基于频率比-随机森林模型的滑坡易发性评价

本文以陕西省洋县为研究区,通过搜集资料、实地调查获得研究区滑坡分布状况。结合研究区地质环境特征与前人研究经验,初步选取海拔、坡度、坡向、地形起伏度、曲率、距水系距离、距道路距离、降雨量及岩土体类型,共九种滑坡影响因子展开滑坡易发性研究。首先,采用皮尔森相关系数法对各因子间的相关性进行分析。其次,按照70/30的比例将滑坡数据随机划分为模型训练集与模型验证集。然后,采用模型训练集对频率比模型(FR)、随机森林模型(RF)及两者的耦合模型(FR-RF)进行训练,利用模型验证集对模型训练结果进行检验,并绘制ROC曲线。最后,利用验证后的模型绘制研究区滑坡易发性分区图。结果表明：(1)所选取的9个滑坡影响因子是相互独立的；(2)本研究所采用的三个模型均表现良好,其中FR-RF模型预测准确度最高(0.901),其次为RF模型(0.863),最后为FR(0.833)；(3)本研究所绘制的滑坡易发性分区图可为当地政府制定土地利用规划、预防滑坡等方案提供参考借鉴。

Landslide Susceptibility Research Based on FR-RF Model :A Case study of Yang County, Shaanxi Province

This paper took Yang county, Shaanxi province as the study area and the distribution of landslides in the study area was obtained through field investigation. According to the geological environment characteristics of the study area and previous researches, nine landslide conditioning factors were selected, including elevation, slope angle, slope aspect, topographic relief, curvature, distance to rivers, distance to roads, rainfall and lithology, to carry out the landslide susceptibility research. First of all, the Pearson correlation coefficient method was introduced to analyze the correlation among factors. Next, dividing the landslides into training dataset and validation dataset with the ration of 70/30. In addition, Using training dataset to train to the frequency ratio(FR) model, random forest(RF) model and their ensemble model (FR-RF). Applying the validation dataset to verify the training results and drawing ROC curve. Finally, using the verified models to generated landslide susceptibility maps of study area. The results show that: (1) nine landslide conditioning factors are independent from each other. (2)three models adopted in this study all performed well, with FR-RF model obtained the highest prediction accuracy (0.901), followed by RF model (0.863) and FR (0.833). (3)the landslide susceptibility maps generated in this study can provided references for the local government to make land use planning and landslide prevention.