基于信息量–随机森林模型的邛崃山系滑坡易发性评价
Evaluation of Landslide Susceptibility of the Qionglai Mountain Range Based on the Information-Entropy Random Forest Model

针对目前滑坡易发性评价中非灾害点赖于主观推测或随机选取，导致非灾害点辨识度较低，容易导致分析误差增大，进而影响模型的精度的问题，本文采用基于信息量法支持下的随机森林模型对邛崃山系进行滑坡易发性评价。经过大量统计研究与地质环境条件分析，选取高程、坡度、坡向、归一化植被指数、降雨量、岩组、距断层距离、距道路距离和距水系距离9个影响滑坡的因素进行分析，得出各评价指标权重。利用信息量模型进行易发性预评价和分级，从低易发区随机选取2373个非灾害点作为负样本点。然后通过随机森林模型获取滑坡易发性评价结果，并利用ArcGIS中的自然断点法对邛崃山系滑坡易发性结果进行分类，将研究区域划分为极低易发区(29.4%)、低易发区(30.5%)、中等易发区(20.6%)、高易发区(13.4%)、极高易发区(7.1%)五个等级区。并通过ROC曲线检验评价精度，结果显示模型AUC值为0.940，验证了信息量支持下的随机森林模型在邛崃山系滑坡易发性评价中具有较高的可信度。
In the current evaluation of landslide susceptibility, the identification of non-landslide points is often based on subjective estimation or random selection. This leads to low discernibility of non-landslide points, increasing analytical errors and affecting model accuracy. A random forest model supported by the information value method was adopted to evaluate landslide susceptibility in the Qionglai mountain system. Extensive statistical research and analysis of geological and environmental conditions were conducted. Nine factors affecting landslides were selected: elevation, slope, aspect, normalized vegetation index, rainfall, lithology, distance to fault, distance to road, and distance to water system. The weights of each evaluation index were determined. The information value model was used for preliminary susceptibility evaluation and classification. From low susceptibility areas, 2373 non-landslide points were randomly selected as negative sample points. Landslide susceptibility results were obtained using the random forest model. The natural break method in ArcGIS was employed to classify the results. The study area was divided into five susceptibility zones: very low susceptibility (29.4%), low susceptibility (30.5%), moderate susceptibility (20.6%), high susceptibility (13.4%), and very high susceptibility (7.1%). The evaluation accuracy was tested using the ROC curve. The results showed an AUC value of 0.940, verifying that the random forest model supported by the information value method has high reliability in evaluating landslide susceptibility in the Qionglai mountain system.