Evaluation of landslide susceptibility in Beichuan County based on grids
王潇 1魏秋燕 1董建辉 1冉培廉 2刘亮 2黄秋香 3徐湘涛 3李少达2
扫码查看
点击上方二维码区域，可以放大扫码查看
作者信息
1. 成都大学建筑与土木工程学院,成都 610106
2. 成都理工大学地球与行星科学学院,成都 610059
3. 成都理工大学环境与土木工程学院,成都 610059
折叠
摘要
滑坡易发性评价的实质就是以历史滑坡数据为基础,进行特定区域滑坡灾害发生的概率评估.易发性评价结果多数取决于样本的精细程度.传统的样本制作方法会丢失滑坡的部分位置信息,为最终评价结果带来不确定性.本研究提出了一种全新的网格样本制作方法,尽可能完整地保留滑坡的边界位置信息.将不同的机器学习模型(逻辑回归模型、深度神经网络)与本文提出的样本制作方法结合,并通过受试者工作特征(receiver operating characteris-tic,ROC)曲线实现精度验证.ROC曲线中2个模型的AUC(area under curve)值分别为0.878,0.963.最终的易发性分区结果显示:深度神经网络在对于极高滑坡易发区的划分更为精细,便于节约人力、物力资源,集中关注于滑坡真正高发的那些区域.
Abstract
The essence of landslide susceptibility assessment is to use historical landslide data as a basis and to conduct a probability assessment of landslide occurrence in a given area.Most of the results of susceptibility evaluations depend on the resolution of the sample.The traditional sample production method loses part of the location information of landslides,which brings uncertainty to the final evaluation results.In this study,we propose a new method of producing grids samples to preserve the boundary location information of landslides as completely as possible.Different machine-learning models(a logistic regression model and a deep neural network)are combined with the sample production method proposed in this paper,and accuracy validation is achieved through receiver operating characteristic(ROC)curves.The area under the curve values for the two models are 0.878 and 0.963,respectively.The final results of the susceptibility partitioning show that the deep neural network is much more refined in the partitioning of very high landslide susceptibility zones.This approach allows us to save human and material resources and focus on those areas with very high landslide susceptibility.
关键词
滑坡/易发性评价/深度神经网络/逻辑回归模型
Key words
landslides/susceptibility assessment/deep neural network/logistic regression model
万方数据