基于BP神经网络模型的滑坡易发性评价方法

滑坡是中国频发的地质灾害，滑坡的易发性评价涉及多种影响因素，如何利用多影响因素进行精确、有效的滑坡易发性评价是滑坡减灾防灾工作的重点和前提。为探讨基于反向传播(back propagation, BP)神经网络模型的不同滑坡易发性评价方法的适用性，以川西蒲江县为研究区，通过实地调查与编录，筛选地质、地貌、环境等12类影响因子，分析各影响因子与滑坡的相关性，确定影响因子的权重大小，构建BP神经网络模型，完成因子权重法和栅格赋值法的滑坡易发性评价图编制和精度评价。结果显示：研究区筛选的12类滑坡影响因子不存在线性相关，坡度、地形湿度指数(topographic wetness index, TWI)和距道路距离对区内滑坡发育影响明显，利用滑坡影响因子构建的BP神经网络模型可对滑坡易发性进行有效的定量评价。综合现场调查与接收者操作特征(receiver operating characteristic, ROC)曲线精度分析，结果表明：基于BP神经网络模型的栅格赋值法和因子权重法曲线下面积(area under curve, AUC)分别为0.86和0.798,栅格赋值法评价精度优于因子权重...

Study on Landslide Susceptibility Evaluation Methods Based on BP Neural Network Model

Landslides are frequent geological disasters in China, landslide susceptibility evaluation involves many influencing factors. How to use multiple influencing factors to conduct accurate and effective landslide susceptibility evaluation is the key and premise of landslide disaster reduction and prevention. In order to discuss the applicability of different landslide susceptibility evaluation methods based on BP neural network model, Pujiang County in western Sichuan were took as the study area, and 12 types of impact factors such as geology, landform and environment were chose after field survey and cataloging, the correlation between each impact factor and landslide were analyzed, the weight of the impact factor were determined, the BP neural network model were constructed, and the preparation and accuracy evaluation of landslide susceptibility evaluation map by factor weight method and grid assignment method were completed. The results show that there are no linear correlations among the 12 types of landslide influencing factors selected in the study area, and the slope, TWI and distance from the road have obvious effects on the landslide development in the area. The BP neural network model constructed by the landslide influencing factors can effectively carry out the quantitative evaluation of landslide susceptibility. Based on the field investigation and ROC accuracy analysis, the grid assignment method based on BP neural network model (AUC value is 0.86) is superior to the factor weight method (AUC value is 0.798) in evaluation accuracy. The grid assignment method based on BP neural network model is more suitable for landslide susceptibility evaluation in the study area.