基于逻辑回归的地震滑坡易发性评价——以汶川地震、鲁甸地震为例

准确评估地震诱发的滑坡风险,并及时绘制滑坡易发风险图是灾害应急救援的科学前提和理论基础。目前机器学习在滑坡敏感性评估中具有广泛应用,但大多数研究缺乏对模型的普适性探讨,且该类预测模型缺乏定量评价地震动参数对模型精度的影响。该文以2008年5月12日的汶川8级地震和2014年8月3日的鲁甸6.5级地震为例,先通过相关系数及方差膨胀因子选择地震滑坡的影响因子构建数据库,并随机按照7∶3的比例分为训练集和测试集,再分析影响因子在滑坡和非滑坡样本中的频数分布,最后分别利用两次地震的训练集建立逻辑回归模型(Logistic Regression, LR)进行精度验证和易发性评估。结果显示模型在同一次地震的测试集下均达到较高的预测精度(>90%);但是基于汶川地震构建的模型对鲁甸地震诱发滑坡的预测精度整体下降了14%。此外,地震动参数(Modified Mercalli Intensity Scale, MMI)对模型预测精度贡献在5%~29%。结果表明基于历史地震事件建立的模型对未来地震引发滑坡的预测中仍具有较大的局限性,需要增加不同地区不同震情的样本量和新的机器学习方法提高预测模型的普适性。

Assessing Earthquake-induced Landslide Susceptibility based on Logistic Regression in 2008 Wenchuan Earthquake and 2014 Ludian Earthquake

Susceptibility mapping of earthquake triggered landslides is the scientific premise and theoretical basis of emergency rescue.Currently,machine learning has been widely applied in the assessment of earthquake triggered landslide susceptibility.However,previous studies have not considered universality of model.Furthermore,these studies lack the quantitative evaluation of the effect of ground motion parameters on the accuracy of the model.We took the May 12,2008 Wenchuan earthquake and the August 3,2014 Ludian earthquake as study cases.We constructed the database including the landslide inventory and the influencing factors which are selected using correlation coefficient and variance inflation factor.In this study,70%of the landslides are randomly chosen for training the logistic regression model and 30%for testing.The frequency distribution of landslide and non-landslide samples is analyzed for different variables.Then the model performance is evaluated and the susceptibility map is produced.The results show that the model achieves higher prediction accuracy(>90%)using the test samples in the same earthquake event.However,when we applied the model constructed using the training dataset of Wenchuan earthquake to the testing data of Ludian earthquake event,the prediction accuracy of declined by 14%.In addition,we found that the ground motion parameter(MMI)contributes 5%to 29%to the model prediction accuracy.The results show that the model based on historical earthquake events still has relatively limitations in the prediction of future earthquake-induced landslides.It is necessary to increase the sample size of different earthquakes in various regions and use new machine learning methods to improve the universality of prediction models.