摘要
以贵州尖山营滑坡为工程背景,通过对深度学习的总结与分析,建立多层感知器模型以对该滑坡危险区范围进行非线性预测研究。通过对深度神经网络算法的优化,构建64-128-32-1四层多层感知器模型,并以滑坡最大高差、滑坡体积、滑源区坡度、坡脚坡度、地层倾角作为输入量,以滑坡最大水平运动距离作为输出量对该模型进行训练,实现影响因素与运动距离的非线性映射。根据对贵州省尖山营滑坡调查和研究,尖山营滑坡区域面积约648700 m^(2),体积约1200万m^(3),属于特大型滑坡。依据最优模型对该滑坡进行滑距预测,滑坡平面直线距离1769 m区域内为危险区域。
Geological hazards frequently occurred in China due to its complex geographical location.Landslides are characterized by the outbreak and destructive impacts, threaten the lives and property of civilians are regularly reported in mountainous areas of China.Therefore, predicting dangerous areas of the landslide has significant importance to estimate the affected area and loss, and to protect people′s property and life safety.A method is tried to find for the prediction of the longest horizontal travel distance of the landslide combining K-fold cross-validation and deep learning theory due to complex factors that are hard to find.Jianshanying landslide is a high slope in Faer Town, Shuicheng County, Guizhou Province, and its failure may cause a great loss.A deep neural network model of the high slope′s hazard zone was established to carry out a nonlinear prediction.Five factors are integrated based on its locomotion length studied by foreign and domestic academics, which do not exist multicollinearity, including the largest discrepancy in elevation, the volume, slope gradient, slope angle, and stratum dip angle.Typical rainfall-causing landslide cases in the southwest part of China are also collected by reviewing relevant literature and documentation.The parameters selected from these cases are evaluated by multiple linear regression and random forest regressor to sufficiently understand their feature.Besides, due to the size of the data set, the deep learning theory is applied to the dangerous areas prediction of the landslide with the K-fold cross-validation method.The deep neural network model is built and continuously optimized base on the training data of these cases.Generalization ability is tested by the K-fold cross-validation method, and the best model has been selected.The largest locomotion length, namely the dangerous areas of the landslide, is predicted.The largest discrepancy in elevation of the front edge and back edge of the landslide is the most significant factor in predicting the longest horizontal travel distance of the landslide.Multiple linear regression model is not suitable to solve this issue, except for the discrepancy in elevation.The P-values of its intercept and other indicators are all larger than 5%,so there is no enough evidence to reject the hypothesis H0,their regression coefficients are equal to zero.Besides, according to the model training and prediction, the loss curve shows a staged decrease, dropping to 0.17,which achieves a convergence and displays its good generalization ability.After model training, using the nine examples of testing data to input into the model and compared with the actual target distance, except for the example in Liena, Tibet, the relative errors of the others are between-14% and 14% while their absolute errors are between-160 m and 160 m.The elevation is the main factor, while the volume, slope gradient, slope angle, and stratum dip angle are of equal importance in prediction indicators.The deep neuron network model is highly authoritative and has good generalization ability, so it tends to predict the travel distance of the landslide.After the relevant factors of the Jianshanying landslide are input into model, the result illustrates that the hazard zone starting from its back edge within the linear distance of landslide direction is 1 769 m.
作者
董建辉
钱珂江
赵建军
谢飞鸿
李海军
朱要强
DONG Jianhui;QIAN Kejiang;ZHAO Jianjun;XIE Feihong;LI Haijun;ZHU Yaoqiang(College of Architecture and Civil Engineering,Chengdu University,Chengdu 610106,China;State Key Laboratory of Geological Disaster Prevention and Geological Environment Protection,Chengdu University of Technology,Chengdu 610059,China;Guizhou Geological Environment Monitoring Institute,Guiyang 550000,China)
出处
《南水北调与水利科技(中英文)》
CAS
北大核心
2021年第5期972-981,共10页
South-to-North Water Transfers and Water Science & Technology
基金
国家自然科学基金项目(41877273)。
关键词
滑坡滑距预测
深度神经网络
非线性预测
landslide distance prediction
deep neural network
nonlinear prediction