Combined application of numerical simulation and machine learning in debris flow hazard mapping.

Area of Science:

• Geosciences
• Computational Science
• Environmental Engineering

Background:

• Debris flow hazard mapping (DFHM) is crucial for mitigating debris flow risks.
• Traditional DFHM relies on time-consuming numerical simulations to determine debris flow intensity.

Purpose of the Study:

• To develop and evaluate a combined framework of numerical simulation and machine learning for efficient DFHM.
• To assess the speed and accuracy of machine learning models as alternatives to traditional hydraulic simulations in DFHM.

Main Methods:

• Utilized the FLO-2D model to simulate debris flows for 20, 50, and 100-year recurrence intervals.
• Collected debris flow depth and velocity as dependent variables, and topographic and land cover data as independent variables.
• Trained a gradient boosted decision tree (GBDT) machine learning model using 20 and 50-year datasets and validated it with the 100-year dataset.

Main Results:

• Machine learning predictions closely matched numerical simulation results, validating the proposed framework.
• The machine learning approach demonstrated over a tenfold increase in speed compared to traditional numerical simulations.
• The study confirmed the validity and rationality of integrating machine learning for enhanced DFHM efficiency.

Conclusions:

• Machine learning models show significant potential as efficient alternatives to hydraulic simulations for large-scale DFHM.
• The proposed combined framework offers a substantial improvement in the efficiency of debris flow hazard assessment.
• This research paves the way for faster and more scalable debris flow risk management strategies.