Comparing Machine Learning Algorithms for Nationwide Flood Susceptibility Mapping in South Korea: Spatial Generalization to an Unseen City

Increasing heavy rainfall associated with climate change and the continued advance of urbanization have heightened the importance of spatially identifying flood-prone areas and responding proactively. This study constructed a nationwide 100 m grid flood susceptibility map and compared the spatial generalization performance of CatBoost, XGBoost, and a patch-based CNN under the same input data and spatial partitioning. Flood labels were constructed from 38,003 flood-trace polygons collected through the National Disaster and Safety Portal API. Eleven spatial variables related to topography, hydrology, land cover, vegetation, soil, and geology were used as inputs following quality inspection, correction, and reconstruction. Busan Metropolitan City was excluded entirely from training, early stopping, and model selection, and its 65,727 valid grid cells (natural flood rate, 3.50%) were used as a geographic external evaluation set. In non-Busan areas, 16 km spatial blocks and a 1.6 km buffer were applied to prevent spatial overlap between training and evaluation patches. In the Busan external evaluation, the ROC-AUC values of the three models were similar, ranging from 0.831 to 0.835, whereas the CNN achieved the highest PR-AUC (0.142) and F1-score (0.204)—4.1 times the random-prediction baseline, compared with 2.7 and 2.8 for CatBoost and XGBoost—indicating a reduction in false positives while comparable recall was maintained. The relative advantage of the models varied by evaluation metric, indicating the need to interpret ROC-AUC, PR-AUC, and fixed-threshold metrics together under conditions of natural class imbalance and when evaluating on a city unseen during training.