Flash-flood hazard assessment using ensembles and Bayesian-based machine learning models: Application of the...

Area of Science:

• Hydrology and Environmental Science
• Geospatial Analysis
• Natural Hazard Assessment

Background:

• Flash floods pose a significant global natural hazard, with Iran being particularly vulnerable.
• Existing temporal forecasting models are insufficient for spatial hazard assessment, crucial for mitigation and policy.
• There is a need for more accurate flash-flood hazard mapping models to support disaster risk reduction.

Purpose of the Study:

• To develop and evaluate advanced ensemble models for accurate flash-flood hazard mapping.
• To identify key environmental variables influencing flash-flood occurrence.
• To provide tools for effective watershed management and flood damage remediation.

Main Methods:

• Application of ensemble models: boosted generalized linear model (GLMBoost) and random forest (RF).
• Utilization of Bayesian generalized linear model (BayesGLM) for comparative analysis.
• Employing simulated annealing (SA) for feature selection to optimize model performance.

Main Results:

• Both GLMBoost and RF models demonstrated high accuracy (90-92%) and skill scores (Kappa: 79-84%).
• Key contributing variables identified include distance from stream, vegetation, drainage density, land use, and elevation.
• The models achieved excellent performance metrics: Success Ratio (94-96%) and Threat Score (80-84%).

Conclusions:

• The proposed ensemble models offer a significant advancement in flash-flood hazard mapping accuracy.
• These models are valuable tools for watershed managers, especially in data-scarce regions.
• The findings support improved adaptation, mitigation strategies, and disaster risk reduction for flash floods.