基于多层感知器模型的区域滑坡敏感性评价研究——以四川低山丘陵区为例

Research on Regional Landslide Susceptibility Assessment Based on Multiple Layer Perceptron——Taking the Hilly Area in Sichuan as Example

以四川省低山丘陵区为研究区,基于滑坡编目数据,在深入分析研究区滑坡孕灾环境的基础上,选取坡度、地形起伏度、岩土类型和断裂构造等八类孕灾环境因子。通过融合确定性系数和多层感知器,提出CF-MLP模型,并对研究区滑坡的敏感性进行评价,计算得到的滑坡敏感性指数在ROC曲线中的线下面积为0.84,说明该模型预测结果对滑坡具有较好的识别作用;基于模型预测结果对研究区进行敏感性区划,共分为高敏感性、中敏感性和低敏感性三个区域,与历史滑坡的分布现状相一致。实践证明,CF-MLP模型在一定程度上解决了滑坡孕灾环境因子数据的合理量化问题,提高了多层感知器网络的收敛效果,有利于建立更为准确的滑坡敏感性分析模型。

Research on the hazard assessment of regional landslide contributes essential practical significance and study value to disaster preventing and rescuing,land planning,and disaster managing and decision,and so on.The hilly region of Sichuan Province is located in the east of the province,this area is in the subtropical monsoon climate zone,and its abundant rainfalls provide sufficient water conditions for the development of landslides,which make the highest density of landslides in our country.This paper,taking it as the study area,selects 8factors（including height,slope,relief degree of land surface,lithology,fault,road,river,NDVI etc.）as disaster pregnant environmental factors of the landslide with analyzing the pregnant environment in depth based on the landslide catalog database,and then introduces the CF-MLP model by fusion of certainty factor and multiple layer perceptron to assess landslide susceptibility in the study area,the better recognition results to landslide are proved because that the AUC under the ROC of the landslide susceptibility index calculated by CF-MLP model is 0.84;the study area is divided into three regions which are high sensitivity,medium sensitivity and low sensitivity on the basis of the forecast results of the model and it is consistent with the historical landslide data.Practice has proved that the CF-MLP model has solved the quantification problem of the landslide disaster pregnant environmental factors,and is beneficial to build a more accurate model for landslide susceptibility analysis by improving the convergence effect of the multiple layer perceptron networks.