摘要
崩塌、滑坡等形成的松散物质常为山洪灾害提供重要泥沙物源。汶川“5·12”地震灾区存在大量不稳定边坡与潜在的滑坡、崩塌风险区。构建崩塌、滑坡易发性评价模型对于该区域复合山洪灾害的早期防范具有重要意义。本文从地形、地质及气象水文等方面筛选出10个评价因子,应用极端梯度提升(XGBoost)与轻量级梯度提升机(LightGBM)两种先进的集成学习算法和逻辑回归、随机森林两种常见算法分别构建汶川县崩塌滑坡易发性评价模型,通过准确率、精确率、受试者工作特征曲线(ROC)下面积等定量指标对比各模型评估结果。结果表明:根据不同分类评价指标,两种集成学习模型相较于传统模型拥有更高的分类预测能力;分类准确率方面,XGBoost模型(0.903)与LightGBM模型(0.903)优于随机森林模型(0.900)与逻辑回归模型(0.864);精确率方面,LightGBM模型(0.887)略优于XGBoost模型(0.882),优于随机森林模型(0.872)与逻辑回归模型(0.802);根据不同模型ROC曲线下面积计算结果,XGBoost模型(0.904)与LightGBM模型(0.904)具有近乎同等的分类性能,略优于随机森林模型(0.902),逻辑回归模型最差(0.869);对易发性图进一步对比分析发现,两种集成学习模型的易发性分区结果与逻辑回归、随机森林模型结果存在一定差异,根据对各分区崩滑点密集程度的计算,两种集成学习模型的结果较为可靠,LightGBM模型在识别和预测崩滑高易发区域方面的性能最佳。
Objective The 5·12 Wenchuan earthquake triggered extensive secondary geological disasters and cascading effects.Wenchuan County,which was severely impacted by the earthquake,exhibits widespread unstable slopes and areas prone to landslides and collapses.In mountainous regions,the occurrence of extreme rainfall events precipitates extensive landslides and collapses.The copious loose material produced constitutes a substantial sediment source,exacerbating the magnitude of flash flood disasters under the coupling effect of water and sediment movement,and particularly heightening the risk of debris flows and debris floods.Given these circumstances,it is imperative to develop assessment models for landslide and collapse susceptibility to facilitate early prevention of compound flash flood disasters in Wenchuan County.Conventional susceptibility assessment approaches often rely on expert experience and subjective judgment;alternatively,they encounter difficulties in adequately fitting high-dimensional complex data.As a result,the precise delineation of the actual spatial distribution of areas susceptible to landslides and collapses remains a formidable challenge.Recent advancements in data science and machine learning provide promising solutions.Two state-of-theart ensemble learning algorithms,eXtreme Gradient Boosting(XGBoost)and Light Gradient Boosting Machine(LightGBM),are introduced to formulate dependable models for appraising susceptibility to landslides and collapses within the confines of Wenchuan County.Methods A comprehensive evaluation of factors related to topography,geology,meteorology,and hydrology was conducted to select ten evaluative factors:Elevation,slope,aspect,terrain relief,distance to rivers,distance to faults,normalized difference vegetation index(NDVI),land cover type,average annual precipitation,and lithology.Data preprocessing procedures were implemented to ensure the effectiveness and stability of model training.The data were standardized to mitigate the impact of differing scales among the dependent factors on the model.Factors displaying significant multicollinearity were identified and excluded using the Variance Inflation Factor(VIF),ensuring the independence of each feature in the analysis.In addition,the Information Gain Ratio(InGR)was utilized as a metric to evaluate the importance of each factor,facilitating the preliminary selection of explanatory variables.Then,two advanced ensemble learning algorithms(XGBoost and LightGBM)were applied alongside two traditional algorithms(logistic regression and random forest)to construct landslide and collapse susceptibility assessment models for Wenchuan County.Quantitative metrics,including accuracy,precision,recall,F_(1) score,and receiver operating characteristic(ROC)curves,were employed to enable a comparative and evaluative analysis of the performance of each model.These models were then utilized to predict the probabilities of landslide and collapse occurrences across the designated study area.The natural breakpoint method was employed to demarcate susceptibility zones,resulting in the development of a map delineating areas vulnerable to landslides and collapses.Additional qualitative and quantitative analyses were performed on the resulting susceptibility maps,with particular attention given to the correspondence between predicted