Rapid prediction of flood velocity in steep rivers using machine learning and DEM-derived river geometry descriptors

ABSTRACT

Schematic overview of the proposed machine learning framework for rapid flood-velocity prediction. High-resolution digital elevation models (DEMs) are used to derive river geometry descriptors, while HEC-RAS 2D simulations provide flood-velocity data for multiple flood scenarios. These geomorphological and hydraulic parameters are combined into a dataset used to train and test the ensemble learning models, including AdaBoost, gradient boosting, random forest, and XGBoost. The trained models learn the relationship between river geometry descriptors and HEC-RAS flood velocity and are subsequently applied to predict flood velocities along river networks without requiring full hydrodynamic simulations. The framework demonstrates how DEM-derived river geometry descriptors and machine learning algorithms can be integrated to provide computationally efficient flood-velocity estimates for unseen river systems.

Two-dimensional (2D) hydrodynamic models are widely used for flood simulation; however, their computational requirements can limit their application for rapid flood analysis. This study evaluates four ensemble machine learning (ML) algorithms, such as XGBoost, gradient boosting, random forest, and AdaBoost, for predicting flood velocity across multiple flood scenarios. The models utilize river geometry descriptors derived from high-resolution digital elevation models (DEMs) and HEC-RAS simulations from 13 steep Norwegian rivers. Gradient boosting achieved the highest accuracy in reproducing the training dataset, with R2 values ranging from 0.91 to 0.93. However, XGBoost showed the best performance on the unseen test dataset, achieving R2 values of 0.20–0.60 and demonstrating relatively good transferability across river systems and flood scenarios. Model performance generally improved with increasing flood magnitude in the test river case. The results also indicate that hyperparameter optimization can improve model transferability. In the test dataset, the ML models were approximately 4.5*105 times faster than the corresponding HEC-RAS simulations. The result of this study highlights the potential of ensemble learning models for rapid flood velocity estimation.