摘要
山洪危险性评估是灾害链式防控体系的核心环节。利用机器学习算法进行山洪建模及评估山洪风险时,单一模型普遍存在过拟合风险、泛化能力不足等缺陷。集成模型通过模型融合可提升预测性能和模型稳健性,但其建模机理与适用性亟待深入探索与验证。本研究提出一种基于堆叠集成学习的山洪危险性评估方法,以汶川地震重灾区为实证对象,系统构建了包含随机森林(Random Forest,RF)、逻辑回归(Logistic Regression,LR)、梯度提升决策树(Gradient Boosting Decision tree,GBDT)和多层感知机(Multilayer Perceptron,MLP)的单一模型与堆叠集成模型,验证集成模型在山洪危险性评估中的对比优势。主要研究发现:(1)堆叠集成模型的准确率(90.51%)与F1分数(90.37%)显著优于单一模型,召回率与精确率差异仅1.33%,展现出更优的均衡性与综合性能。(2)集成模型的AUC值达96.87%,较RF、LR、GBDT和MLP分别提升1.55%、2.75%、2.13%和1.00%,验证了集成策略对预测精度的提升效果。(3)通过特征重要性分析揭示,距水系距离(0.047)、归一化植被指数(0.029)、坡度(0.028)构成山洪灾害的关键驱动因子,水区高差(0.027)、高程(0.017)及不透水面密度(0.016)次之。(4)在汶川地震扰动区,山洪极高与高危险区集中分布于东部低洼地带与西部高海拔河流沿岸。本研究可为地震灾区山洪风险识别提供创新方法,其构建的集成模型框架可为复杂地形区灾害评估研究提供范式参考。
Risk assessment of flash flood is a prerequisite of flood-related geo-disaster chain prevention and control.Although machine learning algorithms were introduced for flash flood modeling to assess their susceptibility with high efficiency,but a single model generally suffer from the defects of overfitting or insufficient generalization capacity.Some integrated model improve the prediction performance and model robustness through model fusion,but its modeling mechanism and applicability need to be deeply explored and verified.In this study,it proposed an innovative approach to flash flood risk assessment based on Stacking algorithm,and took an empirical study of the post-shock area of the Wenchuan earthquake in 2008,china.It constructed four single machine learning models(Random Forest,RF;Logistic Regression,LR;Gradient Boosting Decision Tree,GBDT;Multilayer Perceptron,MLP)and a Stacked Ensemble(SE)model which evolved from a fusion of the four machine learning algorithms.Then it systematically validated the comparative advantages of the SE model in the process of flash flood susceptibility assessment.(1)The stacked ensemble model significantly outperformed single models in terms of accuracy(90.51%)and F1 score(90.37%),with only a 1.33%difference between recall and precision rates,demonstrating superior balance and overall performance.(2)The AUC(Area Under the Curve)value obtained in the SE model reached 96.87%,an improvement of 1.55%,2.75%,2.13%,and 1.00%compared to RF,LR,GBDT,and MLP,respectively,verifying the effectiveness of the ensemble approach in enhancing prediction accuracy.(3)Feature importance analysis revealed that distance to water systems(0.047),Normalized Difference Vegetation Index(NDVI)(0.029),and slope(0.028)were the key driving factors for flash flood occurrences,followed by water area elevation difference(0.027),elevation(0.017),and impervious surface density(0.016).(4)In the post-shock area of Wenchuan earthquake,the extremely high and high-risk of flash flood zones were concentrated in the eastern low-lying areas and along high-altitude rivers in the west.This study provides an innovative approach to flash flood risk identification in earthquake-affected areas,and the ensemble model framework constructed can serve as a paradigm reference for disaster assessment research in complex terrain areas.
作者
贾倩
熊俊楠
尚依炜
肖慧文
王启盛
JIA Qian;XIONG Junnan;SHANG Yiwei;XIAO Huiwen;WANG Qisheng(College of Geosciences and Technology,Southwest Petroleum University,Chengdu 610500,China;School of Civil Engineering and Geomatics,Southwest Petroleum University,Chengdu 610500,China;Sichuan Water Development Investigation,Design&Research Co.,Ltd.,Chengdu 610500,China;The 4th Geological Brigade of Sichuan,Chengdu 611130,China)
出处
《山地学报》
北大核心
2025年第2期321-334,共14页
Mountain Research
基金
国家重点研发计划专项课题(2023YFC3006701)
四川省科技厅重点研发项目(2024YFHZ0134)。
关键词
山洪
集成学习
汶川地震扰动区
flash flood
ensemble learning
the post-shock area of the Wenchuan earthquake
