Author(s): Kexin Liu; Ryosuke Akoh; Shiro Maeno
Linked Author(s): Ryosuke Akoh, Shiro Maeno
Keywords: Flood inundation prediction augmented random forest model synthetic training dat
Abstract: The need for fast nowcasting of floods has been increasing during the past decades due to frequently occurring extreme events. The application of machine learning (ML) models is gradually gaining attention for the ability efficient predictions compared to conventional hydrodynamic (HD) models. Yet, the scarcity of observations on extreme flood as training data for ML models is still a major limitation. This study proposes an augmented random forest (RF) model using simulation results from HD models with synthetic rainfall events as training data. Different combinations of shapes and distributions of rainfall events were used to train RF model, and predictions were made to evaluate model accuracy. Prediction on rainfalls whose shape and distribution were included in the training data resulted in high performances. When predicting for untrained shapes and/or distribution, the accuracy of models decreases, and models trained with less accumulated rainfall indicated an overall underestimation of water depth, suggesting that the amount of accumulated rainfall is also considered necessary for the construction of augmented RF model.
Year: 2025