| 基于层数自适应加权卷积神经网络的川藏交通廊道沿线滑坡易发性评价 |
| |
| 引用本文： | 黄武彪,丁明涛,王栋,蒋良文,李振洪.基于层数自适应加权卷积神经网络的川藏交通廊道沿线滑坡易发性评价[J].地球科学,2022,47(6):2015-2030. |
| |
| 作者姓名： | 黄武彪 丁明涛 王栋 蒋良文 李振洪 |
| |
| 作者单位： | 1.长安大学地质工程与测绘学院, 陕西西安 710054 |
| |
| 基金项目： | 国家自然科学基金(Nos.41941019,42090053)；;中央高校基本科研业务费专项资金(Nos.300102269208,300102260404)； |
| |
| 摘 要： | 开展铁路沿线滑坡易发性评价对川藏交通廊道工程建设及运维过程中的风险管理具有重要意义.提出一种层数自适应、通道加权的卷积神经网络(layer adaptive weighted convolutional neural network，LAW-CNN)，对川藏交通廊道沿线滑坡易发性进行评价.依据野外调查和影响因素分析筛选出影响滑坡发生的影响因子，绘制滑坡编目，构造用于易发性评价的实验数据集；针对卷积神经网络的权重初值、网络层数等超参数难以优化设置的问题，提出基于影响因子信息熵的通道加权方法和网络层数优选策略，通过多通道加权和层数自适应分类卷积的方式提出滑坡易发性制图的LAW-CNN架构；搜索最优LAW-CNN网络结构并训练网络参数，获取研究区滑坡发生概率并进行易发性分级评价.所提的LAW-CNN模型可以不同权重和不同深度挖掘影响因子的深层特征，实验结果表明，模型曲线下面积(area under curve，AUC)值为0.852 8，极高易发区滑坡点密度为1.251 9，均优于SVM(support vector machine)和CNN模型；川藏交通廊道沿线滑坡极高和高易发区主要集中在大江大河两侧以及横断山区.LAW-CNN模型可较好评价川藏交通廊道滑坡易发性，能够为川藏交通廊道的建设和灾害防治提供科学的依据.
|
| 关 键 词： | 滑坡易发性 川藏交通廊道 层数自适应 多通道加权 卷积神经网络 滑坡 |
| 收稿时间： | 2021-09-30 |
Evaluation of Landslide Susceptibility Based on Layer Adaptive Weighted Convolutional Neural Network Model along Sichuan-Tibet Traffic Corridor |
| |
| Abstract: | It is of great significance for disaster risk management in the process of railway engineering construction, operation and maintenance to carry out precise landslide susceptibility assessment along the Sichuan-Tibet traffic corridor. In this paper, a layer adaptive weighted convolutional neural network (LAW-CNN) is proposed to evaluate the landslide susceptibility along the Sichuan-Tibet traffic corridor. According to the field investigation and influencing factor analysis, the influencing factors are selected, the landslide catalogue and the spatial database is constructed.To optimize the initial weight and the layer number of the CNN network, the channel weighted method and the network layer optimization strategy based on the influence factor information entropy are proposed, and the LAW-CNN architecture is constructed by multi-channel weighted convolution and multi-layer classification convolution. The optimal LAW-CNN structure is searched and the network parameters are trained to obtain the landslide occurrence probability in the study area, followed by a susceptibility classification evaluation.The proposed LAW-CNN model can fully represent the deep characteristics of the factor layers with different weights and depths.The experimental results show that the area under curve value of the proposed model is 0.852 8 and the landslide density in the very high susceptibility area is 1.251 9, which are better than the SVM and CNN models.The very high and high susceptibility areas are mainly concentrated on both sides of large rivers and the Hengduan Mountain Range.The LAW-CNN model can precisely assess landslide susceptibility, and then provide a scientific basis for the construction of the Sichuan-Tibet traffic corridor and disaster prevention. |
| |
| Keywords: | |
|
| 点击此处可从《地球科学》浏览原始摘要信息 |
|
点击此处可从《地球科学》下载全文 |
|
