Integrated machine-learning and clustering framework for selecting temporary shelters in flood-prone cities

Abstract Urban flooding has become a growing concern due to rapid land development and increasingly unpredictable climate patterns. Effective emergency response—particularly the strategic placement of temporary shelters—is crucial for minimizing the impacts of disasters. This study presents a data-driven framework for identifying and spatially optimizing emergency shelters in flood-prone urban environments. The methodology was implemented in a major metropolitan area in southern Iran, which was selected as a representative case due to its high flood risk and urban complexity. Flood susceptibility was modeled using 16 explanatory variables encompassing climatic, geographical, urban development, and urban infrastructure factors. Six machine learning algorithms were tested and compared: Random Forest, K-Nearest Neighbors (KNN), Gradient Boosting Machine (GBM), eXtreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and Categorical Boosting (CatBoost). Among them, CatBoost achieved the strongest test performance, with an F1-score of 0.8627 and an AUC of 0.9218, and was used to generate the final flood-susceptibility map. Buildings with below-median building-level susceptibility scores were retained and spatially allocated using K-Means clustering. The selected k = 28 was supported by cluster-validity and stability analyses, with a Silhouette Score of 0.5991, a Davies–Bouldin Index of 0.4578, and a mean adjusted Rand index of 0.8938. A supplementary route-exposure assessment further identified access-route segments intersecting very high flood-susceptibility zones. The proposed framework provides a transferable decision-support approach for prioritizing preliminary flood-emergency shelter.