A Machine Learning-based Assessment of Urban Flood Susceptibility and Priority Location of Urban Water Detention Facilities: A Case Study of Busan, South Korea

Recent increases in intense rainfall have exacerbated urban flooding, driven by impervious surfaces, drainage limitations, and topography. For predicting urban flood susceptibility, models have to consider the spatial configuration of urban hydrological infrastructure, such as urban water detention facilities (UWDF). Conventional physics-based hydrological and hydraulic models are constrained by extensive data requirements and long setup times. In contrast, machine learning (ML) models have been increasingly applied to large-scale flood prediction due to their ability to capture complex relationships among multiple factors. This study aims to assess flood susceptibility in Busan, Republic of Korea using ML models, categorize the characteristics of high-susceptibility areas, and propose optimal locations for UWDF. In this study, the dependent variable for binary classification was constructed by extracting flooded (1) and non-flooded (0) points at a 1:1 ratio, based on flood inventory maps from 2019 to 2023. The explanatory variables consisted of topographic, meteorological, land-use, and drainage infrastructure factors related to flooding (a total of 16 variables). All input datasets were prepared in raster format and resampled to a spatial resolution of 50 m, consistent with the Digital Elevation Model. The constructed dataset was randomly divided into training and testing sets at an 8:2 ratio and applied to Random Forest, Extreme Gradient Boosting, and Support Vector Machine models. Hyperparameter optimization was conducted for each model via Random Search. Then, model performance was evaluated using Accuracy and ROC-AUC metrics. For the best-performing model, Variable Importance and Partial Dependence Plot analyses were performed to interpret the relationships between key explanatory variables and flood susceptibility. Subsequently, the calculated flood susceptibilities were classified into five levels. K-means clustering was applied to high-susceptibility areas to categorize flooding event types based on shared topographic and environmental characteristics. Based on these results, area types with high potential effectiveness for UWDF were identified, and optimal installation sites were derived. The ML–based flood susceptibility analysis would quantitatively reveal and visualize the complex drivers of flooding in high-susceptibility areas.