Integrated assessment of flood susceptibility and population exposure using machine learning: a case study of Wajima, Japan

Floods increasingly disrupt daily life and cause substantial social and economic losses. Although flood susceptibility mapping (FSM) is widely used to identify flood-prone areas, the spatial relationship between FSM and population exposure has received relatively little attention. This study addresses this gap by integrating machine-learning-based FSM with census-based population distribution to assess population exposure in Machino district, Wajima, Ishikawa, Japan, which was severely affected by the September 2024 Noto Peninsula heavy rainfall event. Four models were compared: Support Vector Machine (SVM), Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Autoencoder Multilayer Perceptron (AE-MLP). Model performance was evaluated using five-fold spatial cross-validation, the area under the receiver operating characteristic curve (AUC), and other classification metrics. Mean cross-validation AUCs were 97.2% for RF, 97.1% for XGBoost, 95.4% for SVM, and 95.6% for AE-MLP. Pairwise DeLong tests using pooled out-of-fold predictions showed no significant difference between RF and XGBoost, whereas both significantly outperformed AE-MLP and SVM. The resulting FSMs consistently identified river corridors and adjacent lowlands as highly susceptible. Depending on the model, 33.34–58.08% of the population was located in moderate- to very-high-susceptibility zones. These findings underscore the importance of integrating flood susceptibility and population exposure assessments for urban planning and disaster management.