A Framework for Modeling Flood Depth Using a Hybrid of Hydraulics and Machine Learning

Area of Science:

• Environmental Engineering
• Computational Fluid Dynamics
• Water Resources Management

Background:

• River engineering requires accurate flow characterization (depth, velocity, extent) for problem-solving.
• Traditional hydraulic models are computationally expensive and infeasible for large-scale, high-resolution simulations.
• Machine Learning (ML) offers potential for efficient prediction in water resources, learning from data to forecast new scenarios.

Purpose of the Study:

• To present an efficient flood simulation framework for large-scale applications.
• To develop a novel, quick, and versatile model combining hydraulic and ML methods for flood mapping and depth prediction.
• To reduce computational time, resources, and expenses in complex hydraulic modeling.

Main Methods:

• A two-dimensional hydraulic model (iRIC), calibrated with measured data, was used to generate training data.
• Two ML models were trained: a random forest (RF) classification model for wet/dry node identification and a multilayer perceptron (MLP) for river depth estimation.
• The models were trained using iRIC simulation results for arbitrary discharge scenarios.

Main Results:

• The RF classification model achieved an overall accuracy of 98.5% in identifying wet or dry nodes.
• The MLP model demonstrated a regression coefficient of 0.88 for river depth prediction.
• The hybrid framework successfully predicted river depth and identified flooded areas.

Conclusions:

• The developed framework efficiently couples hydraulic and ML models for large-scale flood simulations.
• This approach significantly reduces computational demands, making previously infeasible 2D/3D hydraulic modeling practical.
• The study provides a versatile tool for real-time, large-scale river flood prediction and management.