基于逻辑回归树和旋转森林模型的滑坡易发性评价

为进一步探索集成模型在滑坡易发性评价中的适用性,以陕西省汉中市汉台区为例,结合相关资料与野外调查圈定40处滑坡,通过地质类、水文类、人类工程活动类中选取12个影响因子构建逻辑回归树(logistic model tree, LMT)和旋转森林(rotation forest, ROF)模型,分别生成滑坡易发性分区图,采用ROC(receiver operating characteristic)曲线进行模型精度验证与比较。结果表明,研究区滑坡受地形地貌、平面曲率与岩土体类型影响最大;两种模型预测率均较高,易发性等级分区结果与历史滑坡位置分布趋势基本一致;ROF模型的训练集正确率和验证集预测率分别为77.4%和93.1%,高于LMT模型的75.5%和84.0%;ROF模型滑坡极高易发区频率比为6.52,多于LMT模型(2.07),可见ROF模型对研究区滑坡易发性更加敏感,预测结果可靠度高;本文ROF模型滑坡易发性分区结果可为后期研究区防灾减灾与土地规划提供依据。

Evaluation of Landslide Susceptibility Based on Logistic Model Tree and Rotation Forest Model

Hantai District, Hanzhong City, Shaanxi Province was used as the study area in order to further explore ensemble model in landslide susceptibility. 40 landslides were delineated through relevant data and field investigation. Logistic model tree (LMT) and rotation forest (ROF) model were constructed by selecting 12 impact factors from geology, hydrology, and human engineering activities, then landslide susceptibility maps were generated respectively. ROC curve was used to verify and compare the model accuracy. The results showed that landslides in study area are most affected by topography, plan curvature and type of rock and soil. The prediction rates of the two models are both high, and the results of classification are basically consistent with the distribution trend of the historical landslide locations. The accuracy of training set and prediction rate of validation set of ROF model were 77.4% and 93.1%, respectively, which were higher than that of LMT model (75.5% and 84.0%). ROF model has a frequency ratio of 6.52, which is higher than LMT model (2.07), indicating that ROF model is more sensitive to landslide susceptibility in study area, and the prediction results have high reliability. The results of landslide susceptibility zoning in the ROF model can provide a basis for disaster prevention and land planning in the future.