Flood susceptibility mapping of urban flood risk: comparing autoencoder multilayer perceptron and logistic regression models in Ubon Ratchathani, Thailand

Abstract Riverine flooding in Southeast Asian cities increasingly affects their residents, often causing prolonged negative consequences due to their geographic position in lowland areas. The rapid expansion of urban areas into highly vulnerable zones has only exacerbated this issue. To mitigate flood risk, it is crucial to assess and map flood susceptibility and for incorporation into future urban planning. This study evaluates and compares the predictive performance of logistic regression (LR) and the autoencoder multilayer perceptron (AE-MLP), an advanced model, for urban flood risk assessment in Ubon Ratchathani, Thailand. The models were assessed using an area under the receiver operating characteristic curve (AUC) and several indexes (accuracy, F1-score, precision, Recall). This study also mapped vulnerable populations based on estimated flood risk from each model. The AE-MLP model outperformed the LR model, achieving an AUC of 0.950 compared to 0.860. An analysis of flood risk distribution revealed that the LR model estimated 1.79% of the population residing in high-risk areas and 0.83% in very high-risk areas. In comparison, the AE-MLP model estimated 24.00 and 7.97%, respectively, demonstrating its superior sensitivity in identifying vulnerable zones. These findings indicate that the AE-MLP model can significantly improve flood risk prediction and accurately identify high-risk areas. Integrating these models into urban planning and disaster management frameworks can enhance resilience to flooding.