不同分级方法对区域滑坡易发性区划的影响

易发性分区是开展区域地质灾害风险评价的基础步骤，选取合理的分级方法对有效绘制区域滑坡易发性图意义显著，但鲜有研究对比了各易发性分级方法的优缺点，尤其是未能将历史滑坡与预测出的易发性指数相联接。针对该问题，以陕西省延长县为例采用3种机器学习模型计算滑坡易发性指数，即分类和回归树、随机森林和径向基函数；设计了5种易发性分级方法，划分不同的滑坡易发性等级，包括4种常规的基于地理信息系统的分级方法（自然断点、等间隔、分位数和几何间隔），同时考虑了滑坡与易发性指数间的非线性关联性的频率比阈值法。结果表明：3种模型的受试者工作特征（ROC）曲线下面积均大于0.75，但划分的易发性分级图分布模式却存在较大差异，使用几何间隔和分位数法的易发性图能在极高易发区中识别出更多滑坡，但这两种方法划分的极高和高易发区的总面积过大；使用等间隔法和频率比阈值法在极高和高易发区中的滑坡比率更大，说明识别出的滑坡更为集中。本文提出的频率比阈值法用于滑坡易发性分级，能为易发性的准确分区提供思路，为边坡稳定性较差区域的工程选址以及土地利用规划提供科学参考，提高地质安全评估及应急管理能力。

Effects of Different Classification Methods on Regional Landslide Susceptibility Zonation

Susceptibility zoning is the basic step of regional geological hazard risk assessment, and a reasonable classification method for the susceptibility levels is of significance for obtaining effective regional landslide susceptibility maps. However, few studies have compared the advantages and disadvantages of the susceptibility classification method, especially the failure to link landslide with the predicted susceptibility index. Yanchang County in Shaanxi Province was taken as the study area, and three machine learning models were applied to calculate the landslide susceptibility index of the region, which were classification and regression tree (C&RT), random forest (RF), and radial basis function (RBF) models. Then four GIS-based classification methods (natural breaks, equal interval, quantile, geometrical interval) were used to classify susceptibility levels and to generate landslide susceptibility maps. Aiming at the nonlinearity correlation between landslide inventory distribution and susceptibility index was not considered in the classification of susceptibility, a frequency ratio threshold method was proposed to classify susceptibility levels innovatively. The results showed that although the accuracies of the three models expressed by the receiver operating characteristic (ROC) curve were all greater than 0.75, large differences in landslide distribution patterns among different landslide susceptibility maps were observed. Different classification methods have a comparative analysis role in the final landslide susceptibility mapping. The landslide susceptibility maps using geometrical interval and quantile methods identified more landslides in very high susceptibility areas, while the total area of very high and high susceptibility areas was too high. Moreover, the density of landslides from equal interval and frequency ratio threshold methods was larger. That means the landslides identified are more concentrated. This paper innovatively proposed the frequency ratio threshold method for landslide susceptibility classification, which could provide a new idea for accurate zoning of susceptibility, provide scientific reference for project site selection and land use planning in areas with poor slope stability, and improve the ability of geological safety assessment and emergency management.