基于信息量模型和机器学习方法的滑坡易发性评价研究——以四川理县为例

以阿坝藏族羌族自治州地质灾害频发的理县为研究区，从地形地貌、地质环境、水文条件和人类工程活动等方面选取11个影响因子，通过皮尔森相关系数研究各因子之间的相关性，从而构建滑坡易发性评价指标体系。利用信息量模型计算各影响因子的信息量值，从信息量模型得出的极低和低易发性分区中选取非滑坡样本，在此基础上将样本数据代入随机森林和径向基函数神经网络2种机器学习模型开展滑坡易发性评价，并通过接收灵敏度（Receiver Operating Characteristic，ROC）曲线进行精度验证。结果显示：随机森林模型预测出的高易发区单位面积内分布的滑坡点数量更为集中，在仅占6.666%的区域分布了74.026%的灾害点，评价结果优于径向基函数神经网络模型。ROC曲线中两模型AUC（Area Under Curve）值分别为0.893、0.874，说明随机森林模型具有更高的可靠性，比径向基函数神经网络在该区域地质灾害易发性评价中更具优势。

Landslide Susceptibility Evaluation Based on Information Value Model and Machine Learning Method: A Case Study of Lixian County,Sichuan Province

Because there are many mountainous areas and complex terrain in China, geological disasters are widely distributed. Frequent geological disasters not only cause great damage to the ecological environment, but also seriously threaten the life and property safety of local residents. Therefore, it is of great significance to evaluate the landslide disaster susceptibility by zoning. Taking Lixian County, which is prone to frequent geological disasters, as the study area, 11 affecting factors are selected from four aspects: landform, geological environment, hydrological conditions and human engineering activities. The correlation between each factor is studied and the landslide susceptibility evaluation index system is constructed through Pearson Correlation Coefficient. According to the number of landslides, the Information Value model is used to calculate the information value of each affecting factor and select the non-landslide samples. On this basis, two machine learning models, the method of Random Forest and Radial Basis Function Neural Network model, are applied to carry out landslide susceptibility evaluation and the accuracy of landslide hazard susceptibility evaluation results is verified by ROC curve. The results show that the area of high-prone areas and extremely high-prone areas predicted by the Random Forest model is small, but the number of disaster points is more and the disaster plot ratio is higher. The area of high-prone areas and extremely high-prone areas predicted by Neural Network model is larger than that predicted by Random Forest model. In the predicted results of Random Forest evaluation model, the distribution of high-prone areas and extremely high-prone areas is more concentrated, and there are more disaster points distributed per unit area. By verifying the accuracy of the two models and through the comparison of ROC curve, the AUC values of Random Forest Algorithm and Radial Basis Function Neural Network model are greater than 0.8. The AUC value of Random Forest model is higher than that of Radial Basis Function Neural Network model, which is more suitable for landslide susceptibility evaluation in this area. Using the existing landslide statistical data to quantitatively study the landslide susceptibility can effectively improve the prediction accuracy, explore the optimal prediction model and provide a scientific basis for the prevention and management of geological disasters in this area.