基于SMOTETomek和机器学习的粤东山区边坡崩滑地质灾害危险性评价研究

1.广东省地质环境监测总站,广州 510510;2.华南理工大学,广州 510641

地质灾害;危险性评价;SMOTETomek;机器学习;深度学习

Risk Assessment of Slope Collapse Hazards in Mountainous Areas of Eastern Guangdong Based on SMOTETomek and Machine Learning
LI Jinxiang1,2,LIAO Jiale2,SU Wenji2

1.Guangdong Geological Environment Monitoring Station , Guangzhou 510510, China;2.South China Universi⁃ty of Technology , Guangzhou 510641, China

Geological disaster;Risk assessment;SMOTETomek;Machine learning;Deep learning

DOI: 10.13512/j.hndz.2025.02.07

备注

为开展粤东山区边坡崩滑地质灾害危险性评价研究,以梅西镇作为代表性研究区,选取坡度、坡向、平面曲率、剖面曲率、地层岩性、断裂距、水系距、NDVI、建筑物距、道路距、土地利用类型等构建评价指标体系,采用SMOTETomek混合采样结合随机森林(RF)、梯度提升决策树(GBDT)、3D卷积神经网络(3DCNN)、图注意力网络(GAT)等机器学习、深度学习模型开展易发性评价,并在最优模型的基础上叠加降雨量分析,完成危险性评价。结果得出:基于SMOTETomek混合采样的模型精度普遍更高,其中SMOTETomek-GAT模型精度最高,AUC值为0.93。
To assess the geological disaster hazard of slope collapse in mountainous areas of eastern Guangdong, this study selected Meixi Town as the representative research area. An evaluation index system was constructed com⁃prising the following factors:slope,aspect,plan curvature,profile curvature,stratigraphic lithology,distance from fault,distance from water system,normalized difference vegetation index(NDVI),distance from building, distance from road,and land use types. Then,a susceptibility evaluation was conducted by SMOTETomek hybrid sampling integrated with machine learning and deep learning models, including random forest(RF), gradient boosting decision tree(GBDT),3D convolutional neural network(3D CNN),and graph attention network(GAT).Based on the optimal model, precipitation analysis was incorporated to complete the risk assessment. The results show that models based on SMOTETomek hybrid sampling generally achieve higher accuracy, and the SMOTE⁃Tomek-GAT model has the highest precision,with an AUC value of 0.93.
·