一化植被覆盖指数;博山区滑坡敏感性概率最小为0.136、最大为0.841;极高敏感区、高敏感区、中敏感区、低敏感区和极低敏感区分别占博山区总面积的8.08%(56.4 km2)、17.62%(123.0 km2)、25.33%(176.8 km2)、32.87%(229.4 km2)和16.10%(112.4 km2),极高敏感区主要分布在博山区西北部、南部、东北部及其他地区,其中,自然滑坡主要分布于西北部极高敏感区,工程滑坡主要分布于东北部极高敏感区。

       

      Abstract: Landslide susceptibility regionalization aims to analyze the combined characteristics of hazard-inducing factors and their impact on the probability of landslide occurrence, thereby dividing the study area into different susceptible zones. This provides a theoretical basis for land use planning and the formulation of policies for landslide prevention and control. A landslide database was established for Boshan District by integrating various data sources and geological survey data. Hazard-inducing factors were selected using the Pearson correlation coefficient method, and these factors were classified for all landslides, natural landslides, and engineering landslides using the information content method. An improved 7-layer Convolutional Neural Network (CNN) with two rounds of convolution and pooling was constructed. Evaluation models were trained and verified separately for all landslides, natural landslides, and engineering landslides, and the accuracy of the models was assessed using the AUC method. The landslide susceptibility probabilities of all grid cells in Boshan district were calculated, and a landslide susceptibility regionalization map was generated using ArcGIS12.0. The results indicate that the hazard-inducing factors include elevation, slope, aspect, profile curvature, plane curvature, distance from rivers, STI, TWI, distance from roads, land use, distance from faults, lithology, and normalized difference vegetation index (NDVI). The minimum and maximum landslide susceptible probabilities in Boshan district are 0.136 and 0.841, respectively. The extremely high susceptibility, high susceptibility, medium susceptibility, low susceptibility, and extremely low susceptibility zones account for 8.08% (56.4 km2), 17.62% (123.0 km2), 25.33% (176.8 km2), 32.87% (229.4 km2) and 16.10% (112.4 km2) of the total area. The extreme high susceptibility zones are primarily distributed in the northwest, south, northeast and other areas. Natural landslides are mainly concentrated in the extremely high susceptibility zones, while engineering landslides are mainly found in the extremely high susceptibility zones in the northeast.

       

    /

    返回文章
    返回