基于Logistic回归与随机森林的和龙市地质灾害易发性评价

Geological Disaster Susceptibility in Helong City Based on Logistic Regression and Random Forest

为了科学地进行地质灾害易发性分析,基于和龙市地质灾害调查与区划,结合地质灾害分布规律和影响因素,综合考虑地形、地质、气象水文、土壤植被和人类工程活动5个因素,结合GIS技术与方法,提取高程、坡度、坡向、曲率、岩性、与断层距离、降雨量、与水系距离、NDVI(归一化植被指数)、土壤质地、水力侵蚀、人口密度、与道路距离等13个致灾因子,利用Logistic回归与随机森林模型进行地质灾害易发性评价,获得了区域地质灾害易发性分区图。其中,随机森林模型的结果表明:极低易发区面积占比最高,达56.98%,多位于研究区南部,高、极高易发区面积占比共12.89%,位于中部与东北部,该区是地质灾害预防和管理的关键区域;NDVI、高程、人口密度和降雨量等是灾害发育主要影响因素,累积贡献率为58.12%。Logistic回归和随机森林模型的ROC(受试者工作特征)曲线、已有灾害密度统计等验证结果均表明易发性分析和分区与实际灾害分布情况吻合度高,其AUC(ROC曲线下面积)值分别为0.856和0.907,说明两种方法均可实现有效预测,均具有较好的适用性,而随机森林模型表现出更高的准确性与稳定性,预测性能优于Logistic回归模型。

In order to scientifically analyze the geological disaster susceptibility.This paper is based on the investigation and zoning of geological disasters in Helong City,and through the analysis of the distribution rules and influencing factors of geological disasters.Considering terrain,geology,meteorology,hydrology,soil vegetation and human engineering activities,combined with GIS technology and methods,13 disaster causing factors including elevation,slope,aspect,curvature,lithology,distance from fault,rainfall,distance from water system,NDVI,soil texture,water erosion degree,population density and distance from road are extracted.Logistic regression and random forest model were used to evaluate the susceptibility of geological disasters,and the susceptibility zoning map was drawn.The result of random forest model shows:The very low susceptibility area is highest,reaching 56.98%of the total area,which is mostly located in the south of the study area;And the high and very high susceptibility areas account for up to 12.89%,which are located in the central and northeast,it is the key area for geological disaster prevention and management.NDVI,elevation,population density and rainfall are the main factors affecting disaster development,with a cumulative contribution rate of 58.12%.The ROC(receiver operating characteristic)curves and existing disaster density statistics of Logistic regression and random forest models show that the susceptibility zoning map are highly consistent with the actual disaster distribution,and their AUC(area under ROC curve)values are 0.856 and 0.907,respectively,which can achieve effective prediction and have good applicability.However,random forest model shows higher accuracy and stability,and its prediction performance is better than Logistic regression model.