基于堆叠模型的滑坡易发性评价——以商洛市丹凤县为例

Evaluation of landslide susceptibility based on Stacking:Take Danfeng County,Shangluo City as an example

陕南地区滑坡灾害频发,该地区进行有效的滑坡易发性评价对防灾减灾意义重大。文中以陕西省商洛市丹凤县257个滑坡灾点数据为评价样本,选取高程、坡度、坡向、剖面曲率、平面曲率、地形湿度指数(TWI)、植被归一化利用指数(NDVI)、距道路距离、距断层距离和年均降雨量共10个影响因子进行评价。通过相关属性评估分析,与研究区滑坡发生最为密切的是高程、TWI和距道路距离。基于GIS随机提取70%栅格数据作为样本,采用Stacking模型集成RA、DECORATE和RS模型,在WEKA软件中分别进行训练,对剩余30%栅格数据进行验证,采用AUC值进行模型性能比较,最终分别生成各单独分类器与堆叠后的滑坡易发性图。结果表明:4种模型AUC值均大于0.7,训练准确率和验证预测率均较高,适合研究区滑坡易发性评价。其中,Stacking模型的预测能力最高,说明不同基分类器组合提升了堆叠模型的拟合能力,增强了对滑坡易发性的辨识度。

Landslide disasters occur frequently in southern Shaanxi.Effective landslide susceptibility evaluation in this area is of great significance for disaster prevention and reduction.Based on the data of 257 landslide disaster points in Danfeng County,Shangluo City,Shaanxi Province,10 influencing factors such as elevation,slope angle,slope aspect,profile curvature,plan curvature,TWI,NDVI,distance to roads,distance to faults and average annual rainfall were selected for evaluation.Through the evaluation and analysis of relevant attributes,the most closely related to the occurrence of landslide in the study area are elevation,TWI and distance to roads.Based on GIS,70%of the grid data were randomly extracted as samples,and the Stacking model was used to integrate RA,DECORATE and RS models.They were trained respectively in WEKA software to verify the remaining 30%of the grid data,and the AUC value was used to compare the model performance.Finally,the landslide susceptibility maps of each separate classifier and stacked were generated respectively.The results showed that the AUC values of the four models are greater than 0.7 and the training accuracy and verification prediction rate are high.Four models are suitable for the evaluation of landslide susceptibility in the study area.Among them,the Stacking model has the highest prediction ability,indicating that the combination of different base classifiers improves the fitting ability of the Stacking model and enhances the identification of landslide susceptibility.