基于信息量与机器学习耦合模型的滑坡易发性评价对比分析  

Comparative analysis of landslide susceptibility evaluation based on the coupling model of information quantity and machine learning

作　　者：吴雅睿[1] 娄春辉 侯龙君 刘峰[1] WU Yarui;LOU Chunhui;HOU Longjun;LIU Feng(College of Geomatics,Xi’an University of Science and Technology,Xi’an 710054,China;Northwest Nonferrous Survey Engineering Co.,Ltd.,Xi’an 710038,China)

机构地区：[1]西安科技大学测绘科学与技术学院,陕西西安710054 [2]西北有色勘测工程有限责任公司,陕西西安710038

出　　处：《西安科技大学学报》2024年第6期1140-1153,共14页Journal of Xi’an University of Science and Technology

基　　金：国家自然科学基金项目(41977059);陕西省软科学研究计划项目(2022KRM034);西北有色地质矿业集团科技项目(XBDKKJ202211)。

摘　　要：陕西省商洛市镇安县地质环境复杂,滑坡灾害频发。以镇安县为研究区,提出将信息量模型分别与4种机器学习方法(SVM、RBF、ELM、BPNN)相结合,利用信息量模型的定量化和机器学习方法的非线性拟合能力,构建耦合模型进行滑坡易发性评价对比分析。首先,基于县域地质灾害调查数据,从地形地貌、地质环境、气象水文和人类工程活动等方面选取9个影响因子,通过皮尔森相关系数分析各因子之间的相关性,构建滑坡易发性评价指标体系;然后,利用信息量模型量化影响因子,以各影响因子的信息量值作为样本输入数据,代入支持向量机、极限学习机、径向基函数网络和BP神经网络4种机器学习模型开展滑坡易发性评价,并通过接收灵敏度(ROC)曲线进行精度验证。结果表明:区内滑坡主要分布于道路两侧、河流沿岸及山体破碎带等地质环境恶劣区域,其中人类工程活动、气象、地质构造是影响区内滑坡发育的主要因素;4种耦合模型中IV-BPNN模型更适用于研究区滑坡易发性评价,其预测的极高易发区和高易发区单位面积内分布的滑坡点数量更为集中,在仅占21.43%的区域分布了85.15%的滑坡灾害点,评价结果优于其他耦合模型;ROC曲线中IV-SVM、IV-RBF、IV-ELM和IV-BPNN模型的AUC值分别为0.841、0.813、0.838、0.863,其中IV-BPNN模型精度最高,与信息量模型的AUC值(0.722)相比提高了19.5%,具有更高的可靠性。研究提出的IV-BPNN模型可以更好地解决传统滑坡易发性评价方法中影响因子量纲不统一及因子间复杂非线性关系而导致评价结果适应性一般的问题,能更加精确地识别潜在滑坡灾害高风险区域,可为当地滑坡灾害防治工作提供参考依据。The geological environment of Zhen'an County,Shangluo City,Shaanxi Province is complex,and landslide disasters occur frequently.This study focused on Zhen'an County and proposed combining the Information Value Model(IVM) with four machine learning methods(Support Vector Machine,Radial Basis Function,Extreme Learning Machine,and Back Propagation Neural Network) to construct coupled models for comparative landslide susceptibility assessment.First,based on geological disaster survey data of the county,nine influencing factors were selected from aspects such as topography,geological environment,meteorology,hydrology,and human engineering activities.The Pearson correlation coefficient was used to analyze the correlations between these factors,forming a landslide susceptibility evaluation index system.Secondly,the Information Value Model was used to quantify these factors,with the information value of each factor serving as input data for the machine learning models.Landslide susceptibility assessments were then conducted using the four machine learning models(SVM,RBF,ELM,and BPNN),and accuracy was verified through Receiver Operating Characteristic(ROC) curves.The results indicate that:Landslides in the study area are mainly distributed along roadsides,riverbanks,and fractured mountain zones with poor geological conditions.Human engineering activities,meteorology,and geological structures are the main factors influencing landslide development in the area.Among the four coupling models,the IV-BPNN model is more suitable for evaluating the landslide susceptibility assessment in the study area.This model shows the number of landslide points distributed per unit area in extremely high and high-risk areas is more concentrated,with 85.15% of landslide disaster points distributed in only 21.43% of the area,outperforming the other coupling models.The AUC values for the IV-SVM,IV-RBF,IV-ELM,and IV-BPNN models are 0.841,0.813,0.838,and 0.863,respectively.Among them,the IV-BPNN model has the highest accuracy,with an increase of 1

关 键 词：滑坡 信息量模型 机器学习模型 神经网络模型 易发性评价 

分 类 号：P642[天文地球—工程地质学]