Machine Learning-Based Urban Flood Prediction Utilizing 1D-2D Modeling and Statistical Characteristics of Rainfall

Recent changes in rainfall patterns have led to an increasing trend of extreme weather events. As a result, urban flooding damge has become more severe. To predict such damage, this study developed a random forest-based urban flood prediction model that considers the statistical characteristics of rainfall distribution, focusing on Hadan-dong, Saha-gu, Busan, South Korea, a flood-prone area near the Nakdong River estuary. For model training, flood depth data from a one-dimensional and two-dimensional numerical model was utilized. First, a 1D drainage system was constructed using EPA-SWMM, and then flood depth was calculated using FLO-2D. To address data processing challenges in training on the entire rainfall time series, this study used statistical characteristics of rainfall distribution including Kurtosis, Skewness, and the Coefficient of variation as training data. A Random Forest-based urban flood prediction model was then developed using training dataset that included both flood depth and statistical characteristics. The impact of statistical characteristics was assessed by evaluating the model's performance using metrics such as R², RMSE (Root Mean Squared Error), and MAE (Mean Absolute Error). The results showed that both models, with and without statistical characteristics, demonstrated high prediction accuracy. However, the model incorporating statistical characteristics exhibited superior performance, with evaluation metrics R² = 0.9761, RMSE = 2.7354m, and MAE = 0.8942m, indicating improved predictive capability.