Scholarlink

<p>Floods and landslides are pervasive natural hazards that threaten lives, infrastructure, ecosystems, and economies. Although they often co-occur within the same region, flooding in low-lying river corridors and landslides on steep slopes are commonly assessed separately, which limits integrated risk management. To address this gap, this study developed advanced ML models, including Extreme Gradient Boosting (XGB), Bat Algorithm-XGB (BAT_XGB), and Whale Optimisation Algorithm-XGB (WOA-XGB), to build landslide and flood susceptibility maps. The historical flood and landslide database, along with topographical, hydrological, geological, and environmental factors, serves as input data for modelling each type of susceptibility map. Various standard quantitative metrics were utilized to assess the forecasting performance of ML models. The verification results show that the WOA_XGB model achieves the highest prediction performance (AUC = 0.967) for flood susceptibility modelling, and the BAT_XGB model achieves an AUC of 0.983 for the landslide susceptibility map. Finally, the multi-hazard susceptibility map was created in the GIS environment by combining the best landslide and flood susceptibility models based on the bivariate matrix approach. Analysis results on this map indicate that Dinh Hoa, Vo Nhai, Dai Tu, and Phu Luong districts frequently face dual hazards (landslides and floods) at extremely high levels. The proposed framework demonstrates the added value of metaheuristic-enhanced boosting models for improving multi-hazard forecasting and provides actionable spatial evidence to support regional planning, resource allocation, and climate change adaptation strategies in Thai Nguyen Province, Vietnam.</p> Graphical Abstract <p></p> <p>The graphical abstract illustrates the full workflow used to identify areas simultaneously susceptible to floods and landslides in Thai Nguyen province, Vietnam. The process begins with the compilation of historical flood and landslide inventory data, which provide essential evidence of past hazard occurrences and form the basis for model training and validation. In parallel, a comprehensive set of causative factors, including topographic, geological, hydrological, and environmental parameters, is assembled to represent the conditions influencing hazard formation. These factors are then examined for multicollinearity using Tolerance and Variance Inflation Factor (VIF) criteria to ensure that only the most relevant predictors are retained, thereby improving the robustness of the modelling process. The refined datasets are used to develop three machine learning models: a baseline Extreme Gradient Boosting (XGB) model and two hybrid metaheuristic-enhanced models, BAT_XGB and WOA_XGB. Model performance is evaluated through ROC curves and AUC metrics, revealing that WOA_XGB provides the highest accuracy for flood susceptibility mapping, while BAT_XGB performs best for landslide susceptibility. The outputs from these models are subsequently integrated within a GIS environment to generate a multi-hazard susceptibility map. The final panel highlights regions where flood and landslide risks overlap at extremely high levels. This multi-hazard approach provides a clearer understanding of compound hazard exposure and delivers actionable insights for local authorities, planners, and communities seeking to implement proactive disaster risk reduction and adaptation strategies in the face of increasing climate pressures.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Metaheuristic-Enhanced XGBoost Models for Improving Flood and Landslide Susceptibility Assessment

  • 期刊论文

  • 发表日期：

    2026年2月18日

查看全文
  • Quoc-Hung Vu,
  • Quynh Duy Bui,
  • Dinh Chieu Vu,
  • Ngoc-Dung Luong,
  • Cong-Hieu Duong,
  • Chinh Luu,
  • Hang Ha

摘要

Floods and landslides are pervasive natural hazards that threaten lives, infrastructure, ecosystems, and economies. Although they often co-occur within the same region, flooding in low-lying river corridors and landslides on steep slopes are commonly assessed separately, which limits integrated risk management. To address this gap, this study developed advanced ML models, including Extreme Gradient Boosting (XGB), Bat Algorithm-XGB (BAT_XGB), and Whale Optimisation Algorithm-XGB (WOA-XGB), to build landslide and flood susceptibility maps. The historical flood and landslide database, along with topographical, hydrological, geological, and environmental factors, serves as input data for modelling each type of susceptibility map. Various standard quantitative metrics were utilized to assess the forecasting performance of ML models. The verification results show that the WOA_XGB model achieves the highest prediction performance (AUC = 0.967) for flood susceptibility modelling, and the BAT_XGB model achieves an AUC of 0.983 for the landslide susceptibility map. Finally, the multi-hazard susceptibility map was created in the GIS environment by combining the best landslide and flood susceptibility models based on the bivariate matrix approach. Analysis results on this map indicate that Dinh Hoa, Vo Nhai, Dai Tu, and Phu Luong districts frequently face dual hazards (landslides and floods) at extremely high levels. The proposed framework demonstrates the added value of metaheuristic-enhanced boosting models for improving multi-hazard forecasting and provides actionable spatial evidence to support regional planning, resource allocation, and climate change adaptation strategies in Thai Nguyen Province, Vietnam.

Graphical Abstract

The graphical abstract illustrates the full workflow used to identify areas simultaneously susceptible to floods and landslides in Thai Nguyen province, Vietnam. The process begins with the compilation of historical flood and landslide inventory data, which provide essential evidence of past hazard occurrences and form the basis for model training and validation. In parallel, a comprehensive set of causative factors, including topographic, geological, hydrological, and environmental parameters, is assembled to represent the conditions influencing hazard formation. These factors are then examined for multicollinearity using Tolerance and Variance Inflation Factor (VIF) criteria to ensure that only the most relevant predictors are retained, thereby improving the robustness of the modelling process. The refined datasets are used to develop three machine learning models: a baseline Extreme Gradient Boosting (XGB) model and two hybrid metaheuristic-enhanced models, BAT_XGB and WOA_XGB. Model performance is evaluated through ROC curves and AUC metrics, revealing that WOA_XGB provides the highest accuracy for flood susceptibility mapping, while BAT_XGB performs best for landslide susceptibility. The outputs from these models are subsequently integrated within a GIS environment to generate a multi-hazard susceptibility map. The final panel highlights regions where flood and landslide risks overlap at extremely high levels. This multi-hazard approach provides a clearer understanding of compound hazard exposure and delivers actionable insights for local authorities, planners, and communities seeking to implement proactive disaster risk reduction and adaptation strategies in the face of increasing climate pressures.