Event-based machine learning for extreme flood depth mapping using multi-source data: Evidence from central Vietnam

Accurate prediction of extreme flood inundation depth is essential for flood risk assessment in data-scarce coastal environments. Yet, most machine learning (ML) studies focus on flood susceptibility and rely on random data partitioning rather than event-based validation. This study proposes an event-based ML framework for continuous flood depth mapping using multi-source geospatial data, applied to the Tam Giang-Cau Hai lagoon system in Central Vietnam. Model training used the 1999 flood inventory, comprising 285 field-verified flood-depth samples and 680 non-flood samples assigned a depth of 0 m. Event-based evaluation used 42 field-surveyed observations from the 2020 flood, while three 2025 hydrological station records were retained as a supplementary consistency check. Four regression models, Random Forest (RF), Extreme Gradient Boosting (XGB), Support Vector Regression (SVR), and Multi-layer Perceptron (MLP), were developed using terrain, hydrological, surface, anthropogenic, and climatic predictors. Predictor screening used the variance inflation factor and mutual information, while feature importance was used to interpret the retained predictors. Model hyperparameters were optimized using Random Search, Particle Swarm Optimization, and the Puma Optimizer. Under 5-fold cross-validation on the 1999 training dataset, XGB optimized with the Puma Optimizer achieved the best overall performance among the tested configurations (R² ≈ 0.86; RMSE ≈ 0.24 m; MAE ≈ 0.09 m). Validation against the 42 observations from the 2020 flood yielded an R² of 0.764, an RMSE of 0.438 m, and an MAE of 0.209 m, indicating predictive skill for the evaluated event. The proposed framework demonstrates the potential of ML regression with event-based validation as a practical alternative for extreme flood depth mapping in data-scarce coastal environments.