基于机器学习模型与特征筛选的安徽省滑坡灾害易发性分布研究

Landslide Susceptibility Distribution in Anhui Province Based on Machine Learning Model and Feature Selection

ES评分 0

DOI 10.14075/j.jgg.2024.12.596
刊名
Journal of Geodesy and Geodynamics
年,卷(期) 2026, 46(1)
作者
作者单位

1 同济大学土木工程学院, 上海, 200092;
2 安徽省交通规划设计研究总院股份有限公司, 合肥, 230088

摘要
以安徽省历史滑坡与水文地质条件为原始数据,利用5种典型机器学习模型,选取9种环境致灾因子作为初始输入变量,针对机器学习输入变量筛选组合对不同模型的性能影响问题,采用多种因子筛选方法及数种评价指标,考察各筛选组合下每种算法模型对滑坡易发性的预测精度。评价结果显示,全选组与轻量梯度提升机(LightGBM)模型的组合具有最优的性能;在研究区的滑坡易发性研究中,对机器学习输入因子的筛选并不一定对模型的性能提升产生正面影响;集成模型比单评估器模型性能更优,且LightGBM由于采用了基于叶子节点的决策树算法,相较于其他集成算法模型性能有一定的提升。
Abstract
This paper uses historical landslide and hydrogeological data from Anhui province as the original dataset, employing five typical machine learning models and selects nine environmental disaster-inducing factors as initial input variables. To explore the impact of different machine learning input variable selection combinations on the performance of various models, multiple factor selection methods and several evaluation metrics are used to examine the prediction accuracy of each algorithm model under different selection combinations for landslide susceptibility. The evaluation results show that the combination of the full selection set and the lightweight gradient boosting machine (LightGBM). model yields the best performance. Therefore, this combination is chosen for the landslide susceptibility mapping in the study area. Moreover, the study suggests that factor selection for machine learning inputs does not necessarily improve model performance in landslide susceptibility studies of the research area. Additionally, ensemble models outperform individual estimator models, and LightGBM, using a leaf-node-based decision tree algorithm, shows improved performance compared to other ensemble algorithms.
关键词
滑坡; 机器学习; 特征筛选; 易发性; 集成模型
KeyWord
landslide; machine learning; feature selection; susceptibility; ensemble models
基金项目
页码 68-77
  • 参考文献
  • 相关文献
  • 引用本文

陆子康, 陈宏信, 冯世进, 赵勇, 沈国栋, 陈修和. 基于机器学习模型与特征筛选的安徽省滑坡灾害易发性分布研究 [J]. 大地测量与地球动力学. 2026; 46; (1). 68 - 77.

  • 文献评论

相关学者

相关机构