基于逻辑回归的巫山县滑坡易发性区划研究

Landslide Susceptibility Mapping Based on Logistic Regression in Wushan County

【目的】在已有文献研究基础上,构建基于逻辑回归的滑坡易发性评价模型,并与基于随机森林的模型进行分析比较,探讨出适用于三峡库区巫山县的滑坡易发性评价模型。【方法】选取地质条件、地形地貌、环境条件、人类工程活动等4方面影响下的22个因子作为滑坡易发性影响因子,根据研究区963处历史滑坡点,建立30m×30m栅格地理空间数据库;进一步在数据库中,按历史滑坡与非滑坡1∶10的比例构建地理空间数据样本集,基于十折交叉验证法选择最佳样本数据后,进行逻辑回归模型训练;最后,利用最佳样本训练后的模型,进行全域滑坡易发性仿真分析,并将结果划分为低、较低、中、较高、高共5种易发性等级。【结果】逻辑回归模型训练集、测试集的受试者工作特征曲线下面积(Area under curve,AUC)值分别为0.803和0.787;有63.24%历史滑坡落入面积比为26.26%的较高滑坡易发性区划中。【结论】基于逻辑回归和随机森林构建的滑坡易发性评价模型均具有良好的稳定性及预测功能,但随机森林模型性能优于逻辑回归模型性能,更适用于三峡库区滑坡易发性区划研究。

[Purposes]Based on previous studies,it constructed a landslide susceptibility mapping model on account of logistic regression,analyzed and compared its performance with the random forest model,and explored the most suitable landslide susceptibility mapping model in Wushan county.[Methods]22 factors under the influence of geological conditions,geomorphology,environmental conditions,and human engineering activities were selected as the conditioning factors of landslide susceptibility,and combined with 963 historical landslide data in the study area,ageospatial database was constructed.Then,in the geospatial database,10-fold cross-validation were used to train all the datasets(including landslide points and non-landslide points).The best optimized model was chose to simulate the landslide susceptibility in the study area.The results were divided into five classes,that is,very low susceptibility,low susceptibility,moderate susceptibility,high susceptibility and very high susceptibility region.[Findings]The area under the curve(AUC)value of receiver operating characteristic(ROC)curve showed that the prediction accuracy of training dataset,validation dataset were 0.803 and 0.787.63.24% of the historic landslides occurred in the high-very areas.Elevation,annual average rainfall and POI kernel density did more for the occurring of landslide.[Conclusions]The landslide susceptibility mapping models were constructed on account of logistic regression and random forest both have great stability and predictive functions,but the performance of the random forest is better than that of the logistic regression,which is more suitable for the study of landslide susceptibility mapping in the Three Gorges reservoir area.