Flood Susceptibility Mapping Based on Supervised Machine Learning Algorithms with Integrated Uncertainty and Explainability Analysis: A Case Study in Quang Nam, Vietnam

Abstract

Flood is among the natural hazards that have occurred with increasing frequency and have changed in complex ways, with serious impacts on people's lives. A flood susceptibility map the basis for zoning areas at risk of flooding became a necessary tool not only for managers but also for residents to adapt to the broader context of climate change. This study aims to present the use of a supervised machine learning model to create a flood susceptibility map of Quang Nam province (now Da Nang city, Vietnam). Four algorithms, including Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), and Extreme Gradient Boosting (XGBoost) were tested. Based on the evaluation metrics and uncertainty analysis, RF not only achieved the best predictive performance (AUC of 0.9418) but also demonstrated higher stability and reliability. An explainability analysis using Shapley additive explanation (SHAP) and Leave-One-Covariate-Out (LOCO) was conducted to assess the contribution of conditioning factors and interpret the RF model’s predictions. The result is a flood susceptibility map based on the RF model, with levels ranging from very low to very high. Low-lying plains in the lower reaches of the Vu Gia - Thu Bon River frequently flood and are the most heavily affected areas in Quang Nam. This map was validated with 2025 flood marks data, which served as an independent data source not involved in the model construction process. The results show that the majority of actual flooding points are concentrated in moderate-, high-, and very high-sensitivity areas, accounting for 98.2%, while points in low-sensitivity areas are negligible. Generally, the results will help disaster management agencies develop effective adaptation and risk mitigation policies at the commune level in flood-prone areas. Future studies should integrate flood susceptibility with exposure and vulnerability indicators to support a more comprehensive flood risk assessment