人民黄河

为筛选诱发泥石流的致灾因子，评估RF、GBDT、XGBoost三种机器学习算法对泥石流易发性进行预测的性能，给泥石流多发区地质灾害预测及防灾减灾工作提供参考，以青海省海南藏族自治州为研究区，基于历史泥石流灾害数据，依据初选的17个影响因子与泥石流灾害的皮尔逊相关系数筛选诱发泥石流发生的致灾因子，对筛余致灾因子进行分级并设置8种致灾因子组合，采用RF、GBDT、XGBoost三种机器学习算法对泥石流易发性进行预测，采用准确率、精确率、召回率、F1分数、ROC-AUC等指标评价预测效果。结果表明：1)与河道距离、高程、土壤可蚀性、地形湿度指数、年降雨量、归一化植被指数、高程变异系数、岩性、地形粗糙度、剖面曲率、曲率、坡度、高程变异系数、坡向是研究区诱发泥石流的致灾因子，其中高程、与河道距离、土壤可蚀性、地形湿度指数、年降雨量、归一化植被指数为主要致灾因子；2)采用RF、XGBD、XGBoost三种机器学习算法对泥石流易发性进行预测时，基于致灾因子组合C7(该组合不考虑坡向)的预测效果最佳，基于致灾因子组合C5(该组合不考虑曲率、高程变异系数、坡向)和C8(该组合考虑所有致灾因子)的预测效果也较好；3)基于致灾因子组合C7进行泥石流易发性预测时，3种算法的优劣排序为XGBoost、GBDT、RF。

In order to screen the disaster-causing factors that induce debris flows, evaluate the performance of the three machine learning algorithms of RF, GBDT and XGBoost, in predicting the susceptibility of debris flows, and provide references for geological disaster prediction and disaster prevention and mitigation in the areas prone to debris flows, taking Hainan Tibetan Autonomous Prefecture, Qinghai Province as the study area and based on historical debris flow disaster data, based on the initially selected 17 influencing factors and the Pearson correlation coefficient of debris flow disasters, the disaster-causing factors that induce debris flow were screened. The remaining disaster-causing factors were classified and 8 combinations of disaster-causing factors were set. The three machine learning algorithm of RF, GBDT, and XGBoost were used to predict the susceptibility of debris flow. The prediction effect was evaluated by indicators such as accuracy rate, precision rate, recall rate, F1 score, and ROC-AUC. The results show that a) distance from the river channel, elevation, soil erodibility, topographic moisture index, annual rainfall, normalized vegetation index, coefficient of variation of elevation, lithology, topographic roughness, profile curvature, curvature, slope, coefficient of variation of elevation, and aspect are the disaster-causing factors that induce debris flows in the study area. Among them, elevation, distance from the river channel, soil erodibility, topographic moisture index, annual rainfall, and normalized vegetation index are the main disaster-causing factors. b) When the three algorithms of RF, GBDT and XGBoost are used to predict the likelihood of debris flow, the prediction effect based on the disaster-causing factor combination C7(this combination does not consider slope direction) is the best, and the prediction effects based on the disaster-causing factor combinations C5(this combination does not consider curvature, coefficient of variation of elevation and slope direction) and C8(this combination considers all disaster-causing factors) are also good. c) When predicting the susceptibility of debris flows based on the disaster-causing factor combination C7, the ranking of the advantages and disadvantages of the three algorithms is XGBoost, GBDT and RF.