Flood prediction in urban areas based on machine learning considering the statistical characteristics of rainfall
Se-Dong Jang, Jae-Hwan Yoo, Yeonsu Lee, Byunghyun Kim
Abstract
Urbanization has increased impervious surfaces, while climate change has intensified rainfall, leading to more frequent urban flooding. Traditional numerical models for flood prediction are accurate but time-consuming due to extensive parameter calibration and data processing. This study addresses these limitations by proposing a machine learning-based flood prediction method using a Random Forest model. By utilizing past rainfall data, 1D drainage system simulations, and 2D flood analyses, we trained the model to predict flood patterns for various rainfall events. To enhance prediction accuracy, statistical characteristics of rainfall, such as temporal distribution, were incorporated into the model. Performance metrics (RMSE, R 2 , MAE) for the test dataset showed values of 3.1573, 0.9682, and 0.9484 for the total rainfall model, and 2.7354, 0.9761, and 0.8942 for the model with statistical characteristics. Both models displayed high predictive accuracy relative to the numerical model, with the Random Forest model using statistical characteristics showing slightly improved performance. This method provides faster, reliable flood predictions, offering a valuable tool for real-time urban flood management and decision-making during emergency situations. • A Random Forest-based method for rapid flood prediction in urban areas. • Incorporates rainfall's statistical characteristics to enhance prediction accuracy. • Quantitative evaluation of the predictions of numerical model results and Random Forest model predict.