results and actual landslide and collapse events,evaluating the predictive reliability of the proposed models.Results and Discussions The results indicated that both ensemble learning models demonstrated superior classification prediction capabilities when compared to traditional models.XGBoost and LightGBM achieved accuracies of 0.903,surpassing random forest(0.900)and logistic regression(0.864).In terms of precision,LightGBM(0.887)slightly outperformed XGBoost(0.882),while both outperformed random forest(0.872)and logistic regression(0.802).The F_(1) score metric placed XGBoost at the forefront with 0.899,closely followed by LightGBM(0.898)and random forest(0.897),while logistic regression yielded the lowest F_(1) score(0.866).Evaluation of the area under the ROC curve(AUC)indicated that XGBoost and LightGBM achieved nearly identical high classification performance(0.904),outperforming random forest(0.902),with logistic regression trailing at the lowest AUC(0.869).The examination of the constructed susceptibility zoning maps,coupled with quantitative analysis of the area proportions attributed to each zone,disclosed disparities in the partitioning outcomes from the XGBoost and LightGBM models in comparison to those produced by logistic regression and random forest models.These disparities were primarily attributed to the divergent data processing strategies inherent to each algorithm.In an effort to substantiate the reliability of the models’predictions,the density of landslide and collapse points within each susceptibility zone was quantitatively scrutinized.XGBoost,LightGBM,and random forest models consistently reflected the general trend of increasing landslide and collapse point density with higher susceptibility levels,aligning with the typical pattern of disaster susceptibility.LightGBM performed best in identifying high and extremely high susceptibility areas,with landslide and collapse point density ratios of 1.844 and 3.079,respectively,the highest among all models evaluated.In contrast,logistic regression did not adhere to this increasing trend,presenting an anomalous ratio of 0.588 in zones of very low susceptibility,a figure surpassing that within zones of high susceptibility(0.528).This anomaly indicated the presence of prediction bias in the logistic regression model,potentially ascribable to the limitations of the logistic regression algorithm and the lack of representative data.Conclusions The predictive capabilities of the advanced ensemble learning models in assessing landslide and collapse susceptibility in Wenchuan County surpassed those of the two traditional models.These models outperformed the traditional approaches in terms of accuracy,precision,F_(1) score,and area under the Receiver Operating Characteristic.LightGBM demonstrated higher precision,while XGBoost yielded superior results in the F_(1) score.In terms of reliability,both ensemble learning models,particularly LightGBM,exhibited advantages in identifying high and very high susceptibility areas,reinforcing their superiority in landslide and collapse susceptibility assessment.The research findings provide a more accurate tool for evaluating landslide and collapse susceptibility in Wenchuan County and similar areas affected by earthquakes,supporting the development of disaster prevention and mitigation measures.Future research can involve more comprehensive data collection methods and investigate broader applications of ensemble learning models,improving the reliability and practical implementation of predictions in disaster management.
作者
丁嘉伟
王协康
DING Jiawei;WANG Xiekang(State Key Laboratory of Hydraulics and Mountain River Engineering,Sichuan University,Chengdu 610065,China)
出处
《工程科学与技术》
北大核心
2025年第4期52-61,共10页
Advanced Engineering Sciences
基金
国家自然科学基金委员会-中华人民共和国水利部-中国长江三峡集团有限公司长江水科学研究联合基金项目(U2340201)
国家自然科学基金重点项目(52239006)
四川省自然科学基金项目(2024NSFSC0005)。
关键词
汶川震损区
滑坡
崩塌
山洪灾害
机器学习
Wenchuan earthquake-damaged area
landslide
collapse
flash flood disaster
machine learning
